Remote Sensing based detection of landmine suspect areas ...
INTEGRATION OF MULTISOURCE REMOTE SENSING DATA FOR ...
-
Upload
khangminh22 -
Category
Documents
-
view
0 -
download
0
Transcript of INTEGRATION OF MULTISOURCE REMOTE SENSING DATA FOR ...
Integration of multisource remote sensing data forimprovement of land cover classification
Author:Chu, Hai Tung
Publication Date:2013
DOI:https://doi.org/10.26190/unsworks/16056
License:https://creativecommons.org/licenses/by-nc-nd/3.0/au/Link to license to see what you are allowed to do with this resource.
Downloaded from http://hdl.handle.net/1959.4/52524 in https://unsworks.unsw.edu.au on 2022-06-01
INTEGRATION OF MULTISOURCE REMOTE
SENSING DATA FOR IMPROVEMENT OF
LAND COVER CLASSIFICATION
Hai Tung Chu
Supervised by A/Prof. Linlin Ge Co-supervised by Prof. Chris Rizos
A thesis submitted to the University of New South Wales for the degree of Doctor of Philosophy
School of Civil and Environmental Engineering (CVEN) Faculty of Engineering
The University of New South Wales Sydney, NSW 2052, Australia
March 2013
PLEASE TYPE
Surname or Family name: CHU
First name: HAl TUNG
THE UNIVERSITY OF NEW SOUTH WALES Thesis/Dissertation Sheet
Other name/s:
Abbreviat ion for degree as given in the Univers ity calendar: PhD
School: School of Surveying and Spatial Engineering Faculty: Engineering Faculty
Title: Integration of multisource remote sensing data for improvement of land cover classificat ion
Abstract 350 words maximum: (PLEASE TYPE)
The use of multisource remote sensing data for land cover classification has attracted the attention of researchers because the complementary
characteristics of different kinds of data can potential ly improve classificat ion results . However, using more input data does not necessarily
increase the classification performance. On the contrary, it increases data volumes, includ ing noise , redundant information and uncertainty with in
the dataset. Therefore it is essentia l to select relevant input features and combined datasets from the multisource data to achieve the best
classification accuracy. Other challeng ing tasks are the development of appropriate data process ing and classificat ion techniques to exploit the
advantages of multisource data. The goal of th is thesis is to improve the land cover classification process by using multisource remote sensing
data with recent advanced feature selection , data processing and classificat ion techniques.
The capabilit ies of non-parametric classifiers , such as Artificial Neural Network (ANN) and Support Vector Machine (SVM), were investigated using
various multisource datasets over different study areas in Vietnam and Austral ia. Resu lts showed that the non-parametric classifiers clea rly
outperformed the common ly used Maximum Likelihood algorithm.
The feature se lection techn ique based on Genetic Algorithm (GA) was proposed to search for appropriate combined datasets and classifier's
parameters . The integration of GA and SVM classifier was employed for classify ing multisource data in Western Australia . It was revealed that the
SVM-GA model gave significantly higher classification accuracy with less input data than the traditional method. The GA algorithm also performed
better than the conventional Sequential Forward Floating Search Algorithm.
To further increase the performance of multisource data classification the Multiple Classification System (MCS) was proposed and evaluated.
Moreover, the synergistic model us ing GA and the MCS was developed . An experiment was carried out with the MCS consisting of ANN , SVM and
Self-Organising Map (SOM) classifiers . Results confirmed that this newly hybrid model of GA and MCS outperformed other methods and
significantly improve the performance of both GA and MCS algorithm.
Resu lts and analysis presented in this thesis emphasise that using multisource remote sensing data is an appropriate approach for improvement
of land cover classification . The proposed methodologies are very efficient for handling combined multisource datasets.
Declaration relating to disposit ion of project thesis/dissertation
I hereby grant to the University of New South Wales or its agents the right to archive and to make available my thesis or dissertation in whole or in part in the University libraries in all forms of media , now or here after known, subject to the provisions of the Copyright Act 1968. I retain all property rights, such as patent rights . I also reta in the right to use in future works (such as articles or books) all or part of this thesis or dissertation .
I also authorise University Microfilms to use the 350 word abstract of my thes is in Dissertat ion Abstracts International (th is is applicable to doctoral theses only) .
Witness . .2~?3/~ ... .1.3>. .. ..
Date ... .. . ·~ ... .. ... .. .... . ..
The University recognises that there may be exceptional circumstances requ iring restrictions on copying or cond itions on use. Requests for restriction for a period of up to 2 years must be made in writing . Requests fo r a longer period of restriction may be considered in exceptiona l circumstances and re uire the a roval of the Dean of Graduate Research .
FOR OFFICE USE ONLY Date of completion of requ irements for Award:
THIS SHEET IS TO BE GLUED TO THE INSIDE FRONT COVER OF THE THESIS
ORIGINALITY STATEMENT
'I hereby declare that this submission is my own work and to the best of my knowledge it contains no materials previously published or written by another person, or substantial proportions of material which have been accepted for the award of any other degree or diploma at UNSW or any other educational institution, except where due acknowledgement is made in the thesis. Any contribution made to the research by others, with whom I have worked at UNSW or elsewhere, is explicitly acknowledged in the thesis. I also declare that the intellectual content of this thesis is the product of my own work, except to the extent that assistance from others in the project's design and conception or in style, presentation and linguistic expression is acknowledged.'
Signed
Date ........................ u;~ ... /.~ ..... ...t>...
COPYRIGHT STATEMENT
'I hereby grant the University of New South Wales or its agents the right to archive and to make available my thesis or dissertation in whole or part in the University libraries in all forms of media, now or here after known, subject to the provisions of the Copyright Act 1968. I retain all proprietary rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or dissertation. I also authorise University Microfilms to use the 350 word abstract of my thesis in Dissertation Abstract International (this is applicable to doctoral theses only). I have either used no substantial portions of copyright material in my thesis or I have obtained permission to use copyright material; where permission has not been granted I have applied/will apply for a partial restriction of the digital copy of my thesis or dissertation.'
Signed
Date .. ... ........ ....... .. li/Q.)/ ..... MJ.J;) ....... .
AUTHENTICITY STATEMENT
'I certify that the Library deposit digital copy is a direct equivalent of the final officially approved version of my thesis. No emendation of content has occurred and if there are any minor variations in formatting, they are the result of the conversion to digital format.'
Signed
Date ....................... 2(/D>./. Q!j ...
i
ABSTRACT
The use of multisource remote sensing data for land cover classification has attracted the
attention of researchers because the complementary characteristics of different kinds of data
can potentially improve classification results. Such a multisource approach becomes
increasingly important with the ready availability of a variety of satellite imagery. However,
using more input data does not necessarily increase the classification performance. On the
contrary, using too many remotely sensed input datasets increases data volumes, including
noise, redundant information and uncertainty within the dataset. Therefore it is essential to
select relevant input features and combined datasets from the multisource data to achieve the
best classification accuracy. Other challenging tasks are the development of appropriate data
processing and classification techniques to efficiently exploit the advantages of multisource
data. The goal of this thesis is to improve the land cover classification process by using
multisource remote sensing data with recent advanced input feature selection, data processing
and classification techniques.
The capabilities of non-parametric classifiers, such as the Artificial Neural Network (ANN)
and Support Vector Machine (SVM), were investigated using various multisource datasets
over different study areas in Vietnam and Australia. Results showed that the multisource
datasets always gave higher classification accuracy than the single-type datasets. The non-
parametric classifiers clearly outperformed the commonly used Maximum Likelihood
algorithm.
The feature selection (FS) technique, specifically the wrapper approach, based on Genetic
Algorithm (GA) was proposed to search for the appropriate combined datasets and
classifier’s parameters. The integration of GA and SVM classifier was employed for
classifying multisource data in Western Australia, including multi-date, multi-polarised SAR
and optical images. It was revealed that the SVM-GA model gave significantly higher
classification accuracy with less input data than the traditional method. The GA algorithm
also performed better than the conventional Sequential Forward Floating Search algorithm.
The Multiple Classifier Systems (MCS) or classifier ensemble technique, which can
potentially improve classification performance by exploiting the strengths and alleviating the
ii
weaknesses of different classifiers, was also evaluated using different algorithms and
combination rules. An experiment was carried out with the MCS technique using the ANN,
SVM and Self-Organising Map (SOM) classifiers over a study area in New South Wales,
Australia. The investigation shows that the MCS technique, in general, provided higher
classification accuracy than individual classifiers.
Finally, a synergistic model using the FS based on GA and MCS techniques was developed to
further increase the performance of multisource data classification. Results confirmed that the
hybrid model of FS-GA and MCS outperformed other methods and significantly improved on
the performance of both the FS-GA and MCS algorithm.
Results and analyses presented in this thesis emphasise that using multisource remote sensing
data is an appropriate approach for improvement of land cover classification. The proposed
methodologies are very efficient for handling high dimensional, complex datasets such as
combined multisource data.
iii
ACKNOWLEDGEMENT
First of all I would like to thank Vietnamese Ministry of Education and Training (MOET) for
granting me a scholarship without which I would not be able to carry out this research. I
would like also to thank the National Centre for Remote Sensing, the Ministry of
Environment and Natural Resources (MONRE) of Vietnam for giving me an opportunity to
attend the PhD program and providing relevant image data for my study.
I would like to thank for my supervisor A. Prof. Linlin Ge for his valuable support and advice
during my research. I would also like to express my gratitude to my co-supervisor, Prof.
Chris Rizos for his guidance, and patiently correcting my writing.
I would like to show my thanks to the University of New South Wales and the School of
Surveying and Geospatial Engineering for allowing me to study there. I am also grateful to
our GEOS team members, including Dr. Jean Li, Dr. Mahmood Salah, Dr. Kui Zhang, Dr.
Alex Ng and Mr. Alex Hu for their enthusiasm and support.
I thank my friend Mr. Lawrence Greaves for his enthusiasm in helping me correct my
writing.
I wish to thank the European Space Agency (ESA) and Japan Aerospace Exploration Agency
(JAXA) for providing ENVISAT/ASAR and ALOS/PALSAR data that have been used in
this study. I would like also to thank for United State Geological Survey/Earth Resources
Observation and Science Center (USGS/EROS) for providing Landsat 5 TM data.
Last but not least, my special thanks go to my parents, my wife and daughters for their love,
constant support and understanding during the long journey of my study.
iv
LIST OF PUBLICATIONS
Journal papers
1. Chu, H. T., Ge, L., Ng, A. H-M., and Rizos, C., 2012. Application of Genetic
Algorithm and Support Vector Machine in Classification of Multisource Remote
Sensing Data. International Journal of Remote Sensing Application, Vol. 2, No. 3, pp.
1- 11.
2. Chu, H. T., and Ge, L., (in preparation). Integration of Feature Selection and Multiple
Classifier System for improvement of land cover classification using multisource
remote sensing data. International Journal of Applied Remote Sensing.
Conference papers
1. Chu, H. T., and Ge, L., 2010, Synergistic use of multi-temporal ALOS/PALSAR with
SPOT multispectral satellite imagery for Land cover Mapping in The Ho Chi Minh
City area, Vietnam. International Geoscience and Remote Sensing Symposium
(IGARSS), Honoluulu, HI, 25-30 July.
2. Chu, H.T., Li, X., Ge, L., & Zhang, K., 2010. Monitoring the 2009 Victorian
bushfires with multi-temporal and coherence ALOS PALSAR images. 15th
Australasian Remote Sensing & Photogrammetry Conf., Alice Springs, Australia, 13-
17 September, 313-325 (http://www.15.arspc.com/proceedings).
3. Chu, H.T., & Ge, L., 2010. Land cover classification using combinations of L- and C-
band SAR and optical satellite images. 31st Asian Conf. on Remote Sensing, Hanoi,
Vietnam, 1-5 November, paper TS39-2, CD-ROM procs.
4. Chu, H.T., Ge, L., & Wang, X., 2011. Using dual-polarized L-band SAR and optical
satellite imagery for land cover classification in Southern Vietnam: Comparison and
combination. Proc. published 2011 in Australian Space Science Conference Series,
ed. W. Short & I. Cairns, 10th Australian Space Science Conf., Brisbane, Australia,
27-30 September 2010, 161-173. (Refereed paper)
v
5. Chu, H.T., & Ge, L., 2011. Improvement of land cover classification performance in
Western Australia using multisource remote sensing data. 7th Int. Symp. on Digital
Earth, Perth, Australia, 23-25 August.
6. Chu, H.T., & Ge, L., 2012. Combination of genetic algorithm & Dempster-Shafer
theory of evidence for land cover classification using integration of SAR & optical
satellite imagery. XXII Int. Society for Photogrammetry & Remote Sensing Congress,
Melbourne, Australia, 25 Aug - 1 September.
7. Chu, H.T., Ge, L., & Cholathat, R., 2012. Evaluation of multiple classifier
combination techniques for land cover classification using multisource remote sensing
data. 33rd Asian Conference on Remote Sensing (ACRS2012), Pattaya, Thailand, 26-
30 November, paper F5-1, CD-ROM procs.
vi
ACRONYMS AND ABBREVIATIONS
ALOS Advanced Land Observing Satellite
ANN Artificial Neural Network
ASAR Advanced Synthetic Aperture Radar
BP Back Propagation
CCRS/CCT Canada Centre for Remote Sensing/Centre Canadien de teledetection
DEM Digital Elevation Model
ENVISAT Environmental Satellite
ERS European Remote Sensing Satellite
ERS- ½ 1 /2 European Remote-Sensing Satellites
ESA European Space Agency
ETM+ Enhanced Thematic Mapper Plus
FS Feature Selection
GA Genetic Algorithm
GLCM Grey Level Co-occurrence Matrix
HH Horizontal transmit and horizontal receive
HV Horizontal transmit and vertical receive
JERS-1 Japanese Earth Resources Satellite-1
JAXA Japan Aerospace Exploration Agency
LiDAR Light Detection and Ranging
MLP Multilayer Perceptron
MODIS Moderate Resolution Imaging Spectroradiometer
NDVI Normalized Difference Vegetation Index
PALSAR Phased Array type L-band Synthetic Aperture Radar
P (Pan) Panchromatic
PCA Principle Component Analysis
RADAR Radio Detection And Ranging
SAR Synthetic Aperture Radar
SFS/SBS Sequential Forward Search/Sequential Backward Search
SFFS/SBFS Sequential Floating Forward/Backward Search
SIR-L/C/X Shuttle Imaging Radar-L/C/X
vii
SOM Self-organizing Map
SPOT Systeme Probatoire d’Observation de la Terre
SVM Support Vector Machine
TM Thematic Mapper
VH Vertical transmit and horizontal receive
VV Vertical transmit and vertical receive
XS Multispectral
viii
CONTENTS ABSTRACT i
ACKNOWLEDGEMENT iii
LIST OF PUBLICATIONS iv
ACRONYMS AND ABBREVIATIONS vi
CONTENTS viii
LIST OF TABLES xiii
LIST OF FIGURES ix
CHAPTER 1: INTRODUCTION 1
1.1. Problem statement 1
1.2. Objectives 3
1.3. Main contributions 4
1.4. Thesis outline 4
CHAPTER 2: BACKGROUND AND LITERATURE REVIEW 6
2.1. Image classification methods 6
2.1.1. Traditional supervised classification and Maximum likelihood
algorithm
7
2.1.2. Unsupervised classification 8
2.1.3. Non-parametric classification 9
2.2. Data pre-processing 20
2.2.1. Geometric correction 20
2.2.2. Radiometric calibration 22
2.2.3. Speckle noise filtering 24
2.3. Advantages of multisource remote sensing data, selection and integration
issues
26
2.3.1. Advantages of the combined approach 26
2.3.2. Selection and integration of remote sensing data and feature
parameters
37
2.4. Related work on land cover classification using multisource remote sensing
data
42
2.5. Summary 49
ix
CHAPTER 3: EVALUATION OF LAND COVER CLASSIFICATION USING NON-PARAMETRIC CLASSIFIERS AND MULTISOUCE REMOTE SENSING DATA 52
3.1. Application of multi-temporal/polarized SAR and optical imagery for land
cover classification 52
3.1.1. Study area and data used 53
3.1.2. Methodology 54
3.1.3. Results and discussion 58
3.1.4. Summary remarks 72
3.2. Combination of L- and C-band SAR and optical satellite images for land
cover classification in the rice production area 73
3.2.1. Study area and data used 74
3.2.2. Methodology 75
3.2.3. Results and discussion 77
3.2.4. Summary remarks 86
3.3. Use of multi-temporal SAR and interferometric coherence data for
monitoring the 2009 Victorian bushfires 86
3.3.1. Introduction 86
3.3.2. Study area and data used 88
3.3.3. Methodology 89
3.3.4. Results and discussion 91
3.3.5. Summary remarks 98
3.4. Concluding remarks 98
CHAPTER 4: APPLICATION OF FEATURE SELECTION TECHNIQUES FOR MULTISOURCE REMOTE SENSING DATA CLASSIFICATION 99
4.1. Significance of data reduction and feature selection for classification of
remote sensing data 99
4.2. Feature selection techniques used for classification of remote sensing data 101
4.3. Application of genetic algorithm and support vector machine in classification
of multisource remote sensing 108
4.3.1. Introduction 108
4.3.2. Study area and data used 109
4.3.3. Methodology 110
x
4.3.4. Results and discussion 116
4.3.5. Conclusions 123
CHAPTER 5: APPLICATION OF MULTIPLE CLASSIFIER SYSTEM FOR CLASSIFYING MULTISOURCE REMOTE SENSING DATA 125
5.1. MCS in remote sensing 125
5.1.1. Creation of MCS or classifier ensemble 126
5.1.2. Combination rules 127
5.2. Use of MCS and a combination of GA and MCS for classifying multisource
remote sensing data – a case study in Appin, NSW, Australia. 130
5.2.1. Introduction 131
5.2.2. Study area and used data 133
5.2.3. Methodology 129
5.2.4. Results and discussion 141
5.2.5. Conclusions 151
CHAPTER 6: CONCLUSIONS 152
6.1. Summary and conclusions 152
6.2. Main findings 155
6.3. Future work 155
REFERENCES 157
xi
LIST OF FIGURES
Figure 2.1: Spectral classes represented by normal probability distribution ............................. 8
Figure 2.2: Artificial Neural Network classifier ...................................................................... 11
Figure 2.3: Example of linear support vector machine ............................................................ 14
Figure 2.4: SVMs projecting the training data to higher dimensional space. .......................... 16
Figure 2.5: Effect of platform/sensor position and orientation on geometry of image ........... 21
Figure 2.6: Relief displacement in optical and radar satellite image ....................................... 22
Figure 2.7: Spectral signatures of soil, vegetation and water in visible/infrared region of
spectrum ................................................................................................................................... 27
Figure 2.8: Microwave bands used in SAR systems ............................................................... 28
Figure 2.9: Penetration of SAR signals at different wavelengths or frequencies .................... 31
Figure 2.10: Radar transmission in vertical and horizontal polarisation ................................. 32
Figure 2.11: Surface scattering and volume scattering ............................................................ 33
Figure 3.1: Location of the study area ..................................................................................... 53
Figure 3.2: Land cover feature characteristics in SPOT 2 multi-spectral and ALOS/PALSAR
multi-temporal images ............................................................................................................. 56
Figure 3.3: Part of classification results using SVM and ANN classifiers on multi--date,
single-polarised (HH or HV), and multi-date, dual-polarised HH+HV images; ..................... 63
Figure 3.4: Part of the classification results using SVM and ANN classifiers on a SPOT 2 XS
image and the combination of SPOT 2 XS and multi-date ALOS/PALSAR HH polarised
images. ..................................................................................................................................... 65
Figure 3.5: Part of classification results using SVM and ANN classifiers on a SPOT 2 XS
image and combination of SPOT 2 XS and multi-date PALSAR HH, HV polarised datasets.
................................................................................................................................................. 68
Figure 3.6: Comparison of classification accuracy of SVM, ANN and ML classifiers .......... 71
Figure 3.7: Comparison of Kappa coefficients generated by SVM, ANN and ML classifiers 71
Figure 3.8: SPOT 4 multispectral false colour image (left) and ALOS/PALSAR HH
polarisation image (right), both images acquired on December 09, 2007 ............................... 75
Figure 3.9: Spectral properties of different land cover classes in the SPOT 4 XS image ....... 77
Figure 3.10: Backscatter properties of different land cover classes in the multi-date L- and C-
band SAR images .................................................................................................................... 77
xii
Figure 3.11: SVM Classification using four-date PALSAR images (left) and the combination
of four-date ASAR and four-date PALSAR images (right) .................................................... 80
Figure 3.12: ANN Classification using four-date PALSAR images (left) and the combination
of four-date ASAR and four-date PALSAR images (right) .................................................... 81
Figure 3.13: ANN classification using the SPOT 4 multi-spectral image (left) and
combination of SPOT 4 and four-date PALSAR images (right) ............................................. 82
Figure 3.14: SVM classification using the SPOT 4 multi-spectral image (left) and
combination of SPOT 4 and four-date PALSAR images (right) ............................................. 82
Figure 3.15: Comparison of classification accuracy of SVM, ANN and ML classifiers after
merging rice classes (only combined datasets with at least two input features are presented) 85
Figure 3.16: A) and C) showed fire affected areas are in purple to pink colour in multi-date
colour composites in 1st and 2nd study sites. B) and D) showed burnt area (dark colour) in
corresponding false colour Landsat TM images. ..................................................................... 91
Figure 3.17: A) and C) are RGB colour composites of Average SAR image (R), Temporal
Backscatter Change (G) and across-fire coherence data (B) in 1st and 2nd study sites. Burnt
areas appear Yellow to Reddish in colour in A) and C) RGB composites. ............................. 93
Figure 3.18: Backscatter values for burnt and unburnt classes over the first test site ............. 94
Figure 3.19: Coherence values for burnt and unburnt classes over the first test site .............. 94
Figure 3.20: Backscatter values for burnt and unburnt classes over the second test site ........ 95
Figure 3.21: Coherence values for burnt and unburnt classes over the second test site .......... 95
Figure 3.22: A) Burnt scars extraction from Landsat TM images, B) Burnt scars extraction
based on SVM classification (case 3), the Red areas represent burnt scars ............................ 97
Figure 4.1: An example of the Hughes phenomenon with increasing of input data
dimensionality........................................................................................................................ 100
Figure 4.2: Feature selection with filter (left) and wrapper (right) approach ........................ 104
Figure 4.3: Illustration of the crossover and mutation operators ........................................... 105
Figure 4.4: Landsat 5 TM (left) and ALOS/PALSAR HH (right) over the study area acquired
on 07/10 and 20/10/2010, respectively .................................................................................. 105
Figure 4.5: The binary coding of the chromosome ................................................................ 113
Figure 4.6: Major steps of feature selection using the Genetic Algorithm…………………114
Figure 4.7: Land cover feature characteristics in Landsat 5 TM (left) and multi-date PALSAR
(right) images ………………………………………………………………………………115
Figure 4.8: Improvement of accuracy by incorporating textural information with original
datasets using stack-vector and FS-GA approach. ………………………………………...119
xiii
Figure 4.9: Classification of the Landsat 5 TM image (left) and the integration of Landsat 5
TM and four-date PALSAR HH/HV images (right) using the GA technique……………..120
Figure 4.10: Classification accuracy using FS (SFFS and GA) compared to Non-FS (stack-
vector) approaches…………………………………………………………………………..122
Figure 4.11: Number of input features for the Non-FS (stack-vector) and FS (SFFS and GA)
approaches…………………………………………………………………………………..122
Figure 4.12: Impact of the proposed fitness function on overall classification accuracy
compared to the commonly used fitness function………………………………………….123
Figure 5.1: ENVISAT/ASAR VV polarised image acquired on 25/09/2010 ....................... 132
Figure 5.2: ALOS/PALSAR HH polarised image acquired on 07/10/2010 ......................... 132
Figure 5.3: Landsat 5 TM images (false colour) acquired on 10/09/2010 ............................ 133
Figure 5.4: Example of the SOM architecture with three input neurons and 25 (5x5) output
neurons……………………………………………………………………………………...136
Figure 5.5: The integration of feature selection based on GA and MCS techniques for
classifying multisource data………………………………………………………………...140
Figure 5.6: Results of classification of the 4th dataset using FS-GA with SVM classifier…143
Figure 5.7: Improvements of accuracy by applying the FA-GA approach for the SVM, ANN
and SOM classifiers…………………………………………………………………………143
Figure 5.8: Comparison of SVM classification with SVM-Bagging and SVM-Adaboost.M1
methods……………………………………………………………………………………..146
Figure 5.9: Comparison of ANN classification with ANN-Bagging and ANN-Adaboost.M1
methods……………………………………………………………………………………..146
Figure 5.10: Comparison of SOM classification with SOM-Bagging and ANN-Adaboost.M1
methods……………………………………………………………………………………..146
Figure 5.11: Comparison between the best classification results obtained by original
classifiers, MCS based and FS-GA approaches…………………………………………….148
Figure 5.12: Comparison between the BIC, FS-GA, the best results MCS, the best (FS-GA-
MCS1) and the poorest (FS-GA-MCS2) results of FS-GA-MCS methods for different
datasets……………………………………………………………………………………...149
Figure 5.13: Improvements of overall classification accuracy achieved by using FS-GA, MCS
and FS-GA-MCS approaches……………………………………………………………….150
xiv
LIST OF TABLES
Table 2.1: Specification of some optical satellite imagery…………………………………..40
Table 2.2: Specification of different SAR satellite imagery…………………………………41
Table 3.1: ALOS/PALSAR images for the study area………………………………………53
Table 3.2: Average values of Transformed Divergence (TD) and Jefferies-Matusita
separability indices of different combined datasets…………………………………………59
Table 3.3: Land cover classification accuracy of different SAR combination datasets….....61
Table 3.4. Producer and user accuracy (%) for the SVM classifier applied to five-date
PALSAR HH, HV, and five-date PALSAR dual-polarised (HH+HV) images…………......62
Table 3.5. Producer and user accuracy (%) for the ANN classifier applied to five-date
PALSAR HH, HV, and five-date PALSAR dual-polarised (HH+HV) images……………..62
Table 3.6: Land cover classification accuracy of different combinations of SPOT 2 XS and
PALSAR polarised images…………………………………………………………………..64
Table 3.7: Producer and user accuracy for ANN and SVM classifier applied to SPOT 2 XS
multispectral images…………………………………………………………………………66
Table 3.8: Producer and user accuracy for ANN and SVM classifier applied to combination
of SPOT 2 XS multi-spectral and five-date PALSAR HH polarised images……………….66
Table 3.9: Comparison of land cover classification with multi-date SAR images including
like-, cross- and dual-polarised data and combination of these images with their best textural
features………………………………………………………………………………………69
Table 3.10: ALOS/PALSAR and ENVISAT/ASAR images for the study area…………….74
Table 3.11: Overall classification accuracies (%) for different datasets…………………....78
Table 3.12: Confusion matrix for SVM classification using four-date ENVISAT/ASAR HH
polarised image………….…………………………………………………………………...79
Table 3.13: Confusion matrix for ANN classification using four-date ENVISAT/ASAR HH
polarised image………………………………………………………………………………79
Table 3.14: Confusion matrix for SVM classification using four-date ALOS/PALSAR HH
polarised images……………………………………………………………………………..79
Table 3.15: Confusion matrix for ANN classification using four-date ALOS/PALSAR HH
polarised images……………………………………………………………………………..80
xv
Table 3.16: Confusion matrix for SVM classification using four-date ALOS/PALSAR HH &
four-date ENVISAT/ASAR polarised images……………………………………………….81
Table 3.17: Confusion matrix for ANN classification using four-date ALOS/PALSAR HH &
four-date ENVISAT/ASAR polarised images……………………………………………….81
Table 3.18: Confusion matrix for ANN classification using SPOT 4 XS images…………...83
Table 3.19: Confusion matrix for SVM classification using SPOT 4 XS images…………...83
Table 3.20: Confusion matrix for ANN classification using SPOT 4 XS images and four-date
ALOS/PALSAR polarised images…………………………………………………………...83
Table 3.21: Confusion matrix for SVM classification using SPOT 4 XS images and four-date
ALOS/PALSAR polarised images…………………………………………………………...83
Table 3.22: ALOS/PALSAR images for the 1st study area…………………………………..88
Table 3.23: Interferometric pairs over 1st study area…………………………………………88
Table 3.24: ALOS/PALSAR images for the 2nd study area………………………………….88
Table 3.25: Interferometric pairs over 2nd study area……………………………………...…89
Table.3.26: Overall classification accuracy assessment for various combined datasets and
classifiers over the second test site…………………………………………………………...96
Table 4.1: ALOS/PALSAR images of the Western Australia study area…………………..109
Table 4.2: Contents of training and testing datasets………………………………………..116
Table 4.3: Classification accuracy of different datasets using SVM classifiers with the
traditional stack-vector approach…………………………………………………………...117
Table 4.4: Producer and user accuracy (%) of four-date PALSAR HH, HV and dual-polarised
images………………………………………………………………………………………117
Table 4.5: Comparison of land cover classification performance between SVM-GA, SVM-
SFFS and stack-vector approach; nf = number of selected features………………………..121
Table 5.1: ENVISAT/ASAR and ALOS/PALSAR images for the study area……………..131
Table 5.2: Combined datasets for land cover classification in the study area………………134
Table 5.3: Contents of training and testing datasets………………………………………...140
Table 5.4: The classification performance of SVM, ANN and SOM algorithms
on different combined multisource datasets………………………………………………...141
Table 5.5: Comparison of classification performance between FS-GA approach and
the Non-FS pproach………………………………………………………………………...142
Table 5.6: Results of classification using single SVM classifier, bagging and Adaboost.M1
techniques based on SVM classifier for different combined datasets………………………145
Table 5.7: Results of classification using single ANN classifier, bagging and
xvi
Adaboost.M1 techniques based on ANN classifier for different combined datasets……….145
Table 5.8: Results of classification using single SOM classifier, bagging and
Adaboost.M1 techniques based on SOM classifier for different combined datasets………145
Table 5.9: Comparison between the best classification results obtained by individual
classifiers and MCS based on the decision rules of Majority Voting, Sum and Dempster-
Shafer theory………………………………………………………………………………..147
Table 5.10: Comparison of best classification results using single classifier (Non-FS), FS-GA
and FS-GA-MCS classifier combination approaches………………………………………148
Table 5.11: Classification accuracies by applying MV, Sum and DS algorithm for
combination of FS-GA and MCS approaches……………………………………………...150
1
CHAPTER 1
INTRODUCTION
1.1. PROBLEM STATEMENT
Land cover information plays an important role in sustainable management, development and
exploitation of natural resources, environmental protection, planning, scientific analysis,
modelling and monitoring. These data become even more essential when there are rapid
changes on the Earth’s surface due to dynamic human activities as well as natural factors.
Remotely sensed data, in particular satellite images, with distinct advantages such as large
ground coverage, synoptic view, repetitive capabilities, multiple spectral bands or multiple
frequency/polarisation are one of the most effective tools for land cover mapping and have
been applied extensively for land cover monitoring and classification.
Satellite imagery can be acquired in various regions of the electromagnetic spectrum, from
the visible-near infrared (optical) to the microwave (radar) parts of the spectrum.
Consequently, different kinds of satellite imagery detect different characteristics of ground
surfaces. For instance, the optical images from missions such as Landsat, SPOT, MODIS,
IKONOS or Quick Bird provide information essentially on the reflectivity and absorption
capability of land cover features, since the imaging sensors are sensitive to the visible to near-
infrared regions of the spectrum. On the other hand, Synthetic Aperture Radar (SAR)
imagery, provided by missions such as RADARSAT, ENVISAT/ASAR, ALOS/PALSAR or
TerraSAR-X sensitive to the microwave region of the spectrum, contain information on
surface roughness, dielectric content and the structures of the illuminated ground or
vegetation. Thus integration of different types of satellite images could provide
complementary information and consequently improve the land cover classification results.
This approach has been fuelled by the large variety of remote sensing sensors which are now
2
more readily available than at any time in the past. Furthermore, derived information from
original remote sensing data, such as textural information, ratios, indices, coherence data,
Digital Elevation Model (DEM), or from other ancillary data sources can also contribute to
an increase in classification accuracies (Lu and Weng 2007, Tso and Mather 2009).
However, the increase in the amount of input data does not necessarily mean an increase in
classification performance. In addition, such an approach will create a large data volume with
more noisy and redundant data, which may reduce the classification accuracy (Lu and Weng
2007). Waske and Benediktsson (2007) suggested that data sources may not be at the same
level of reliability, where one source is more appropriate for one specific feature and another
source is more applicable to mapping another feature.
Therefore, the challenging tasks are to understand the contribution of each dataset, to select
the most useful input features and to determine the combined datasets which can maximise
the benefits of multisource remote sensing data and give the highest classification accuracy
(Peddle and Ferguson 2002). However, limited research has explored ways to determine
variables from multisource data in order to increase the classification accuracy (Lu and Weng
2007, Li et al. 2011). The Feature Selection (FS) techniques, which have been used
effectively in many applications including classification of remote sensing images (mainly
multispectral and hyperspectral data) but have not been well studied for classifying
multisource remote sensing data, could be useful for addressing this problem.
The classification algorithms are vital for classification. Appropriate selection of
classification algorithms can result in a substantial improvement in the quality of the
classification results. The traditional classification algorithms such as Minimum Distance,
Maximum Likelihood classifiers have been used widely to classify remote sensing images.
These classifiers can produce relatively good classification results in a comparatively short
time. However, the major limitation of these classifiers is their reliance on statistical
assumptions which may not sufficiently model remote sensing data. Furthermore, it is
difficult for statistical-based classifiers to incorporate different kinds of data for
classification. Recently, non-parametric classification techniques, based on machine learning
theory, have been developed. Unlike conventional classifiers, the new classification
algorithms, such as Artificial Neural Network (ANN), Support Vector Machine (SVM) or
Self Organising Map (SOM), do not rely on statistical principles and therefore can handle a
3
complex dataset more effectively (Pal and Mather 2005, Waske and Benedikson 2007).
Because of these properties, non-parametric classifiers are considered as important
alternatives to traditional methods, and the application of non-parametric classifiers to
classify remotely sensed data has become an attractive subject for research.
The classification performance could be enhanced further by applying a multiple classifier
systems (MCS) since it may take advantages of, and compensate weaknesses of, different
classifiers. Despite the robustness of such techniques, few studies have been conducted on the
application of these techniques for classifying combinations of different remote sensing
datasets.
Another strategy which has the potential to improve classification of multisource data is
integration of the MCS techniques with FS, particular the wrapper approach couple with the
Genetic Algorithm (GA). This strategy has not been considered by other researchers, and
therefore is in need of investigation.
1.2. OBJECTIVES
The objectives of this research were to improve the performance of land cover classification
using multisource remote sensing data with different advanced data processing, classification
techniques (including non-parametric classifiers), feature selection, multiple classifier
systems and their integration. The thesis objectives were:
To evaluate the capabilities of various non-parametric classification algorithms such
as ANN, SVM or SOM in classification of different combinations of multisource data,
including multi temporal/polarisation/frequency, coherence (for SAR), multispectral
data (for optical), transformed images, and textured measures; and to analyse the
impacts of these kinds of data and their contributions to final classification results.
To propose and evaluate the combination of FS techniques with the wrapper approach
and non-parametric classifiers for optimising datasets and classifier’s parameters in
land cover classification using multisource remote sensing data.
To evaluate the capabilities of the MCS using various algorithm and decision
approaches for classifying multisource data.
4
To propose and develop models for integration of FS wrapper techniques with MCS
for improvement in the classification performance of multisource remote sensing data.
1.3. MAIN CONTRIBUTIONS
The contributions of this research are:
1. Demonstration of the significance and usefulness of multisource remote sensing data
for land cover classification.
2. Demonstration of the strength and suitability of non-parametric classifiers for
classifying multisource data. These algorithms are capable of high classification
accuracy and often outperformed the traditional parametric classifiers.
3. Resolution of the problems of finding optimal combined datasets and classifier
parameters in the classification of high-dimensional multisource data by introducing
the feature selection technique based on the GA. This is a very effective method
which can increase classification accuracy while using fewer input features.
4. Analysis and comparisons of various MCS algorithms, including boosting and
bagging, majority voting, Bayesian sum and evidence reasoning, for improving the
classifications of complex multisource datasets.
5. Development of an effective method and model for combining the FS wrapper based
on GA and the MCS technique for further improvement of land cover classification
using multisource data. This new approach is capable to incorporate advantages of
both FS based on GA and MCS techniques for increasing classification accuracy.
1.4. THESIS OUTLINE
The problem statement, research objectives, contributions and thesis outline are given in
Chapter 1.
Chapter 2 reviews the principles of digital image classification techniques, potential
advantages of multisource remote sensing data, including multi-temporal/polarised/frequency
SAR data and optical images, their derivatives and ancillary data for land cover classification
and highlights the results of previous studies in this research area.
5
Chapter 3 reports studies on the use of non-parametric classifiers, particularly Artificial
Neural Network (ANN) and Support Vector Machine (SVM) in handling complex spatial
datasets such as combinations of SAR and multispectral satellite imagery. These studies
revealed the advantages of non-parametric classifiers compared to the traditional parametric
classifier for classifying multisource data. Impacts and contributions of each kind of spatial
data on the final classification of integrated datasets were also evaluated and highlighted.
Chapter 4 discusses different FS techniques used for remote sensing data analysis, such as
separability indices, sequential feature selection and GA, filter and wrapper approaches. The
FS based on the wrapper approach and GA technique has been proposed because of its ability
to handle global optimisation with large datasets. In this chapter the results of a case study of
the application of GA and SVM for classification of multisource satellite image data in
Western Australia are presented.
Chapter 5 discusses possibilities of using multiple classifiers systems to improve the
classification performance. The combination of the FS techniques with MCS was proposed
and implemented. This chapter also presents results of a case study in Appin, NSW,
Australia.
The conclusion, main findings and suggestions for future work are presented in Chapter 6.
6
CHAPTER 2
BACKGROUND AND LITERATURE REVIEW
The goal of this thesis is to improve land cover classification using multisource remote
sensing data. However, it does not intend to cover all of the algorithms, methodologies and
aspects of image classification. This thesis concentrates on the pixel-based classification
approach and a number of recently developed supervised classification algorithms.
In this chapter, firstly techniques of image classification using several kinds of classification
algorithms are presented, and advanced techniques such as feature selection and multiple
classifier systems (MCS) – considered as alternatives to traditional classification algorithms
in handling remote sensing data – are also briefly introduced. Major data pre-processing steps
are also described. The advantages of using multisource remote sensing data, particularly a
combination of SAR and optical images, will be discussed in detail. Finally, related studies
which have been undertaken in the utilisation of multisource remote sensing data for land
cover classification are examined in order to suggest directions for future research activities
in this field.
2.1. IMAGE CLASSIFICATION METHODS Image classification is a process of assigning a pixel to a pre-defined feature class.
Supervised and unsupervised classifiers are two classical methods of classification based on
spectral response characteristics of various land cover features present on images (Lillesand
and Kiefer 2004).
7
2.1.1. TRADITIONAL SUPERVISED CLASSIFICATION AND THE MAXIMUM
LIKELIHOOD ALGORITHM
The supervised classification process typically consists of several steps. Firstly, training areas
which represent the cover types or classes are selected. These training areas will be used to
generate the statistics, such as mean value, standard deviation and covariance matrix for each
class. These coefficients will then be subsequently used in particular classification
algorithms. Finally, decision rules, such as maximum likelihood classification, minimum
distance, will be used to assign pixels to classes.
Maximum likelihood algorithm
The maximum likelihood (ML) classifier is one of the most commonly used for remotely
sensed data classification (Jensen 2004, Lu and Weng 2007, Waske and Braun 2009, Lu et al.
2011). This algorithm is based on the assumption that all pixels of the training areas or
classes follow a normal distribution, as indicated in Figure 2.1 (Richards and Jia 2006).
Feature classes can then be statistically described by the mean vector and covariance matrix
of the spectral response pattern. According to these parameters the probabilities of pixels
belonging to a particular class are computed. Finally, the pixel is assigned to the class at
which the probability value is greatest. Among conventional classification algorithms the
maximum likelihood classifier usually provides the best accuracy of classification.
The Bayesian classifier is an important modification of the maximum likelihood classifier.
This classifier introduces a prior probability of occurrence for each class to be classified as
well as the cost for misclassification for each class. The classification will be implemented
with the condition that the loss due to classification is minimum over all classes. In other
words, the classification will be optimum. Nevertheless, in practice the prior probabilities are
usually unknown, hence it is often assumed that the prior probabilities are equal for all
classes (Richards and Jia 2006).
8
Figure 2.1: Spectral classes represented by normal probability distribution (source: Richard
and Jia 2006)
2.1.2. UNSUPERVISED CLASSIFICATION
Unlike supervised classification, in this method cover types to be classified and training areas
are not chosen. Instead, pixels are grouped into classes based on natural clusters present in
the image data. Classes generated by unsupervised classification are solely based on
properties of pixels (spectral or backscatter values), and properties of pixels within the same
class are similar. In the end, corresponding labels for each class will be assigned.
The simplest and most popular algorithm for computing differences between a pixel and
cluster centers is the Euclidean distance metric. The clustering algorithm mainly used in
unsupervised classification is the migrating mean or ISODATA algorithm. In this method, a
number of mean values are initially selected for clusters in multispectral data. Distance from
pixel to cluster mean values are then calculated and pixels will be assigned to the nearest
clusters. After all pixels have been classified, the clusters’ mean values will be recalculated.
The new set of clusters’ mean values then will be used to reclassify the image data. The
iteration process is terminated when the change in value of the cluster mean between two
successive iterations is smaller than some pre-defined threshold (Richard and Jia 2006). It
9
depends on actual contexts whether unsupervised or supervised classification, or a combined
approach (hybrid approach), can be applied.
2.1.3. NON-PARAMETRIC CLASSIFICATION
As mentioned earlier, the commonly used ML classifier is based on the assumption that all
pixels of the training areas or classes follow a normal distribution. Therefore, it is also
referred to as parametric classification. In fact, this assumption is not always true for
remotely sensed images, especially with a complex dataset (Dixon and Candade 2008, Waske
and Braun 2009). Another drawback of the parametric classifier as mentioned by Lu and
Weng (2007) is its difficulty in integrating spectral data with other kinds of data. Waske and
Braun (2009) pointed out that it is reasonable to weight different input data sources for the
classification process because of their different ability to discriminate land cover classes.
However, these weighting factors are not applicable to the conventional parametric
techniques. The non-parametric classifier does not require a normal distribution of the dataset
and, consequently, no statistical parameters are needed to differentiate land cover classes.
Thus, non-parametric classifiers are particularly suitable for incorporation of non-spectral
data into the classification process (Lu and Weng 2007, Tso and Mather 2009, Waske and
Braun 2009). This property has made non-parametric classification techniques very attractive
for land cover mapping. Previous studies revealed that non-parametric classification may give
better classification results than parametric classification, particularly in complex study areas
(Berberoglu et al. 2000, Foody 2002, Kavzoglu and Mather 2003, Waske and Braun 2009,
Kavzoglu and Colkesen 2009). Many non-parametric classification algorithms have been
applied successfully in remote sensing applications, such as Neural Network, Support Vector
Machine, Self-Organising Map, and Decision Trees (Lu and Weng 2007, Waske and
Benediktsson 2007, Salah et al. 2009).
2.1.3.1. Artificial neural network (ANN)
One kind of artificial intelligence technique that has been used for automatic classification is
the Artificial Neural Network (ANN), or Neural Network. This is considered an alternative to
the classical statistical classification methods (Lloyd et al. 2004, Paola and Schowengerdt
1995). The Multi Layer Perception (MLP) model using the Back Propagation (BP) algorithm
is the most well-known and commonly used ANN classifier (Dixon and Candade 2008,
10
Kavzoglu and Mather 2003, Tso and Mather 2009, Ban and Wu 2005). In this type of ANN
model, three layers or more are used with their processing elements as neurons or units. The
input layer supplies input values to all neurons in the next layer. The output layer is the last
processing layer in the network (Lloyd et al. 2004). For land cover classification, the input
variables are often spectral bands or other attributes such as textural information. Layers
between input and output layers are hidden layers. The classification includes three steps,
namely training, allocation and testing. The values of pixels and their known land cover
classes are introduced in the training process. The network weights are adjusted to minimise
error based on measuring the difference between the network generated and real output.
There are forward and backward processes in the back-propagation algorithm. In the forward
process, input training data are supplied to the network, while the weights connecting
network units are set randomly. The network output, which was generated using an activated
function, is then compared with a target and an error is calculated. In the backward process,
this error is fed backward through the network towards the input layer to modify the weights
of the connections in the previous layer in proportion to the error. This process is repeated
iteratively until the total error in the system decreases to a pre-defined level or when a pre-
defined number of iterations are reached (Lloyd et al. 2004, Kavzoglu and Mather 2003). The
most commonly used activation function are sigmoid and tanh functions:
xexsigmoidy −+==
11)( (2.1)
xx
xx
eeeexy −
−
+−
== )tanh( (2.2)
Generally, the three layers (input, hidden and output) architecture is considered appropriate
for remote sensing applications (Foody 1999). Figure 2.2 represents the three layer MLP-BP
network applied for classifying remote sensing data.
11
Figure 2.2: Artificial Neural Network classifier (source: Foody 1999)
For instance, the MLP consists of three layers input, hidden and output with the number of
neuron are Ni, Nh and No, respectively. Firstly the training sample xp=[xp1, xp
2, …xpn] was
presented to the ith neuron of the input layer. As the data propagates to the jth neuron of the
hidden layer, the value at this neuron is computed by:
jN
iiij
j
N
iiijsigmoidj
i
i
xwxwfh θθ −
⎥⎦
⎤⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛−+
=−⎟⎟⎠
⎞⎜⎜⎝
⎛=
∑∑
=
=
1
1 exp1
1 (2.3)
The values of the kth output neuron is calculated by:
kN
jjjk
k
N
jjjksigmoidk
h
h
hhfO θ
ν
θν −
⎥⎥⎦
⎤
⎢⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛−+
=−⎟⎟⎠
⎞⎜⎜⎝
⎛=
∑∑
=
=
1
1exp1
1 (2.4)
where i = 1, 2, …,Ni; j = 0, 1, 2, …, Nh; k = 1, 2, … No; θj and θk are biases associated with
the jth and kth neuron in the hidden and output layer, respectively. The weight connects
the ith neuron in the input layer to the jth neuron in the next (hidden) layer, while the weight
connects the jth hidden neuron to the kth output neuron.
In the backward propagation, weights between neurons are adjusted in order to minimise the
classification errors of the network. In general the least mean square error is used as the
criterion for this process (Tso and Mather, 2009):
2
,)(
21)( kp
pkkp dwE −= ∑ γ (2.5)
where is the value at the kth neuron in the output layer produced by the pth training
sample, and is real target output at the kth output neuron for the pth training sample. The
12
gradient descent method is often used to resolve the minimisation problem of E(w) and the
weights are then adjusted according to:
jkjkjk htt ηδνν +=+ )()1( (2.6)
pijijij xtwtw ηδ+=+ )()1( (2.7)
where w (t) and ν t are weight values at time t. The parameter η is in a range of 0 ≤ η ≤ 1,
and represents the learning rate. and are error value at the kth output neuron and jth
hidden neuron and calculated using:
))(1( kkkkk ydyy −−=δ (2.8)
jkk kjjj hh νδδ ∑−= )1( (2.9)
where is the value calculated by the network at the output neuron kth and is its desired
output. The momentum term α can be applied to make the learning process faster, for
example, the weight connecting the ith input neuron to the jth hidden neuron is modified as:
( ))1()()()1( −−++=+ twtwxtwtw ijijpijijij αηδ (2.10)
where 0 ≤ α ≤ 1.
The ANN has several advantages such as non-parametric base, easy adaptation to different
types of data and input structure. Consequently, it is a promising techniques for land cover
classification (Paola and Schowengerdt 1995, Huang et.al. 2002, Bruzzone et al. 2004, Lloyd
et al. 2004, Ban and Wu 2005, Berberoglu et al. 2007, Aitkenhead et al. 2008, Kavzoglu
2009, Tso and Mather 2009, Pacifici et al. 2009).
Selection of an appropriate network model and parameter values has a major influence on the
performance of the ANN classification process. The important factors are: number of input
neurons, hidden layers and neurons, and output neurons, initial weight range, activation
function, learning rate and momentum, and number of iterations (Zhou and Yang 2008).
Many heuristics have been proposed by researchers to find the optimum parameters for ANN
classifiers, however, none of them are officially accepted (Kavzoglu and Mather 2003). In the
study of Zhou and Yang (2008), 59 ANN models using different parameter values have been
examined for classifying Landsat TM images in the Atlanta metropolitan area, Georgia, USA.
It was found that activation function, learning rate, momentum and number of iterations
strongly influenced classification accuracy. They also reported that using a large number of
hidden layers did not provide a significant increase in classification accuracy.
13
Ban and Wu (2005) compared ANN and ML classifiers for classifying land use/cover
features in the rural-urban fringe of the Greater Toronto area, Canada using multi-temporal
Radarsat images. They found that the ANN classification of filtered SAR images improved
classification accuracies by 4-6% compared to ML. The authors claimed that ANN is more
robust than ML as ANN can minimise conflict of similar signatures between images while
extracting useful information from them. Bruzzone et al. (2004) developed a system for
automatic classification of multi-temporal SAR images based on the ANN classifier with
Radical Basis Function (RBF) as an activated function. The authors claimed the RBF neural
network is a very effective classifier for classifying SAR data. Pradhan et al. (2010)
compared the performance of an ANN classifier using a BP algorithm and an improved k-
Mean unsupervised classifier on land cover classification of the Eastern Himalayan State of
Sikkim. The authors concluded that the ANN-BP classifier produced the best result for
almost all the classes except for the thick forest region, where there was some confusion
between thin vegetation and dense forest. The overall classification accuracy obtained by
ANN-BP model was 90.70%. Heinl et al. (2009) investigated the performance of ANN and
ML classifiers and discriminant analysis (DA) for classifying Landsat 7 ETM+ multispectral
image in combination with ancillary data, including elevation, slope, aspect, sun elevation
angles and NDVI. The MLC and DA gave similar overall accuracy, in a range of 55-60% and
about 75% using only spectral data (Landsat 7 ETM+) and when ancillary data were
incorporated, respectively. The ANN classifier outperformed MLC and DA methods in all
cases with an overall accuracy of about 75% for using only Landsat7 ETM+ images, 85% for
ancillary data, and 86.3% for a combination of multispectral and ancillary data. However, the
authors claimed that while use of ancillary data allowed significant improvement in
classification accuracy, these increases of overall accuracy were observed independent of the
classifier. They went on to suggest that a greater consideration should be given to searching
for appropriate and optimised sets of input variables. The application of an ANN-MLP
network for paddy-field classification using moderate resolution imaging spectroradiometer
(MODIS) data was studied by Yamaguchi et al. (2010). The study showed the effectiveness
of the MODIS data and MLP classifier for mapping paddy-field. The overall accuracy
obtained was 90.8%, which is considered a good result.
Apart from the ANN-MLP network which has been used widely for image classifications,
there are also a few neural network algorithms that have been developed and introduced for
14
classifying remote sensing data, such as Furzy Artmap (Bruzzone et al. 2004, Gao 2009) and
Self-Organising Map (Hugo et al. 2006, Salah et al. 2009).
Results of the studies mentioned above illustrated that the ANN-based algorithm is more
robust and suitable for classifying remote sensing data than the traditional parametric ML
algorithm. However, the application of this technique for the classification of multisource
data has not been deeply studied in the literature.
2.1.3.2. Support vector machine (SVM)
SVM is a recent development of a non-parametric supervised classification technique, which
have proven to be very robust and reliable in the field of machine learning and pattern
recognition (Pal and Mather 2002, Waske and Benediktsson 2007, Anthony and Ruther 2007,
Safri and Ramle 2009, Kavzoglu and Colkesen 2009). SVM separates two classes by
determining an optimal hyper-plane that maximises the margin between these classes in a
multi-dimensional feature space (Kavzoglu and Colkesen 2009). Only the nearest training
samples – namely ‘support vectors’ in the training datasets (Figure 2.3) – are used to
determine the optimal hyper-plane. As the algorithm only considers samples close to the class
boundary it works well with small training sets, even when high-dimensional datasets are
being classified.
Figure 2.3: Example of linear support vector machine (adapted from Mountrakis et al. 2011)
15
As in a case of a binary classification, in n-dimensional feature space, xi is a training set of m
samples, i=1,2,…,m, and their class labels yi = -1 or +1. The optimum separation plane is
defined as:
,1. −≤+bxw i as x belong to class -1 (2.11)
,1. +≥+bxw i as x belong to class +1 (2.12)
or [ ] 1. ≥+bxwy ii i∀ (2.13)
In practice classes are not always fully separated by linear boundaries. Thus, the error
variable iξ is introduced:
[ ] ,1. iii bxwy ξ−≥+ 0≥iξ (2.14)
The optimum hyper-plane is identified based on solving the optimisation problem:
( ) ⎥
⎦
⎤⎢⎣
⎡+ ∑
=
m
iiCw
1
221min ξ (2.15)
where C is the penalty parameter according to the error ξi .
For nonlinear classification the SVM projects input data into a higher dimensional space
using a nonlinear vector mapping function φ so that classes become more separable (Anthony
and Ruther 2007, Tso and Mather 2009). This strategy is appropriate for classification of
remote sensing data, which are usually not linearly separable. This concept is illustrated in
Figure 2.4.
In order to reduce the burden of computation, Vapnik (2000) proposed a kernel function
K(x,y), in which: K (xi, yj) = φ(xi)× φ(yj). Using the technique of Lagrange multipliers, the
optimisation problem becomes:
∑∑∑== =
−m
iijijij
m
i
m
ji xxKyy
11 121 ),(min ααα (2.16)
with ∑=
=m
iiiy
10α and Ci ≤≤α0 , i=1, 2, …,m
where αi is the Lagrange multipliers. The Lagrangian has to be minimised with respect to w, b
and maximised with respect to αi ≥ 0.
Major kernel functions are the Gaussian Radial Basis Function (RBF), linear, polynomial and
sigmoid functions:
16
Linear yxyxK .),( = (2.17)
RBF ( )2exp),( yxyxK −−= γ (2.18)
Polynomial ( )( )dyxyxK 1,),( += (2.19)
Sigmoidal ( )( )1.tanh),( += yxkyxK (2.20)
where x, y represent training samples and their class labels, respectively, in a feature space,
and γ is the width of a kernel.
The final decision function is defined as:
⎟⎠
⎞⎜⎝
⎛+= ∑
=
m
iiii bxxKyxf
1),(sign)( α (2.21)
The above theory was developed for separating only two classes. In the presence of multi-
classes, several strategies have been proposed to apply SVMs. The most common approaches
are one-against-all (OAA) and one-against-one (OAO). Let us assume there are N classes. In
the one-against-all strategy, a set of N binary SVM classifiers, each trained to separate one
class from the rest, is applied. The pixel will be labelled with the class in which the pixel has
the maximum decision value. On the other hand, in the one-against-one strategy N(N -1)/2
SVMs are constructed for each pair of classes. Each classifier votes to one class, and the pixel
will be assigned to the class with the most votes.
Figure 2.4: SVMs projecting the training data to higher dimensional space (source:
Markowetz 2003).
17
In SVMs, the problem of over-fitting in high-dimensional feature space is controlled by the
structure risk minimisation principle (Osuma 2005, Anthony and Ruther 2007, Mountrakis et
al. 2011). The SVMs have been applied successfully in many studies using remotely sensed
imagery. In these studies the SVMs often provided better (or at least no worse) level of
accuracy than other classifiers (Huang et al. 2002, Pal and Mather 2005, Waske and
Benediktsson 2007, Kavzoglu and Colkesen 2009, Pacini el al. 2009).
Pal and Mather (2005) compared SVMs, ML and the ANN approach for classifying Landsat
7 ETM+ and hyperspectral (DAIS) data. The results indicate that SVMs obtained higher
classification accuracy than either the ML or ANN classifier. Basili et al. (2008) investigated
the capabilities of SVM classification methods for land cover mapping in the city area of
Rome, Italy, for three periods in 1994, 1996 and 1999, using ERS 1/2 SAR data. The study
made use of different input parameters, including backscattered mean and standard deviation
images, coherence images, GLCM contrast and energy texture measures. The results
indicated that the SVM method is very applicable for land cover classification even in very
complex surfaces such as urban areas. Moreover, in this study the RBF kernel performed
significantly better than the 2nd order polynomial kernel.
Kavzoglu and Colkesen (2009) used Terra ASTER images and SVMs with radial basis and a
polynomial kernel function to classify land cover type in the Gebze District of Turkey. The
performance of SVMs was compared with the ML classifier. Results indicated that SVMs in
most cases outperform the ML algorithm in terms of overall accuracy (by 4%) and individual
classes. It was also found that the radial basis function (RBF) kernel gave higher accuracy
than the polynomial kernel by approximately 2% of overall accuracy.
In the study by Safri and Ramle (2009), SVM classifiers with different kernel functions,
including linear, radial, sigmoid and polynomial, were used to classify a SPOT 5 satellite
image. The performance of SVM classifiers were compared with the Decision Trees (DT).
Results showed that SVMs outperform DT in terms of classification accuracy. The lowest
overall classification accuracy given by SVM classifiers was 73.70% with the linear function
kernel, while the highest accuracy of 76.00 % was obtained by the RBF kernel. The overall
classification accuracy of the DT algorithm was only 68.78%.
18
Selecting the appropriate model is essential for the performance of the SVM classifier. In the
study carried out by Huang et al. (2002), performance of the SVM classifier has been
compared with ANN, DT and ML for land cover classification using Landsat TM images.
Two options for input variables were selected for testing, namely three input variables (Red,
NIR bands and NDVI images) and seven input variables (6 Landsat TM bands and NDVI
images). For the SVM classifier, different kernel functions (linear, RBF, polynomial) with
various relevant parameters were implemented. For the ANN classifier, the three layer model
(input, hidden, output) with a number of hidden neurons equal to one, two and three times
that of the input variables were tested. The classification results revealed that the SVM
classifier outperformed the ML and DT classifiers in most cases. The SVM classifier was
also more accurate than the ANN classifier for the case of using seven input variables.
However, the ANN provided higher classification accuracy when only three input variables
were employed. The possible reason was the limited success of the SVM algorithm in
transforming nonlinear class boundaries in a very low-dimensional space into linear ones in a
high-dimensional space. On the other hand, the complex network structure might allow ANN
to generate complex decision boundaries even with very few variables, and, as a result, have
better comparative performance than the SVM. Huang et al. (2002) also emphasised that
using more input variables resulted in more improvement in classification accuracy than
choosing better classification algorithms or increasing the training data size. Interestingly,
studies by Dixon and Candade (2008) showed that both SVM and ANN classification
produced better accuracy than ML classification, while SVM and ANN classification
produced comparable results.
Similar to the ANN classifier, although the SVM classifier often outperformed the commonly
used ML algorithm in classification of remote sensing data, and is theoretically more
appropriate than the ML algorithm for handling complex datasets, there are few studies using
this technique for classifying multisource data.
2.1.3.3. Feature selection techniques
The feature selection (FS) or feature subset selection (FSS) algorithm is very important for
data processing in general, as well as for classification of remote sensing data. This technique
allows the selection of only relevant, informative variables while removing the least effective,
highly correlated variables to generate the most separable input dataset for the classification
19
process. Therefore, the FS can help to reduce noise and uncertainty within datasets and lead
to improved classification performance. Because of these advantages the FS techniques have
been increasingly used for image classification and have given promising results, particularly
for highly dimensional datasets (Serpico and Bruzzone 2001, Kavzoglu and Mather 2002, Pal
2006, Anthony and Ruther 2007, Bruzzone and Persello 2009, Maghsoudi et al. 2011).
Among of the FS techniques, the Genetic Algorithm (GA) is one of the most efficient
methods, capable of working with a large search space and has more chance to avoid the
problem of local optimal solutions (Huang and Wang 2006, Zhuo et al. 2008). Consequently,
FS based on GA methods is suitable for large datasets such as in the case of multisource
remote sensing data. However, the use of GA techniques for classification of multisource
data has not been adequately researched. Detailed discussion of the concepts and the main FS
techniques, including GA and its application in land cover classification, will be presented in
chapter 4.
2.1.3.4. Multiple classifier systems (MCS)
Recently, the multiple classifier systems (MCS) or classifier ensemble algorithm has been
increasingly used in pattern recognition and remote sensing. This technique generates final
results by combining the output from different classification algorithms (Xie et al. 2006, Du
et al. 2009a). Consequently, the MCS algorithm has the potential to improve the classification
performance by incorporating advantages of various classifiers while reducing the uncertainty
of individual classifiers. The MCS algorithm is often more accurate than the least accurate
constituent classifier (Foody et al. 2007) and in the ideal case should outperform the best
individual classifier (Huang and Lees 2004). For these reasons, the MCS became attractive to
researchers. Many investigations have been carried out using MCS techniques to classify
remote sensing data (Briem et al. 2002, Bruzzone et al. 2004, Foody et al. 2007, Du et al.
2009b, Yan and Shaker 2011). However, there are not many studies using this technique to
classify multisource data. In this thesis, the MCS technique is proposed and evaluated for
classifying multisource remote sensing data. The concepts and applications of MCS for land
cover classification using remote sensing data will be reviewed in detail in Chapter 5.
20
2.2. DATA PRE-PROCESSING
The data pre-processing process, which removes (or at least reduces) the distortions of
satellite imagery, is essential for the success of remote sensing applications such as image
classification. The basic data pre-processing procedures are presented below.
2.2.1. GEOMETRIC CORRECTION
Remote sensing imagery suffers from the geometric distortion caused by different factors
during the image acquisition process. This kind of distortion leads to the displacements of
imaged pixels from their corrected positions. Thus it is necessary to carry out the geometric
correction (or geometric rectification) in order to eliminate or reduce the geometric
distortions to a satisfactory level (Gao 2009). This task is even more important for
multisource data applications since data from different sensors and databases needs to be
precisely transferred to the common reference systems for further integrated analysis. In
general, geometrical errors are considered as systematic or random depending on their nature.
The systematic errors can be predicted and completely eliminated. For example distortion
caused by Earth rotation and curvature, variation in the platform speed, inconsistency in
scanning mirror velocity (optical imagery) or antenna pattern and range spreading loss (radar
imagery) are systematic errors which can be removed prior to image distribution (Richards
and Jia 2005, Gao 2009). On the other hand, random errors are unpredictable and cannot be
removed completely. The image rectification process can only reduce these errors to an
acceptable level. For example, most of errors related to position and orientation (Figure 2.5)
of the sensor are random.
21
Figure 2.5: Effect of platform/sensor position and orientation on geometry of image (Source:
Gao 2009)
The image rectification process suppresses geometrical errors and transforms the imagery to
the reference system, which could be another image (base image) or preferably a cartographic
projection system such as Universal Tranverse Mecator (UTM). The transform model, which
represents relationship between image pixels and the corresponding coordinates in a
reference system, is determined using ground control points (GCPs). The GCPs must be
stable and clearly visible in both image and the reference system. Many geometrical models
have been applied for transformation of local image coordinate system to those of a reference
system. They can be categorised as either non-physical or physical. In the non-physical
22
approach, the transformation model is assumed to be a certain function, such as a polynomial
function. This method is simple and very flexible since it can be applied to any kind of
image. This method can give a very good result in areas of flat terrain. However, this model
is not rigorous and therefore the accuracy of geometric correction is not high in the case of
complex terrain such as mountainous areas. Moreover, this method cannot account for the
relief displacement (Figure 2.6) caused by the height of terrain. In contrast, the method using
a physical model, such as the co-linearity equation, precisely represents the geometry
between the two coordinate systems. The advantages of this method are robustness and high
accuracy. In particular, using a physical model allows ortho-rectification of image, in which
the relief displacement can be reduced by introducing data from a DEM. The limitations of
this method are complexity, time consuming, and not being available in many commercial
software. In addition, this method requires some sensor-specific parameters that can only
provided by the satellite vendors.
Figure 2.6: Relief displacement in optical and radar satellite image (D is a displacement, Δh
is height of points, η and 90o- η are viewing angles) (source: CCRS/CCT)
2.2.2. RADIOMETRIC CALIBRATION
Optical imagery
The radiances received at the sensors are often converted to digital number (DN) values.
However, in many cases the DN is not adequate for representing surface features, particularly
23
when multi-temporal and multi-sensor data are employed. Thus it is preferable to convert the
DN to radiance values. The sensor calibration is described by:
βαλ += DNL . (2.22)
where is spectral radiance, DN is a digital number, α is gain and β is an offset of the
sensor.
The spectral radiance is dependent on the illumination conditions, such as season, time of
day, geographical position, etc. Thus, it is more convenience to convert the radiance to a
reflectance value, which represents the ratio of radiance to irradiance. The reflectance is
unitless and allows direct comparison between images. For example the reflectance
(exoatmospheric reflectance) of Landsat TM images is presented by:
sESUN
dLθ
πρ
λ
λ
cos... 2
= (2.23)
where Lλ is the spectral radiance, d is the Earth-Sun distance in astronomical units, ESUNλ is
the mean solar exoatmospheric irradiance, and θs is the solar zenith angle in degrees. ESUNλ
is derived from tables provided in the Landsat Technical Notes (August 1986).
SAR imagery
The radar backscattering coefficient (Sigma Nought, σ0) and the radar brightness (Gamma
Nought γ), which represent the total return energy from the surface area within the projected
pixel, is given by (Absolute calibration of ASAR Level 1 products, 2004):
);sin( ,
2,0
, jiji
ji KDN
ασ = )cos( ,
0,
,ji
jiji α
σγ = (2.24)
for i = 1… L and j = 1….M.
where σi,j0 is Sigma Nought at image line and column “i,j”
DNi,j2 is the pixel intensity for pixel i, j
K is the absolute calibration constant
αi,j is incidence angle at image line and column “i,j”
γi,j is Gamma Nought at image line and column “i,j”
L,M are the number of image lines and columns
For rough terrain the local incident angle will be adjusted by the slope angle:
αi,j = β – θ
24
where β is look angle of the radar system
θ is local slope angle
2.2.3. SPECKLE NOISE FILTERING
The speckle noise is a natural and commonly seen phenomenon in radar imaging systems.
The radar system transmits in-phase waves to the target and receives backscattered signals.
After interacting with the target surface, these waves are no longer in-phase because of
differences in scattering properties or due to the distances to the targets. These out-of-phase
radar waves can interact in constructive or destructive ways to produce light and dark pixels,
which are the so-called speckle noise (Qiu et al. 2004). Radar speckle noise is often modelled
as a multiplicative process. This means that the noise increases with an increase in signal
strength. Speckle noise strongly affects the quality of SAR images and therefore must be
reduced before using the data for analysis.
Among of the noise suppression techniques the adaptive filters are considered the most
effective and appropriate for processing SAR data. The adaptive filter reduces the noise while
preserving image sharpness. The most commonly used adaptive filters are the Lee & Lee-
sigma, Gamma-Map and Frost filters.
Lee-Sigma and Lee Filters: In these filters Gaussian distribution for the noise in the image
data is assumed. The Lee-Sigma and Lee filters compute the value of the filtered pixel based
on the statistical distribution of the DN values within the moving window. The Lee filter
considers that the mean and variance of the target pixel is equal to the local mean and
variance of all surrounding pixels within the moving window (Lee, 1981):
DNout = [Mean] + K[DNin - Mean] (2.25)
where Mean = average of pixels in a moving window
)(][)(
22 xVarMeanxVarK+
=σ
(2.26)
and
[ ] [ ][ ]
[ ]22
2
1)( windowMeanwithin
SigmawindowMeanwithinthinwindowVariancewixVar −⎟⎟
⎠
⎞⎜⎜⎝
⎛
++
= (2.27)
25
The Sigma filter is based on the sigma probability of a Gaussian distribution. It is assumed
that 95.5% of random samples are within a two standard deviation range. Firstly, the sigma
(i.e., standard deviation) of the entire image is computed. Then the target pixel is replaced by
the average of all intensity values of neighbouring pixels in the moving window that fall
within the designated range (Lee, 1983).
Gamma-MAP Filter: The Maximum A Posteriori (MAP) filter assumes that the original
intensity value of the pixel lies between its actual value and the average intensity values of
the pixels in the moving window. This filter uses a Gamma distribution instead of a Gaussian
distribution. The formula for the Gamma-Map filter is (Frost et al., 1982):
0)(23 =−+−∧∧−∧
DNIIII σ (2.28)
where: Î = sought value; I = local mean; DN = input value; σ = original image variance.
Frost filter: The Frost filter computes the new value of the target pixel based on a weighted
sum of the values within the moving window (Frost et al. 1981, 1982). The weighting
coefficients decrease with distance from the pixel of interest, and increase for the central
pixels as variance within the window increases (Qiu et al. 2004):
∑ −=nxn
tfrost eKDN ||αα (2.29)
where )/)(/4( 222 −= In σϖα (2.30)
and K = normalisation constant; I = local mean; σ =local variance; ϖ = image coefficient
of variation value; |t| = |X-X0| + |Y-Y0| is distance from the centre pixel to its neighbours and
n = moving window size (Lopes et al. 1990).
Enhanced Lee and Enhanced Frost filters: The Enhanced Lee and Enhanced Frost filter
were developed by Lopes et al. (1990). These filters are an adaptation of the Lee and Frost
filters and therefore use local statistics (coefficient of variation) within individual moving
windows in a similar way. These filters reduce speckle in radar imagery while preserving
texture information. Each pixel is assigned to one of three categories, including
homogeneous, heterogeneous or point target. For the Enhance Lee filter, if the pixel is
considered homogeneous, its value is replaced by the average value of the moving window.
In the case of heterogeneous, the new pixel value is equal to a weight average value. For the
26
point target category the pixel value is not changed. The Enhanced Frost filter treats the pixel
the same as the Enhanced Lee filter for the homogeneous and point target cases. For the
heterogeneous case, the pixel value is determined by an impulse response which is used as a
convolution kernel (Lopes et al. 1990). Recently, the multi-temporal speckle filter has been
developed by Quegan et al. (2000), and has been used effectively in applications such as
forest and rice mapping.
2.3. ADVANTAGES OF MULTISOURCE REMOTE SENSING DATA, SELECTION AND INTEGRATION ISSUES 2.3.1. ADVANTAGES OF THE COMBINED APPROACH
Multisource remote sensing data involve the satellite images themselves, derived products
such as textural information, indices, and ratios, and other additional data from, for example,
geographical information system (GIS) or field-based measurements. The obvious advantage
of a multiple data source is that various sources of complimentary information can contribute
to analysis process.
2.3.1.1. Combination of optical and SAR images
As in a case of integration of SAR and optical images, information in different parts of the
electromagnetic spectrum can be employed. Instead of using only the spectral response
characteristics of features in the visible/near-infrared or backscattered properties in the
microwave region, when they are collected separately, both of them are employed for the
classification process (Forster 1987). Haack (1984) claimed that the best classification could
be obtained by the combination of data from the major portions of the spectrum including
visible, near infrared, thermal and microwave. The visible/near-infrared imagery contains
information about surface response rather than structural details (Forster 1987). This surface
response represents the chemical properties of surface cover such as bare soil, urban areas,
water, and forests. For example, vegetation absorbs incident energy in the red channel for its
photosynthetic process and reflects strongly in the near-infrared region, while water
reflectance is higher for a shorter wavelength (blue) and reduces rapidly at longer
wavelengths (red and near-infrared) (Figure 2.7). The reflectance of dry soil rises uniformly
27
through the visible and near-infrared wavelength ranges, peaking in the middle infrared
range. It shows only minor differences in the middle infrared range due to absorption by clay
minerals.
Figure 2.7: Spectral signatures of soil, vegetation and water in visible/infrared region of
spectrum (source: MicroImage, Inc).
SAR imagery, by contrast, is very sensitive to the physical properties of surfaces, such as
their roughness, texture, shape and dielectric properties (Avery 1992, Charlton and White
2006). Therefore the combined approach will incorporate both physical and chemical
properties of surfaces for classification purposes. For example, major roads, pavement or
other flat concrete surfaces can hardly be separated from other features using visible/near-
infrared data due to their similarity in spectral response. However, by employing surface
physical properties (roughness) of radar data these features are more easily discriminated. On
the other hand, in SAR images it is often difficult to distinguish between different surfaces
which have very similar texture or roughness, such as water, grass or car park. Nevertheless,
these features are very easily distinguished in using visible/near-infrared data.
In vegetated areas, saturation in the spectral value of optical data often occurs while imaging
a complex forest stand structure and accompanying canopy shadows (Steininger 2000, Lu et
al. 2003). Radar sensors with longer wavelength (Figure 2.8) allow penetration through the
28
tree canopy, to a certain level, and give information on structures of vegetation stands (Santos
et al. 2003), and reduce the problem of spectral saturation (Lu and Weng 2007).
Figure 2.8: Microwave bands used in SAR systems (source: Henderson and Lewis 1998)
Forster (1986) compared the responses of Landsat MSS band 5 and band 7 data and SIR-B
SAR data over an urban area of Sydney. It was found that there was a high degree of overlap
between natural vegetation classes (forest) and residential areas with some amount of
vegetation such as grass or trees in the visible red and near-infrared bands. The spectral
response values of these two classes still overlapped in the combined data set, but the mean
spectral values were much more separated because the L-band SIR-B data can penetrate
through the tree canopy which covered the residential buildings. According to Forster (1986),
areas which were cleared for development were very similar to the intense urban areas
(commercial and industrial) in the optical image. However, this was not the case for the SAR
data, where the response from soil was low, because of its fairly smooth surface, compared to
the high backscatter from intense, rough urban areas.
Weydahl et al. (1995) also reported that satellite SAR images may add additional information
about hard targets within built-up areas, including building direction, complexity of objects,
materials, when they are used together with optical images. Many bright points in radar
images result from metal objects, roofs of buildings or corner reflectors formed between
29
buildings and the ground. This may, again, provide more information complimentary to
optical data.
2.3.1.2. Uses of multi-temporal data
The ability to collect multi-temporal data is one of the most useful properties of remote
sensing systems. Multi-temporal imagery is a valuable data source for land cover mapping
since it can provide signatures of ground features in the time domain. Many studies have
demonstrated that use of satellite imagery acquired in different seasons leads to improved
classification accuracy, particularly for crop and vegetation cover types (Brisco and Brown
1995, Pierce et al. 1998, Quegan el al. 2000, Agelis et al. 2002, Bruzzone et al. 2004, Chust
et al. 2004, Ban and Wu 2005, Park and Chi 2008, McNairna et al. 2009). McNaima et al.
(2009) evaluated the capabilities of multi-temporal optical and SAR imagery for annual crop
inventories in Canada. They concluded that although there are a variety of cropping systems,
crops can be successfully classified using multi-temporal satellite data. Guerschman et al.
(2003) explored the use of multi-date Landsat TM images for land cover classification in the
south-western part of the Argentine Pampas. It was shown that at least two Landsat TM
scenes from the same growing season are necessary for successfully identifying land cover
types. The acquisition dates of these images should be able to represent the shift between
winter and summer crops. The authors claimed that in order to discriminate winter crops and
rangelands, one spring (September to October) image and one summer (late December to
February) image were sufficient. Additional images acquired in late summer or fall can
improve separability of summer crops or pasture features at a marginal level.
However, it is often difficult to obtain optical multi-temporal data due to its heavy
dependence on weather conditions. Conversely, the capability of obtaining data at any time
under all weather conditions is one of the major advantages of SAR systems, hence they can
provide multi-temporal data on a regular basis. Multi-temporal SAR data permit the temporal
variation of the SAR signal to be taken into account as well as the extraction of additional
information on land cover class, provided that there has been no transition between land
cover classes during the period considered, but only a change in physical conditions (Agelis
et al. 2002). Quegan et al. (2000) investigated the responses of forests in ERS time series
images, and found that the temporal stability of a forest is greater compared with many other
types of land cover, and this kind of information can be used to effectively map forest areas.
30
McNairna et al. (2009), after comparing the performance of multi-temporal SAR with optical
image for crop classification in various sites in Canada, claimed that the multi-temporal SAR
(two ENVISAT/ASAR images) was comparable with the single optical (SPOT) image, while
the combined dataset (two ASAR images + one SPOT image) provided the greatest
classification accuracy. Shang et al. (2008) pointed out that when available, multi-temporal (2
to 3 scenes acquired at different growth stages) optical data are ideal for crop classification.
However, due to cloud and haze interference good optical data are not always obtainable and
a SAR-optical combination offers a good alternative. They also found that when only one
optical image is available, the addition of two ASAR images acquired in VV/HH polarisation
will provide acceptable accuracies. In Blaes et al. (2005) the performance of various multi-
temporal ERS datasets was evaluated. It was found that by increasing the number of images
the classification accuracies were significantly increased. The overall accuracies vary in a
range of 40% to 65%, depending on number of scenes and acquisition dates. Moreover,
single SAR images are often displayed in a grey scale format which causes considerable
difficulties for visual analysis. Uses of multi-temporal can provide data in the form of a
colour composite, and as a result may significantly ease interpretation tasks.
2.3.1.3. Use of multi-frequency data
SAR images can be acquired at different wavelengths (or frequencies) such as P-, L-, C- or
X-bands. These wavelengths allow different degrees of penetrations and sensitivities to
roughness, structures and scale of surface features. Figure 2.9 shows different penetration
capabilities through the tree canopy of L-, C- and X-band SAR signals. The shorter
wavelength (C-, X-bands) SAR with less capability of penetration often reflect information
from the top of surfaces, such as the upper levels of the tree canopy. The longer wavelength
(L-, P-bands) SAR with deeper penetration can provide more information on the layers
beneath (Forster, 1986).
31
Figure 2.9: Penetration of SAR signals at different wavelengths or frequencies (source Jensen
2004)
Uses of single frequency, and or single polarisation, imagery might not provide enough
information for accurate discrimination even when multi-temporal data are exploited. Thus,
integration of multi-frequency SAR images holds the potential of increasing separability
between land cover classes (Shang et al. 2006, 2008).
Lemonie et al. (1994) demonstrated the contributions of multi-frequency radar for increased
agricultural feature separation using AirSAR data. Baronti et al. (1995) implemented an
analysis with three-frequency (P-, L- and C-band) AirSAR data. They found that P-band data
are effective only in discriminating broad classes of agriculture landscapes. The integration of
L- and C-band helps to reveal finer class details. Shin et al. (2007) investigated the
capabilities of L-, C- and X-band SIR-C/X SAR data for mapping deciduous and two
coniferous species in temperate forests in South Korea. They results showed that X-band
SAR data is more useful for mapping tree species than C- or L-band data since the
backscattered signal from the top of the tree canopy carries more information about tree
species. Pierce et al. (1998) also illustrated that to improve forest type classification it is
essential to combine X-band SAR data with C- and L-band SAR data. Shang et al. (2009)
investigated the application of integrated L-, C- and X-band SAR for crop classification in
two study sites in Canada. Results indicate that L-band performed significantly better than C-
band for larger biomass corn crops with accuracies of 83.8% and 61.0%, respectively. The C-
32
band outperformed L-band for lower biomass crops. Integration of multiple frequencies of
SAR data brought significant improvements to crop classification. The combinations of four-
date C-band (Radarsat-2) and five-date X-band (TerraSAR-X) produced the highest overall
classification accuracy of 87.3%. The authors concluded that when multi-temporal, multi-
frequency SAR data are used, satisfactory crop classification (above 85% accuracy) can be
achieved using a SAR-only dataset.
2.3.1.4. Use of SAR polarimetric and SAR interferometric coherence data
One of the advanced properties of SAR systems is the polarisation of radiation. Polarisation
refers to the orientation of the electric field which oscillates in a plane perpendicular to the
direction of radar wave propagation. The common design of radars is to transmit and receive
microwave radar pulses in horizontal (H) or vertical (V) polarizations (Figure 2.10). The
radar systems which transmit and receive the same kind of polarised microwave radiation,
either horizontally polarised (HH) or vertically polarised (VV), are referred to as like-
polarisation systems, while radar systems which transmit and receive differently polarised
signals are referred to as cross-polarised systems (either HV or VH). There are several
transmit/receive polarisation combination modes offered by the radar systems:
- Single-polarisation : HH, VV, HV or VH
- Dual-polarisation: HH and HV, VV and VH, or HH and VV
- Alternating polarisation: HH and HV, alternating with VV and VH
- Fully polarimetric: HH, VV, HV, and VH
Figure 2.10: Radar transmission in vertical and horizontal polarisation (source: CCRS/CCT)
This advanced function of SAR systems could provide more detailed information about the
interaction of SAR signals with surface features and therefore could enhance the land cover
class discrimination. The like-polarised SAR systems are sensitive to the surface scattering
33
mechanism, in which the radar pulse strikes and backscatters from rough surfaces such as
rocks, bare ground and built-up areas. In this scattering mode, the transmitted and
backscattered signals have the same polarisation. In other words, there are no, or little de-
polarisation effects in the backscattered energies. The cross-polarised SAR systems, on the
other hand, are more sensitive to the volume scattering mechanism, where radar signals are
de-polarised by surface materials and the polarisation of returned signals are changed.
Volume scattering often occurs when the radar signal strikes tree canopy, whereby the radar
pulses interact with various tree components such as leaves, branches, and so on, which cause
de-polarisation of the radar signal (Figure 2.11). Hence, synergistic use of both like- and
cross-polarised data can lead to an enhancement of the separability of land cover classes.
Figure 2.11: Surface scattering and volume scattering (source MicroImage, Inc).
Park and Chi (2008) have integrated multi-temporal/multi-polarisation C-band SAR data to
classify land cover in the Yedang plain, Korea (HH from Radarsat-1, and both VV and VH
from ENVISAT/ASAR). Results of this study indicated that the use of multi-polarisation
SAR data could improve classification accuracy significantly. The authors also made a
further suggestion on integration of multi-temporal/polarisation with optical data for land
cover mapping. Skriver (2008) found that single image with either single or dual-polarisation
did not give sufficient classification performance, while using multi-temporal acquisitions the
classification accuracy improves significantly for both kinds of polarisation data.
Dutra et al. (2009) studied capabilities of ALOS/PALSAR polarimetric data for land cover
classification in the Brazilian Amazon, and found that the best channel combination for
mapping tropical classes, consisting of the primary forest, secondary forest, bare soil,
34
agriculture and degraded forest, is dual-polarisation HH-HV images. However, the authors
pointed out that none of combined PALSAR dataset can discriminate one year regeneration
areas, and recommended the complementary use of optical images when possible, since in the
optical images, the ‘one year regeneration’ class can be differentiated from secondary forest.
Bargiel and Herrmann (2011) used 14 dual-polarised spotlight TerraSAR-X images acquired
during the vegetation season to classify two different agricultural study areas in North
Germany and Southeast Poland. Results showed that in all cases the highest classification
accuracy was achieved when both HH and VV polarisation were employed, and the accuracy
was notably reduced if only one polarisation was used in the classification process.
Interferometric coherence information is also useful to detect land cover features (Takeuchi et
al. 2001, Engdahl and Hyyppa 2003, Srivastava et al. 2006, Park and Chi 2008, Nizalapur et
al. 2011). Interferometric coherence images are generated from a pair of SAR single look
complex (SLC) images using both amplitude and phase information, in which one image is
considered as the ‘master’ and the other as the ‘slave’. Coherence between the two SAR SLC
images is calculated from:
2
22
1
*21
)(.)(.
)().(
∑∑∑
=xsxs
xsxsγ (2.31)
where s1 and s2 are two complex co-registered images.
In the coherence image, a feature which is rather stable, such as urban areas and native
forests, will have a high value of coherence and appear brighter, while a feature which is
changed over time, such as rice fields, will have low coherence and appear darker. This
information could be also incorporated with original SAR backscattered or optical images to
enhance land cover classification. Takeuchi et al. (2001) claimed that using coherence images
generated from pairs of ERS 1/2 Tandem mode and JERS repeated pass images can easily
identify plantation or deforestation areas based on their higher coherence values than that of
non-deforested forests. Engdahl and Hyyppa (2003) evaluated multi-temporal ERS 1/2
inteferometric data for land cover classification in southern Finland, and found that coherence
images provide more information on surface cover types than the standard backscatter
intensity images. They also claimed that the use of coherence data has greatly improved
classification accuracy and allowed them to detect two forest and two urban classes instead of
only one of each, as in the case of using backscattered data. Bruzzone et al. (2004) used 8
35
ERS complex SAR images, which were acquired within a period from June 1995 to May
1996, and the ANN classifiers for land cover mapping in the area of Bern, Switzerland. They
found that the combination of long-term coherence and backscatter temporal variability
which were generated from the above dataset gave the greatest classification accuracy.
Nizalapur et al. (2011) used a coherence image generated from two ENVISAT/ASAR C-band
images in conjuntion with backscatter difference and mean backscatter data to classify land
cover features in a forested area of Chattisgarh, Central India. The traditional ML classifier
was applied. Results of the study highlighted the potential of SAR data in land cover
classification over this area with an overall accuray of 82.5%, average producer’s accuracy of
83.69% and average user’s accuracy of 81%.
Interestingly, Thiel et al. (2009) monitored a large forested area in Siberia using
ALOS/PALSAR intensity and coherence data in summer and winter. They found that the
synergistic use of coherence with intensity data improved separation, particularly from forest.
For example, recent and old clear cuts are separated, and even fire scars can be potentially
discriminated from recent clear cuts. Temporal decorrelation in the forest-based classes are
mainly because of factors such as plant growth, wind, and changes of soil moisture. The use
of coherence data also enhanced the separability of urban areas from all other land cover
types, since the urban are least subjected to temporal decorrelation. The authors claimed that
by integrating coherence with original backscatter intensity images many classes can be
better discriminated.
2.3.1.5. Uses of spatial information
One of the main limitations of these classical classifiers is that they rely only on the spectral
response or backscattered values of cover types. In a complex ground surface, with many
cover types which have similar spectral properties, it is difficult to achieve high classification
accuracy by applying these classifiers. The improvement is possible if spatial information
(textures and pattern of image features) are taken into account since they can provide
additional information about image characteristics (Pacifici et al. 2009). Textures represent
spatial arrangement and variation of patterns on the Earth’s surface (Sheoral et al. 2009). A
large number of texture measures have been generated, analysed and applied for image
classification (Lu and Weng 2007). The most common texture measures are derived from the
Grey Level Co-occurrence Matrix (GLCM). However, Lloyd et al. (2004) claimed that geo-
36
statistical texture measures allow better classification performance than the GLCM texture
data. The texture measures are often incorporated with spectral data as inputs for the
classification process (Gong and Howarth 1990, Lloyd et al. 2004, Nyoungui et al. 2002,
Narasimha Rao et al. 2002, Puissant et al. 2005, Ban and Wu 2005, Berberoglu et al. 2007,
Tassetti et al. 2010). In general, this approach produces better classification accuracy
compared to the conventional method dealing with the original dataset alone. Gong and
Howarth (1990) pointed out that this classification procedure is particularly effective for
differentiation between rural and urban features at urban-rural fringes. For example, Herold et
al. (2005) applied various textured measures derived from Radarsat images to classify land
cover features in Africa. It was found that using a combination of Radarsat images and
texture data (variance) improved the classification accuracy by 16%, 21% and 4% in three
different areas compared to the case that only original data were used. In order to identify
appropriate textures it is crucial to select texture measure, spectral bands, size of moving
windows and other parameters (Chen et al. 2004, Lloyd et al. 2004). Recently, in the study of
Tassetti et al. (2010), textural information, including GLCM and edge-density measures
generated from IKONOS images, was incorporated with spectral data for classification. The
authors pointed out that the spectral/textural approach gave an overall accuracy of 80.01%
compared to 63.44% of accuracy given by using only spectral bands.
The multi-scale texture approach was studied by Coburn and Roberts (2004). In this study,
first-order (variance) and second-order (GLCM Entropy) texture measures generated from
different window sizes were employed as additional information for forest stand
classification. It was revealed that all of the different texture measures provided
improvements of from 4 to 13% in overall accuracy. The multi-scale image texture approach
gave significant increases of 4 to 8% compared with the use of single band texture measures.
Results of the study also indicated that there is no single window size that could sufficiently
represent the whole range of textural information in the image. Zhang et al. (2007)
emphasised the robustness of multi-scale texture analysis while using GLCM texture
measures derived from different window sizes and high spatial resolution IKONOS imagery
for urban land cover/use classification. According to Zhang et al. (2007), the overall accuracy
of the multi-scale approach was higher than cases of single scale texture and original spectral
data by ~6% and ~11%, respectively. This conclusion agrees with the result of Pacifici et al.
(2009), who applied multi-scale first- and second-order textural information extracted from
QuickBird and WorldView-1 panchromatic images for classifying land use of four different
37
urban areas: Las Vegas, Washington D.C., San Francisco (USA) and Rome (Italy). They
claimed that different asphalt surfaces such as roads, highway and parking lots can be
separated with the multi-scale approach based on differences of textural information content.
This approach also allows distinguishing residential houses, apartment blocks and towers
with high classification accuracy (Kappa coefficient is above 0.9).
However, adding textural information does not always increase the classification accuracy. It
has been claimed that radar texture did not give any improvement, but rather reduced the
overall accuracy for some classes (Sheoran et al., 2009). In Chu and Ge (2010a),
incorporation of textural information extracted from either optical or SAR images did not
give any significant improvement. In Salah et al. (2009) the GLCM entropy generated from
the intensity image did not make any contribution to the final result.
2.3.2. SELECTION AND INTEGRATION OF REMOTELY SENSED DATA AND
FEATURE PARAMETERS
As mentioned in the previous section, nowadays, there is a wide range of satellite imagery
with different spectral, spatial, and temporal properties, including both SAR and optical data,
available to users. On the one hand, this provides more opportunities for users to work with
various kinds of remotely sensed data, and various data compositions over the same area of
interest. On the other hand, it is a difficult task to select appropriate input datasets, processing
methods and parameters. As pointed out by Gamba et al. (2005) the amount of overlapping
datasets available for a given area may be very large and, accordingly, the exploitation of
such overlapping datasets may represent a serious challenge. For land cover classification, it
is crucial to understand strengths and weaknesses of different types of data so as to select the
most suitable dataset (Jensen 2004). In general, scale, image resolution and user requirements
are the most important factors for the selection of datasets. Gamba et al. (2005) also
suggested that the simplest way to select a suitable dataset would be one involving sensors
with similar resolution, while allowing capture of different characteristics of the same
environment, and the typical case could be for SAR and multi-spectral data. Different kinds
of optical and SAR satellite images are summarised in Tables 2.1 and 2.2.
2.3.2.1. The impacts and selection of the scale and spatial resolution of images on land
cover classification
38
The spatial resolution, which is defined as the dimension of the smallest recognizable and
non-divisible element of a digital image (Jensen 2004), plays an important role in land cover
mapping using satellite imagery. Because of a rapid increasing number of sources of remote
sensing data, the selection of the appropriate spatial resolution becomes a crucial task. The
spatial resolution and scale of images are often selected based on the ability to identify and
quantify the required land cover features and the mapping purposes.
At a local level, fine-scale classification process, the high spatial resolution data such as
SPOT 5, IKONOS, QuickBird, WorldView (optical) or TerraSAR-X, Radarsat fine mode, or
COSMO-Skymed (SAR) images should be applied. These kinds of data are suitable to map
small, detail features such as buildings, streets, car parks, trees and tree species. The lower
resolution data, but with larger spatial coverage, such as AVHRR, MODIS, MERIS, SPOT
Vegetation (optical) or SCANSAR should be employed for large regional or global scale
mapping projects. These datasets can identify broad land cover features such as urban-build
up areas, larger water bodies, forest and agricultural lands. On the other hand, medium
resolution data including Landsat TM, +ETM, SPOT 1-4, Aster (optical) and
ENVISAT/ASAR, Radarsat, PALSAR (SAR) should be used for medium or small region
scales (Lu and Weng 2007, Jensen 2004) to map various land cover features such as
residential, industrial areas different kind of water surfaces, forest, crops and pastures.
However, the definition of low or high spatial resolution is rather relative and depended on
the ground features to be mapped. According to Woodcock and Strahler (1987) if the pixel
size is smaller than size of objects the spatial resolution can be considered as high (H) while
it is considered as low (L) if pixels larger than objects to be mapped. At the optimal condition
for image classification the pixel size and the size of objects in the ground should be at the
same range (Woodcock and Harward, 1992; Marceau et al., 1994a; Marceau et al., 1994b).
The selection of appropriate scale or spatial resolution of image is also a function of the
environment, particularly the spatial structure (homogeneous or heterogeneous) of features on
the ground (Woodcock and Strahler 1987, Chen et al. 2004). Despite the fact that the higher
spatial resolution image can identify more detail land cover features, it does not necessarily
provide an increase in the classification accuracy. Markham and Townshend (1981) evaluated
the impact of spatial resolution on the image classification and founded that the accuracy of
classification was affected by two contradict factors. The first factor involved changes in a
proportion of boundary pixels which are often mixed of land features. As spatial resolution
39
becomes higher, the number of mixed pixels failing on the boundary of land cover classes
will decrease. Consequently, confusions between classes will be reduced and the better
classification accuracy will be achieved. On the other hand, the second factor involved the
increase of spectral variance within land-cover classes caused by the finer spatial resolution
of image data. This within-class variance reduces the spectral separability of classes and led
to lower classification accuracy. Therefore, it is necessary to use not only image spectral
characteristics but also other sources of information such as textures, contextual information,
combination of different sensors or object based data to classify land cover features in the
complex environment (Woodcock and Strahler 1987, Chen et al. 1999).
The benefits of using multisource remote sensing data are not confined to any particular scale
or spatial resolution of images. For examples, at a regional scale, Torbick et al. 2011
integrated MODIS and ALOS/PALSAR ScanSAR Wide-Beam (WB1) images to map paddy
rice attributes in the Poyang lake Watershed, Jiangxi province, China. The Landsat 7 ETM+
data was used to mask natural water and wetland areas from other features. It is found that
high overall accuracy of 89% was obtained for the rice paddy map. The authors claimed that
the combining SAR and optical approach coupled with continental-scale acquisition
strategies should be the solution for the large-area, operational agricultural rice mapping.
Thiel et al. 2007 employed multi-date ENVISAT/ASAR Wide Swath Mode (WSM) data and
ENVISAT/MERIS data for large scale monitoring water and vegetation in Africa. It showed
that, the synergistic use of SAR and optical data enhanced the classification performance by
compensating the limitation of the optical images. The uses of medium resolution
multisource data for land cover classification has been carried out successfully by many
researchers (e.g. Chust et al. 2004, Erasmi & Twelve 2009, Kim & Lee 2005, Huang et al.
2007, Sheoran et al. 2009, Chu & Ge 2010a, Ruiz et al. 2010). In a case of high resolution
data the combination approach also gave positive results (Burini et al. 2008, Amarsaikhan
2010, Salah et al. 2009, Pouteau et al. 2010, Makarau et al. 2011, Nordkvist et al. 2012). For
instance, Burini et al. 2007 combined SPOT 5 (optical) and TerraSAR-X data for land cover
classification in Tor Vergata/Frascati, Italy using the ANN algorithm. Amarshaikhan 2010,
evaluated the synergistic use of very high resolution Quickbird and TerraSAR-X images for
classifying urban land cover types in Ulaanbaatar, the capital city of Mongolia. It was
revealed that the ruled-based approach provided better classification accuracy than the
traditional ML algorithm and the information derived from the combination of very high
resolution optical and SAR images is very reliable.
40
The uses of multisource data comprising of multiple resolution data has been proposed and
evaluated by several authors (Chen and Stow 2003, Coburn and Roberts 2004, Zhang et al.
2008) to deal with diversity in sizes of land cover objects in the complex environment. Chen
and Stow 2003 introduced three different methods for selecting and integrating information
from multiple spatial resolutions datasets for classifying land cover features in rural-urban
fringe of Del Mar, San Diego County, California, USA. Five different simulated spatial
resolution of 4m, 8m, 12m, 16m and 20m were derived from the 1-m USGS color infrared
Digital Ortho Quarter Quads (DOQQ) image. The results showed significant improvement in
the classification accuracy when compared with those from single-resolution approaches. The
authors also suggested potential uses of these techniques to employ the combination of high
spatial resolution data (such as Ikonos) with the coarser spatial resolution data (such as
Landsat TM) for land-use and land-cover mapping. In the study of Coburn & Roberts (2003)
the multi-scale texture approach was applied. The multi-scale image texture approach gave
significant increases of 4 to 8% compared with using single-band texture measures.
Table 2.1: Specification of some optical satellite imagery.
Sensors Spectral bands Spatial resolution (m) Country Coverage (km)
SPOT 2 XS (multi-spectral) PAN (panchromatic)
20 10
France 60 x 60
SPOT 4 XS (multi-spectral)
PAN (panchromatic)
20
10 France 60 x 60
SPOT 5
XS (multi-spectral)
PAN (panchromatic)
PAN (panchromatic)
10
5
2,5
France 60 x 60
LANDSAT TM 1,2,3,4,5,6,7 30 USA 180 x 180
LANDSAT ETM+ 1,2,3,4,5,6,7
8 (panchromatic)
30
15 USA 180 x 180
ASTER
1,2,3N,3B
4,5,6,7,8,9
10,11,12,13,14
15
30
90
Japan 60 x 60
IKONOS
Panchromatic
Multi-spectral
Nadir off-nadir
0,82 1
3,2 4
USA 11 x 11
QUICKBIRD Multi-spectral Panchromatic 2,44
0,61 USA 16,5 x 16,5
41
RapidEye
2008
B 0.44-0.51; G 0.51-0.58
R 0.63-0.685
NIR:0.69-0.73; 0.76-0.90
6.5 77
GeoEye-1
2008
B 0.45-0.51; G 0.51-0.58 R 0.655-0.69
NIR: 0.78-0.92
1.65 15x15
WorldView-2
2009
0.40-0.45
B 0.45-0.51; G 0.51-0.58
Y 0.585-0.625; R 0.655-0.69
NIR: 0.705-0.745; 0.860-1.04
1.8 16.4
ResourceSAT-2
2011
G 0.52-0.59; R 0.62-0.68
0.77-0.86
5.8 (LISS-4)
23.5 (LISS-3) 70
Table 2.2: Specification of different SAR satellite imagery
Sensors Lauch Country Band Resolution (m)
ALMAZ-1 1991 Russia S 15 ERS-1 1991 ESA C 26 JERS-1 1992 Japan L 18 ERS-2 1995 ESA C 26
RADARSAT-1 1995 Canada C 10 SRTM 2000 USA Germany Italy C, X 30
ENVISAT-ASAR 2002 ESA C 30 ( Polarisation); 150 (Wide Swath);
1000 (Global Monitoring) ALOS-PALSAR 2006 Japan L 10 ( Polarisation); 100 (ScanSAR)
TERRASAR-X 2007 Germany X 1 TANDEM-X 2009 Germany X 1
RADARSAT-2 2007 Canada C 3
COSMO-SKYMED COSMO-1&2
COSMO-3 COSMO-4
2007
2008
2010
Italy X 1
RISAT 2008 India C 3
2.3.2.2. Selections of parameters for land cover classification
As well as the different kinds of satellite images mentioned above, many parameters are
involved in the land cover classification process, including spectral bands, vegetation indices,
texture measures, multi-temporal parameters, polarisation, frequencies, transformed images
42
and additional data (Lu and Weng 2007). Unfortunately, large data volumes does not always
produce high classification accuracy. Using too many parameters and too much image data
will also increase uncertainty between datasets and may actually reduce classification
accuracy. For example, in the case of ANN classifiers, Kavzoglu and Mather (2002) have
pointed out that a large amount of inputs decreases the generalisation capabilities of the
networks and produces more redundant and irrelevant data. Therefore it is necessary to select
only the parameters or sets of parameters that may significantly contribute to the separation
of land cover classes and which are less correlated to each other (Jensen 2004, Lillesand and
Kieffer 2004, Lu and Weng 2007, Tso and Mather 2009). It is important to apply techniques
such as feature selection, principal component analysis, discriminant analysis, transformed
divergence, Mahalanobis or Jeffrey Matusita distance, wavelet transform or mixture analysis
to minimise overlap or to extract the most useful parameters or type of data (Kavzoglu and
Mather 2002, Jensen 2004, Lillesand and Kieffer 2004, Richards and Jia 2006). It is even
more important in the case of multisource data since the data volume may be very large with
a significant degree of correlation.
Thus, an important research challenge is to select the best dataset, including original satellite
imaging data (optical and SAR, multi-temporal/frequencies/polarisation), derived data
(texture measures, indices, ratio, inteferometric coherence) and auxiliary parameters, for
classification with maximum uncorrelated content, minimum uncertainty and greatest
effectiveness.
2.4. RELATED WORK ON LAND COVER CLASSIFICATION USING MULTISOURCE REMOTE SENSING DATA
One of the first attempts to evaluate the capabilities of multiple remote sensing datasets was
made by Haack (1984). The work was to compare the performance of urban land cover
classification using combined radar and Landsat TM data, and each of them as a single
dataset. The classification process was carried out for eight major land cover types in the
urban region of Los Angeles, USA. The results indicated that the combined approach
produced better classification accuracy than any single dataset. The authors pointed out that
the best selection of a combined dataset is one which uses all major portions of the
electromagnetic spectrum, including visible, near-infrared, infrared and radar.
43
A comparison of the capabilities of optical (Landsat TM and SPOT) and SAR (JERS-1 and
ERS-1) satellite images for mapping tropical land cover was carried out by Hussin (1997)
over an area of 1750 km2 in Central Sumatra, Indonesia. Supervised classification was carried
out for each single dataset and all possible combinations including fused JERS-1+Landsat
TM, ERS-1 +Landsat TM, JERS-1+SPOT, ERS-1+SPOT, JERS-1+ERS-1+Landsat TM and
JERS-1+ERS-1+SPOT. Results of the classification indicate that although fusion of a SPOT
image with JERS-1 and ERS-1 provided a better accuracy than the single SPOT image, it
decreased the accuracy of results of single SAR data (JERS-1 and ERS-1). In contrast, the
combined dataset produced a slight increase in accuracy of Landsat TM data. This implied
that the increase in data volume does not necessarily increase accuracy of land cover
classification.
Kim and Lee (2005) classified land cover features in the Western coastal region of the
Korean peninsula using several combinations of Landsat ETM+ and Radarsat-1 images. The
standard ML classifier was employed for classification of eleven land cover classes. The
overall classification accuracy using combined datasets was improved to 74.6% as compared
to an accuracy of 69.35% using only Landsat ETM+ data.
Nizalapur (2008) analysed the potential of different fusions of SAR and optical data for land
cover classification in the Dandeli forested region, Karnataka, India. Two false colour
composite (FCC) images were generated, the first one from coherence and backscattering
images of ENVISAT/ASAR data (HH polarisation) on 25th September 2006 and 30th October
2006, the second one with ENVISAT/ASAR data (HH polarisation) on 25th September 2006
along with IRS-P6 LISS-III of 11th January 2005 using the intensity hue saturation (IHS)
fusion technique. The ML classification technique was applied separately to each FCC image.
Results suggested composition of coherence information along with backscatter images
enhanced the delineation capabilities of SAR data. The overall classification accuracy and
kappa coefficient of the FCC were observed to be 78% and 0.75, respectively. However, the
more interesting finding was observed by merging the optical LISS-III data with HH-
polarised ASAR data that allowed better delineation of the forest types from other land cover
classes and minimised the shadow effect. The overall classification accuracy and kappa
coefficient of the merged data was 82% and 0.80, respectively. This indicates the significance
of integrating SAR and optical data for better classification of the land cover classes.
44
Chu and Ge (2010b) investigated the classification of various land cover features in the south
of Vietnam using combinations of multi-temporal ALOS/ PALSAR (L-band),
ENVISAT/ASAR (C-band) SAR and SPOT multi-spectral optical satellite data. Results
demonstrated the advantages of the integration approach and clearly highlighted the
complementary characteristics of multisource datasets. The combination of optical and multi-
temporal SAR images gave increases in classification accuracy of 6.45% and 23.13% (SPOT
+ ENVISAT/ASAR) and 10.01% and 29.40% (SPOT + ALOS/PALSAR) in comparison to
the cases of using only single SPOT 4 multi-spectral, ASAR or PALSAR multi-date images.
Ruiz et al. (2010) also reported that the joint utilisation of optical (Landsat ETM+) and SAR
(Radarsat-1) provided useful information on land cover classes and improved classification
accuracy compared with using either type of images on their own.
As mentioned in the previous section, incorporation of texture information can enhance
classification accuracy. This has been demonstrated by Lu et al. (2007), who investigated
incorporation of Landsat ETM+ (multi-spectral + panchromatic) and Radarsat-1 SAR data
and their textures for classification over the Brazilian Amazon area. In this study, wavelet-
merging techniques were used to integrate Landsat ETM+ multi-spectral with Radarsat-1
SAR images. Texture information derived from GLCM based on Landsat ETM+
panchromatic band or Radarsat-1 data and different sizes of moving windows were
examined. Pearson’s correlation analysis was used to analyse the correlation between the
selected textures. The highly separable textures with low correlation coefficients were
selected. Although eight texture measures were derived with three window sizes (9x9, 15x15
and 21x21), the best three measure are mean (ME) with a 15x15 window, and variance (VA)
and second moment (SM) with a 21x21 window generated from panchromatic band, and ME,
VA and contrast (CO) with a 21x21 window based on the Radarsat C-HH band. These
textures were combined with original data for the ML classification process. Results of
classification on different combinations has shown that although classification accuracy from
other combinations did not improve significantly, or even slightly decreased (Landsat ETM+
multi-spectral +Panchromatic or Landsat ETM+ multi-spectral + SAR data) compared to
ETM-ALL (Landsat ETM+ multispectral), the combination of data fusion and textures from
both panchromatic and Radarsat data improved overall accuracies by 5.8% to 6.9% compared
to ETM-ALL. This study showed that texture measures are an important factor for improving
land cover classification performance. However, selection of suitable textures is a difficult
task since textures vary with the characteristics of the surface landscape and the kind of
45
images to be used. Identifying suitable textures involves the selection of appropriate texture
measures, moving window sizes, and image bands (Chen et al. 2004).
A similar study was performance by Haack and Bechdol (2000) in which multi-sensors
including Landsat TM and Shuttle Imaging Radar (SIR-C, C and L band) were evaluated for
east African landscapes. The main land cover features of interest were settlements, natural
vegetation, and agriculture. The common ML classifier was used for land cover
classification. Three texture algorithms were derived: mean euclidean distance (MED),
variance (VAR), and kurtosis (KRT) from two bands SIR-C data. The texture measures were
examined with eight different window sizes ranging from 3x3 to 17x17. After comparing the
classification accuracy using these texture measures it was concluded that the variance (2nd
order) measure of texture provided the best classification accuracies. The optimum window
size for texture measurement was 13x13. These results showed that the radar data by itself
provides excellent classification accuracies, particularly in combination with texture
measures. Haack and Bechdol (2000) pointed out that some combinations of radar with
optical data (Landsat TM band 4,5 and C-, L-band radar) also increased classification
performance compared to optical data alone, but were not as good as radar data, and L-band
was more useful than C-band for this area.
Huang et al. (2007) investigated the capability of incorporating Landsat ETM+ data with
multi-temporal SAR (Radarsat-1) images to enhance automatic land cover mapping over an
area of St. Louis, Missouri, USA. Two Landsat ETM+ and three Radarsat-1 images acquired
at different dates were used for this study. In addition, the texture measures generated from
SAR data and an ML classifier were applied. Eight texture images were derived (Gaussian
high pass filter (HPF), Gaussian low pass filter (LPF), Laplacian, mean, median, entropy and
data range (DR), and variance), and were applied to the Radarsat-1 data with sizes of moving
window ranging from 3x3 pixels up to 25x25. These data were then added to the Landsat
imagery as additional bands for classification. It was found that integration of multi-sensor
data improves the classification accuracy significantly over a single Landsat image. The
entropy measure consistently provides larger improvements in comparison with the other
seven measures. The integration of entropy derived from a 13x13 window with Landsat
images achieved an increase in classification accuracy of more than 7%. The greatest
accuracy of 84.36% was obtained with a combination of Landsat images and Radarsat-1
feature combinations from entropy, data range and mean generated from 13x13, 9x9 and
46
19x19 window sizes, respectively. The improvement was about 10% compared to the
classification using Landsat data alone.
The capability of integrated Interferometric SAR (InSAR) with ERS-1 and JERS-1 images
and optical SPOT XS data for land cover classification has been evaluated by Amarsaikhan
(2006) using a rule-based classification over an urban area of the city of Ulaanbaatar,
Mongolia. Firstly, seven combination datasets were used, including SPOT XS, coherence
images (generated from a pair of InSAR), JERS-1 and ERS-1 images, six first principal
component (PC-1) images and eight GLCM texture measures generated for standard ML
classification. The results indicated that the combination of SPOT XS multi-spectral and SAR
data (JERS-1, ERS-1) gave an improvement in classification performance. The selected
dataset was then used for initial segmentation with the Mahalanobis distance rule, which is a
modification of the ML classifier. A set of rules were introduced to assign pixels to land
cover classes. Although this is also a kind of parametric classification, the results were
significantly improved, up to 95.16 % of overall accuracy. Amarsaikhan (2006) concluded
that the integrated features of the optical and InSAR images can significantly improve the
classification of land cover types, and that the rule-based classification is a powerful tool for
the production of a reliable land cover map.
The synergy of dual-polarimetric SAR (Envisat/SAR) and optical medium resolution
(Landsat-ETM+) data for land cover classification at the regional level has been evaluated by
Erasmi and Twele (2009) in Central Sulawesi, Indonesia. The main objective of this work
was to identify intensively managed agricultural land (rice crops and cocoa plantations) and
discriminate cropland from forested land. It was expected that the multi-temporal SAR data
would improve the detection and separation of those cropping systems based on the
relationship between the temporal signature of the SAR backscatter signal and the
phenological status of the crops. Moreover, the contribution of dual-polarisation (cross- and
co-polarisation) on classification accuracy was also evaluated. Envisat/ASAR dual-
polarisation images were acquired at six different dates. Accompanying these data channels,
first order (mean, minimum, maximum, standard deviation) and second order (dissimilarity
and homogeneity) statistics were calculated. The SAR index, the Normalised Polarimetric
Difference Index (NPDI) defined as the ratio of the difference to the sum of co-polarised and
cross polarised backscatter values, was also tested. The Jeffries-Matusita Index (JMI)
(Richards and Jia 2006) was used to analyse separability between classes. In this study the
47
object-based classification was used with the nearest neighbour classifier. There were seven
land cover classes selected for classification, including natural forest, cocoa plantations, rice
cropping systems, fallow land/mixed, urban land, water and riverbed. Based on the results
obtained, the authors drew the following conclusions: 1) multi-temporal SAR datasets, in
general, allow differentiation of major land cover features (natural forest, cocoa plantations,
rice cropping systems); 2) multi-temporal SAR data enhance the accuracy of mapping crops
(e.g. rice crops); 3) integration of ASAR with Landsat images increased classification
accuracy significantly, and the combination of like-polarised ASAR time series and Landsat
multi-spectral data produced the best results ; 4) temporal statistical measures do not make a
noticeable contribution to a better separation of rice crops from other land cover classes; and
5) SAR-based texture measures have only minor effects on classification accuracy. The
authors also suggested future studies should investigate the improvement of land cover
classification from the application of satellite-based multi-frequency and full-polarimetric
SAR data that may be obtained in the near future and from the combination of different SAR
platforms (Envisat/ASAR, TerraSAR-X, ALOS-PALSAR).
Haack and Khatiwada (2010) used the Transformed Divergence (TD) separability index to
assess the integration of optical (Landsat TM) and Shuttle Imaging Radar –C (SIR-C) quad-
polarisation imagery for land cover mapping in a study site in Bangladesh. GLCM variance
texture measures with a 7x7 window size were also extracted for both kinds of data. It
revealed that the combination of two bands which gave the highest separability value was the
Landsat TM Mid-infrared (MIR) and the SIR-C L-HV texture. The authors suggested that
fusion of two datasets such as radar and Landsat TM can be valuable for improving the
classification accuracy for land cover/use.
Ruiz et al. (2010) reported that the combined utilisation of optical (Landsat ETM+) and SAR
(Radarsat-1) provided useful information on land cover classes and improved classification
accuracy compared with using either type of original image data on their own. However, in
contrast, Maillard et al. (2008) demonstrated that use of Radarsat-1 and ASTER data did not
improve the classification as compared to the case of using only ASTER data while
classifying wetland and vegetation features in the savanna region of Brazil.
Lehmann et al. (2011) combined Landsat TM and ALOS/PALSAR data for forest monitoring
over a test site in north-eastern Tasmania, Australia. The ML classifier was employed to
48
classify the forest and non-forest classes. The results indicated that the joint use of SAR and
optical data produced a better forest classification than that of either single dataset. The
combination approach gave the highest classification accuracy of 94.32%, with 5.7% and
2.3% improvement in overall accuracy over the PALSAR-only and Landsat-only
classifications, respectively. The authors also claim that HV-polarisation data are the most
useful for separating forest/non-forest in SAR data, and that the contributions of each band of
Landsat TM and PALSAR data to the final results are significantly different.
Recently, Nordkvist et al. (2012) investigated the combination of optical satellite data and
airborne laser scanner data for vegetation classification. In this study, the ML and DT
classifiers were used to classify SPOT 5 and Leica ALS 50-II data from 2009, over a test area
in mid-Sweden. Seven vegetation classes were identified for classification, including clear
cut, young, coniferous ( >15)m and coniferous (5-15m), deciduous, mixed forests and mire. It
was shown that the integration of SPOT and ALS data produced classification accuracy up to
72%, compared with 56% while using only SPOT data. The combined dataset significantly
reduced confusion between classes because of its complementary nature. While the SPOT 5
image was crucial for species separation, the ALS data was useful for distinguishing between
height classes which have similar spectral characteristics. The authors claimed that
integrating the large area laser scanning dataset could lead to significant improvement in
vegetation classification based on satellite imagery.
In Waske and Benediksson (2007), multi-temporal SAR data, including ENVISAT/ASAR
and ERS-2 images acquired from April to August 2005 and a Landsat 5 TM image on 28
May 2005, were employed for mapping land cover features in agricultural areas in Germany.
SAR and multi-spectral data were classified separately using the SVM classifiers. The final
results were obtained by fused outputs of these initial classification processes. The
combinations were carried out by another SVM classifier and several commonly used voting
methods, namely, majority voting and absolute maximum. Eight land cover classes were
determined for classification, including arable crops, cereals, forest, grassland, orchard,
canola, root crops and urban. Results illustrated that the approach based on combining
outputs of SVM classifications on single-source data performed better than all other single
classification algorithms. The fusion SVM approach gave an increase of 2% in overall
accuracy as compared to the SVM classifier which was the best single classifier and trained
by the whole combined multisource dataset. Use of multisource data has led to a significant
49
improvement in all classes and overall classification accuracy. The SVM classification on
multisource data resulted in an increase of total accuracy by up to 11.5% and 6% when
compared to accuracy obtained by classifying SAR and optical data, respectively. However,
the improvements of classification accuracy were observed also for other classifiers, such as
ML and DT.
A similar approach was used by Pouteau et al. (2010), who used three different information
sources: optical very high resolution IKONOS satellite image, SAR images from NASA
PACRIM II AirSAR mission and DEM data; for mapping plant species in a mountainous
tropical region in the Marquesas archipelago, French Polynesia. The three sources of
information were classified separately by the SVM algorithm and then combined by the SVM
fusion method. The fusion approach achieved an overall accuracy of 70%, which is
considered fairly good for such a complex study area, while the single data source provided
very poor results of 54%, 20% and 30% for IKONOS image, AirSAR and DEM data,
respectively.
Salah et al. (2009) used a fusion of lidar data, multi-spectral aerial images and 22 auxiliary
attributes, including texture strength, GLCM homogeneity and entropy, NDVI and standard
deviation of elevations and slope, for building detection in the area of the University of New
South Wales campus, New South Wales Australia. The SOM classifier was employed for this
study. The results show that the combination of lidar and aerial photography gave an
improvement of accuracy by 38% compared with using only an aerial image, and using
derived attributes further improved the result by 10%.
2.5. SUMMARY
Based on the discussion in Sections 2.1.3, 2.3.1, and 2.4, the following conclusions can be
drawn:
- The use of multisource remote sensing data, including various combinations of multi-
temporal/frequency/polarimetric or interferometric coherence SAR data and optical
satellite imagery, is an approach of high potential for land cover classification. This
approach can incorporate the advantages of different types of spatial information and
therefore improve the classification performance. Previous studies have shown that, in
50
general, use of multisource image data often produces better classification accuracy than
any single dataset. However, in some cases, the integrated datasets do not give an
increase in the classification accuracy, particularly when the ML classifier was applied.
This emphasises the need for selecting relevant input features and appropriate
classification techniques for classifying multisource data.
- In spite of its parametric nature and disadvantages in integrating different kinds of data,
the traditional ML is still the most commonly used classifier (Lu et al. 2011), including
for multisource data. Unlike the parametric classifier, the non-parametric classifiers such
as ANN and SVM are advanced techniques based on the principles of machine learning,
which do not require a normal distribution assumption of input data and therefore are
theoretically more suitable than the traditional statistical classifiers for handling complex
datasets. These classifiers have been applied successfully for the classification of single-
type datasets such as optical or SAR images. However, to date, there has not been much
work carried out on applying these classifiers for classifying multisource data.
Consequently, this approach will be evaluated in this thesis.
- In the studies mentioned above, the combined datasets were often generated by adding
different kinds of data together. This approach does not guarantee that one of the
combined datasets will give the best classification performance. The difficult tasks of
how to select relevant input data, optimal, or nearly-optimal combined datasets and
classifier parameters that could provide the highest classification accuracy have not yet
been considered. This becomes even more challenging for the cases of multisource
datasets where the volume and dimension of datasets could be extremely large, with a
high level of diversity (spectral, multi-temporal, polarisation, frequencies). Therefore, this
subject is one of the main targets of this research. In this thesis, the author proposes and
evaluates FS techniques, in particular the GA, to resolve these tasks. These techniques
have been applied successfully in a number of remote sensing applications, including
several studies on classification of multispectral and hyperspectral data. Although they
are considered to be effective tools to reduce data dimensions, their applications for
handling multisource data have previously been limited.
- The other powerful technique which has also been developed and increasingly used for
image classification, but not well addressed for classifying multisource remote sensing
51
data, is MCS or classifier ensemble. The MCS is capable of taking advantage of different
classifiers and compensating for their weaknesses, and therefore enhancing classification
accuracy. Thus, one of the objectives of this thesis is to utilise the MCS to improve the
classification of multisource remote sensing data, in which various classifiers and
combination techniques will be applied and evaluated.
- Finally, although both the MCS and GA based FS techniques are powerful and have been
used in classification of remote sensing data the integration of these techniques has not
been considered by researchers. Hence, one of the main contributions of this thesis is to
propose and evaluate the integration of these two techniques for classifying multisource
data. This newly developed method holds the potential to improve classification
performance by exploiting the strengths of both techniques.
52
CHAPTER 3
EVALUATION OF LAND COVER CLASSIFICATION
USING NON-PARAMETRIC CLASSIFIERS AND
MULTISOUCE REMOTE SENSING DATA
As has been mentioned earlier, although parametric classifiers such as Maximum Likelihood
(ML) are not suitable for handling complex datasets due to their assumption of a normal
distribution, they are still the most commonly used algorithm and there are many studies
using this kind of classifier for multisource remote sensing data (Nizalapur 2008, Erasmi and
Twele 2009, Ruiz et al. 2010, Lehmann et al. 2011, Lu et al. 2011, Nordkvist et al. 2012,
Wang et al. 2012). On the other hand, non-parametric classifiers, such as Artificial Neural
Network (ANN) or Support Vector Machine (SVM), have prominent advantages for
incorporating different type of data, but there are not many applications of these classifiers in
practice. This chapter will evaluate the performance of non-parametric classifiers for land
cover classification using different integrated datasets in study areas in Vietnam and
Australia.
3.1. APPLICATION OF MULTI-TEMPORAL/POLARISED SAR AND OPTICAL IMAGERY FOR LAND COVER CLASSIFICATION
The major objective of this study was to investigate the potential of using multisource remote
sensing data comprised of multi-temporal, dual-polarised SAR (ALOS/PALSAR) data in
combination with optical multi-spectral (SPOT 2 XS) satellite images for land cover
classification in southern Vietnam by using the non-parametric (ANN and SVM)
classification techniques. The classification process was carried out for different single-type
and combination datasets.
53
3.1.1. STUDYAREA AND DATA USED
The study area is located along the Saigon River in southern Vietnam, centred at the
coordinate 106o 46 E; 10o 48’ 30’’ N. The area covers a part of Ho Chi Minh City, and a
small part of the Dong Nai province (Figure 3.1). The terrain is very flat with the main land
cover features being water surface, vegetation, bare land and urban areas.
The SPOT 2 multi-spectral (SPOT 2 XS) imagery has a spatial resolution of 20m for two
visible (Green, Red) and one near-infrared spectral bands. The SPOT 2 XS image used for
this study was acquired on January 10, 2008. Five ALOS/PALSAR dual-polarised (HH/HV)
images acquired for the period from June 08, 2007 to June 10, 2008, were used for this study
(Table 3.1).
Table 3.1: ALOS/PALSAR images for the study area.
Satellite/Sensor Track -
Frame
Acquisition
dates
Polarisation Orbit Spatial
Resolution
ALOS/PALSAR 477_20
08 Jun 2007 HH/HV Ascending 12.5 m 08 Sep 2007 HH/HV Ascending 12.5 m 09 Dec 2007 HH/HV Ascending 12.5 m March 2008 HH/HV Ascending 12.5 m 10 Jun 2008 HH/HV Ascending 12.5 m
Figure 3.1: Location of the study area.
54
3.1.2. METHODOLOGY
ALOS/PALSAR images was geo-rectified to the map coordinates (WGS84, UTM projection,
zone 48) using the in-house developed software from the GEOS group at the University of
New South Wales. The software employed the physical modeling and used the SRTM-DEM
to create simulated SAR images and remove the distortion of images due to terrain
displacement. These geometrically corrected images from multiple dates were perfectly
matched with maximum errors less than 0.6 pixels.
The optical and SAR images were re-sampled to same spatial resolution (12.5m), so that they
can be easily registered and added together to generated combined datasets for classification
in a later step. In order to preserve the original pixel’s values the Nearest Neighbour
algorithm was employed for the re-sampling process. The SPOT 2 XS image was registered
to SAR images manually using the ENVI software version 4.6 with the image-to-image
procedure using Ground Control Points (GCPs). Since the difference of geometry of SAR and
optical images the GCPs had been selected carefully with the reference of Google Earth
images. Only features which can be clearly indentified in both kinds of images with relatively
flat surrounding area were selected as the GCPs. For examples, the cross of relatively small
roads or channels, or small isolate features such as houses. In order to avoid the image
distortion caused by the warping process only the first order polynomial function (linear
function) was used. The GCPs were distributed evenly over the entire study area. The testing
process was carried out after the image registration using test points which are different from
the GCPs to ensure that the maximum error of matching of the two images is less than 1
pixel.
The adaptive Enhanced Lee filter with window size of 3x3 was applied to remove speckle
noise from the SAR images. PALSAR backscatter values were converted to decibel (db):
Db = 10 * log10(DN2) + CF (3.1)
where Db, DN are magnitude values and CF is the Calibration Factor provided by the Japan
Aerospace Exploration Agency (JAXA).
Ten interferometric coherence images have been generated from five PALSAR HH polarised
images. There are four short-term coherence images derived from pairs of SAR single look
complex (SLC) images with 3 months difference in acquisition dates while the other
coherence images were generated from pairs of SLC images with larger differences in
acquisition dates.
55
In this study, the textured data generated from multi-date SAR images were employed. The
main objectives of using textural information were to evaluate the possibilities to improve
land cover mapping accuracy by combining multi-date PALSAR images with their textural
information. Grey Level Co-occurrence Matrix (GLCM) texture measures were extracted
from the First Principal Component (PC1) of each five-date PALSAR like- (HH) and cross-
(HV) polarisation dataset. Four GLCM texture measures were employed, namely, variance,
homogeneity, entropy and correlation. Since there is no preferred direction, average of
texture measures generated at eight different directions – 0o, 45o, 90o, 135o, 180o, 225o, 270o,
315o – were computed and integrated with SAR backscatter data for classification. Various
window sizes were tested for texture generation, including 3x3, 5x5, 7x7, 9x9, 11x11 and
13x13. For each window size the incorporation of GLCM texture measures with multi-date
PALSAR images was carried out using single, two, three and all four measures to generate
different combined datasets. These datasets would be used for classification with both SVM
and ANN classifiers. The two sets of textural features which provide the highest accuracy
when integrated with five-date HH and HV images were identified and combined with five-
date dual (HH+HV) images for classification. The GLCM measures are:
GLCM Variance 21
0,,
2 )( i
N
jijii iP μσ −= ∑
−
=
(3.2)
GLCM Homogeneity ∑−
= −+
1
0,2
,
)(1
N
ji
ji
jiP
(3.3)
GLCM Entropy )ln( ,
1
0,, ji
N
jiji PP −×∑
−
= (3.4)
GLCM Correlation ( )( ) ⎥
⎥
⎦
⎤
⎢⎢
⎣
⎡ −−∑−
=22
1
0,,
))((
ji
jiN
jiji
jiP
σσ
μμ (3.5)
where i, j are the pixel’s grey values; Pij is number of the co-occurrence of grey values i and j,
and N is the size of the moving window.
Two non-parametric classifiers, ANN and SVM, were tested in this study. In order to
compare the performance of non-parametric and parametric algorithm for classifying
multisource data, the commonly used ML classifier was also implemented. For the ANN, the
commonly used Multi Layer Perception (ANN-MLP) neural network with a Back
Propagation (BP) algorithm was applied. The networks consisted of three layers, including
56
input, hidden and output layers, since the use of more than three layers does not improve
accuracy significantly (Mills et al. 2006). Each input feature, such as SPOT XS image bands,
was introduced to the network by one input layer neuron. Land cover classes were presented
in the output layer neurons. The hidden layer neurons were automatically structured and
pruned by the ENVI software used for classification (Argany et al. 2006). The sigmoid
function (Equation 2.1), was used as an activated function. The parameters that were used for
the ANN classification processes were: number of iterations: 1000; learning rate: 0.1; training
momentum: 0.9; training RMS exit criteria: 0.1.
The SVM classifier using the Radial Basic Function kernel presented in Equation (2.18) was
selected for this study because of its robustness (Melgani and Bruzzone 2004, Foody and
Mather 2004, Kavzoglu and Colkesen, 2009). The optimal parameters, including γ (width of
a kernel) and C (penalties coefficient) for the SVM classifier, were determined using the grid
search algorithm and cross-validation technique.
Six land cover classes were identified for classification. These classes were: Water Surface
(WF), Bare Land (BL), Dense High Urban structures (DHU), Low Flat Urban structures
(LFU), High Dense Vegetation (HDV), and Low Sparse Vegetation (LSV). The spectral
properties and backscatter signatures of these features in the SPOT 2 multi-spectral and
multi-temporal PALSAR images are shown in Figure 3.2.
Figure 3.2: Land cover feature characteristics in SPOT 2 multi-spectral and ALOS/PALSAR
multi-temporal images.
57
The training and testing data were randomly selected based on manual interpretation with
reference to land use maps created in 2005, field checking and with the aid of multi-date high
resolution images from Google Earth. Since the land use/land cover map in the study area is
out of date and the field visit was carried out on different period from the image acquiring
dates the additional data including cropping calendar or season growing pattern are also
collected for reference. Testing pixels were selected independently from the training data and
used to construct the confusion table. The samples were collected using the Region of Interest
(ROI) tool of the ENVI 4.6 software. To ensure the reliable of training and testing datasets,
only pixels which are well represented land cover features and distant from the edges were
chosen to avoid confusions between classes. For each land cover class the number of training
and testing pixels were roughly equal. There are 5699 training pixels, including 1071 pixels
for WF, 1002 pixels for HDV, 1021 pixels for LFU, 997 pixels for BL, 971 pixels for LSV
and 637 pixels for DHU while the testing dataset consisted of 5926 pixels, with 1230 pixels
for WF, 978 pixels for HDV, 1054 pixels for LFU, 951 pixels for BL, 864 pixels for LSV and
849 pixels for DHU. In order to reflect the within class variation, samples had been collected
from at least three different locations for each type of features. For a stable land cover classes
such as urban features (DHU, LFU) only pixels which are clearly recognizable and
unchanged in all images were selected. For features which are changed according to season
such as vegetation (HDV, LSV), water surface (WF) or bare land (BL), samples has to be
selected using the old land use map, the cropping calendar and the appearance of these pixels
on multi-date images must agree with the knowledge of their seasonal growing pattern.
In this study, the Transformed Divergence (TD) and Jefferies-Matusita (JM) indices were
also employed to examine separability of combined datasets of multi-spectral and SAR dual-
polarisation imagery for land cover mapping. These indices are widely used in classification
of remote sensing data for analysing separability between classes based on their statistical
distances (Jensen 2004, Erasmi and Twele 2009, Hack and Khatiwada 2010). The TD is
computed using:
( )( )[ ] ( )( )( )[ ]Tjijijijijiij CCtrCCCCtrD μμμμ −−−+−−= −−−− 1111
21
21 (3.6)
⎥⎥⎦
⎤
⎢⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛ −−=
8exp12000 ij
ij
DTD (3.7)
58
where: Dij and TDij are Divergence and Transformed Divergence of the two classes i and j, tr
is trace of the matrices (sum of diagonal elements), Ci, Cj are class covariance matrix, and μ
is the class mean vector. The JM is computed as:
( ) ( ) ( )⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛
×
++−⎟⎟
⎠
⎞⎜⎜⎝
⎛ +−=
−
ji
jiji
jiTji
CC
CCCC 2/ln
21
281
1
μμμμα (3.8)
)1(2 α−−= eJMij (3.9)
where: α is the Bhattacharyya distance, and JMij are Jefferies-Matusita indices of the two
classes i and j.
The Transformed Divergence and Jefferies-Matusita indices are saturated at 2000 (TD) and
1.4142 (JM), which implies excellent between-class separation. Jensen et al. (2004) claimed
that the TD values above 1900 indicate good separation, while the TD values less than 1700
are considered to be poor.
Classification accuracy assessment
The confusion matrix, which represents the relationship between the classification results and
reference data, is widely used for classification accuracy assessment. The performance of
land cover classification was evaluated using several indices, such as overall accuracy,
producer’s and user’s accuracy, and Kappa coefficient derived from the confusion matrices.
The overall accuracy is a rate between the total number of correctly classified samples and
the total number of reference samples. The producer’s accuracy for each class is the
proportion of samples in the reference dataset that are correctly classified by the classifier,
while the user’s accuracy measures the proportion of classified samples that agree with the
reference data. The Kappa coefficient uses all of the elements from the confusion matrix to
measure the difference between the actual agreement of the classification result with the
reference data and the chance agreement of allocated labels matching the same reference data
(Ban and Wu 2005, Gao 2009, Tso and Mather 2009).
3.1.3. RESULTS AND DISCUSSION
3.1.3.1. Uses of separability indices
59
Average values of TD and JM for all datasets with six land cover classes are given in Table
3.2. The averages of TD and JM indices of single-date images are ~1144 (TD) & ~0.9566
(JM), ~1085 (TD) & ~0.9457 (JM) and ~1470 (TD) & ~1.1252 (JM) for HH, HV and dual-
polarised (HH+HV), respectively.
Table 3.2: Average values of Transformed Divergence (TD) and Jefferies-Matusita
separability indices of different combined datasets.
ID Datasets TD JM
1 Five date PALSAR HH polarised images (5Date_HH) 1547 1.178
2 Five date PALSAR HV polarised images (5Date_HV) 1548 1.141
3 Five date PALSAR dual HH+HV polarised images (5Date_Dual) 1812 1.292
4 Five date PALSAR HH+ four coherence images (5date_HH +4 Coh) 1835 1.314
5 Five date PALSAR HV+ four coherence images (5date_HV +4 Coh) 1797 1.290
6 Five date PALSAR dual HH+HV polarised images + four coherence images
(5Date_Dual + 4Coh) 1923 1.351
7 Five date PALSAR HH+ ten coherence images (5date_HH +10 Coh) 1941 1.364
8 Five date PALSAR HV+ ten coherence images (5date_HV +10 Coh) 1930 1.354
9 Five date PALSAR HH + HV + ten coherence images (5date_HH+HV +10 Coh) 1974 1.383
10 SPOT2 multi-spectral images (SPOT 2 XS) 1881 1.331
11 SPOT 2 XS + single PALSAR dual-polarised HH+HV images (XS +1Date_Dual) 1992 1.396
12 SPOT 2 XS + five-date PALSAR HH polarised images (XS + 5Date_HH) 1999 1.407
13 SPOT 2 XS + five-date PALSAR HV polarised images (XS + 5Date_HV) 1914 1.366
14 SPOT 2 XS+ five-date PALSAR dual HH+HV polarised images (XS + 5Date_Dual) 2000 1.411
15 SPOT 2 XS + five-date PALSAR HH polarised images (XS + 5Date_HH+4Coh) 2000 1.412
16 SPOT 2 XS + five-date PALSAR HV polarised images (XS + 5Date_HV + 4Coh) 1951 1.388
17 SPOT 2 XS+ five-date PALSAR dual HH+HV polarised images
(XS + 5Date_Dual+ 4Coh) 2000 1.413
18 SPOT2 multi-spectral + five-date PALSAR HH polarised images + ten coherence
images (XS + 5Date_HH+10Coh) 2000 1.413
19 SPOT 2 XS + five-date PALSAR HV polarised images + ten coherence images
(XS + 5Date_HV + 10Coh) 1988 1.404
20 SPOT 2 XS + five-date PALSAR dual HH+HV polarised images
(XS + 5Date_Dual+ 10Coh) 2000 1.414
60
It appears that these indices in general are good indicator of separability between classes.
Single SAR images, no matter whether like-, cross- and dual-polarised images, gave very
poor separability. All of these images had TD values less than 1500 and JM values less than
1.2. The TD and JM values increased sharply for multi-date SAR images, which reach 1547
(TD) & 1.178 (JM) and 1548 (TD) and 1.141 (JM) for five-date HH and five-date HV data,
respectively. The five-date dual-polarised datasets further improved the separability to 1881
(TD) and 1.331 (JM). The interferometric coherence data also made a substantial
improvement of class separability when combined with multi-date SAR images. In particular,
their integration with multi-date dual-polarised SAR images provided TD and JM separability
values greater than 1900 and 1.350, respectively. These are considered to exhibit good
separation (Jensen, 2004). The SPOT 2 XS image had acceptable separability values of 1881
(TD) and 1.331 (JM), which are very close to the good separation level suggested by Jensen
(2004). It is clear that the integration of SPOT 2 XS image with multi-date SAR data gave
even higher values of separability – all of these combined datasets have TD values greater
than 1900 and six of them provided the saturated TD value of 2000, while their JM value
were also very high, and six of them had the JM values greater than 1.41 (which is very close
to the maximum value). However, it is worth noting that the average TD or JM value only
represents a general indication of separability between classes. For a detailed study the TD
and JM separability between particular pairs of classes should be investigated.
3.1.3.2. Combination of multi-date, multi-polarisation SAR images
The overall classification accuracy and Kappa coefficients for the ANN, SVM and ML
classifiers for different SAR combined datasets is summarised in Table 3.3 and Figure 3.3.
61
Table 3.3: Land cover classification accuracy of different SAR combination datasets.
Datasets
ID
SVM accuracy ANN accuracy ML accuracy
Overall (%) Kappa Overall (%) Kappa Overall (%) Kappa 1 58.12 0.50 56.87 0.48 52.57 0.43
2 54.00 0.45 50.89 0.41 53.00 0.44
3 65.51 0.59 68.16 0.62 63.35 0.56
4 71.04 0.65 65.86 0.59 70.55 0.65
5 68.87 0.63 63.25 0.56 62.98 0.56
6 75.43 0.71 74.03 0.69 62.97 0.56
7 65.51 0.59 62.94 0.55 61.59 0.54
8 58.39 0.50 59.62 0.52 57.39 0.49
9 72.31 0.67 67.99 0.62 67.06 0.61
Single-date and single-polarised SAR images provided very poor classification accuracy,
with averages of 49.35% and 45.53% using the SVM classifier for HH and HV polarised
images, respectively. The classification of a single SAR image was not possible for the ANN
and ML classifiers. The SVM classification of single-date, dual-polarised images provided an
increase of approximately 8.45% in overall accuracy. However, it is still considered to be
rather poor, with just 57.80% accuracy. Multi-date, single-polarised PALSAR images
produced some improvement in classification performance in comparison with single-date
SAR images – the classification accuracy using the SVM classifier increased by 8.77% and
8.47% for HH and HV polarised images, respectively. The SVM classifier provided a 7.71%
increase in accuracy for classifying multi-date, dual-polarised images. The improvement was
13.17% in the case of the ANN classifier. Use of multi-date, dual-polarised data resulted in a
remarkable improvement in classification accuracy as compared with using multi-date,
single-polarised datasets. The accuracy increased by 7.39% and 11.82% as compared with
multi-date HH images using SVM and ANN classifiers, respectively. The improvements
were even more significant for the multi-date HV image, with an increase of 11.51% (for
SVM classifier) and 17.27% (for ANN classifier) in classification accuracy.
Sensitivity of like- and cross-polarisation to surface and volume scattering mechanisms were
clearly demonstrated in the classification results. As can be seen in Figure 3.3 and Table 3.4,
for the SVM classifier, while the like-polarisation (HH) resulted in better accuracy for
62
classifying urban structures and bare land (which are subject to surface scattering), cross-
polarisation (HV) produced higher accuracy for both High Dense Vegetation (HDV) and Low
Sparse Vegetation (LSV) classes. The classification results of multi-date, dual-polarised
PALSAR images (Table 3.4) highlighted the complementary nature of like- and cross-
polarised SAR data. Accuracy for most land cover classes increased significantly, except for
the LSV class, compared to the case of classifying single-polarised data. These
complementary effects are also illustrated in Figure 3.3, which shows results of classification
using five-date, single-polarised (HH or HV), and five-date, dual-polarised (HH+HV) SAR
images. In the results of ANN classification, the use of SAR multi-date, dual-polarised
images also gave a remarkable improvement in classification performance as compared to the
multi-date, single-polarised images (Table 3. 5).
Table 3.4. Producer and user accuracy (%) for the SVM classifier applied to five-date
PALSAR HH, HV, and five-date PALSAR dual-polarised (HH+HV) images.
Land cover classes Five-date HH Five-date HV Five-date HH+HV
Producer User Producer User Producer User
Water Surface (WF) 40.89 76.44 47.48 76.84 47.80 85.89
Bare Land (BL) 57.73 47.91 51.21 55.59 58.15 59.72
Dense High Urban structures (DHU) 87.40 91.27 34.51 37.18 94.46 67.39
Low Flat Urban structures (LFU) 80.08 55.27 76.47 56.25 89.09 69.45
High Dense Vegetation (HDV) 48.88 48.43 54.19 52.95 57.16 65.84
Low Sparse Vegetation (LSV) 37.96 41.26 57.87 46.82 51.04 47.83
Table 3.5: Producer and user accuracy (%) for the ANN classifier applied to five-date
PALSAR HH, HV, and five-date PALSAR dual-polarised (HH+HV) images.
Land cover classes Five-date HH Five-date HV Five-date HH+HV
Producer User Producer User Producer User
Water Surface (WF) 31.30 87.30 35.77 79.28 60.16 77.81
Bare Land (BL) 81.39 43.88 62.15 47.74 62.25 60.59
Dense High Urban structures (DHU) 82.21 97.90 0.00 0.00 83.98 96.61
Low Flat Urban structures (LFU) 79.32 55.88 98.29 43.02 95.54 63.53
High Dense Vegetation (HDV) 57.06 44.50 42.54 53.96 53.99 57.89
Low Sparse Vegetation (LSV) 13.77 46.12 61.69 55.87 53.82 60.94
63
Figure 3.3: Part of classification results using SVM and ANN classifiers on multi--date,
single-polarised (HH or HV), and multi-date, dual-polarised HH+HV images;
A) SPOT 2 XS false colour images, B) PALSAR HH polarised image, C) PALSAR HV
polarised image, D) SVM classification of multi-date, dual-polarised (HH+HV) images, E)
SVM classification of multi-date, single-polarised HH images, F) SVM classification of
multi-date, single-polarised HV images, G) ANN classification of multi-date, dual-polarised
(HH+HV) images, H) ANN classification of multi-date, single-polarised HH images, K)
ANN classification of multi-date, single-polarised HV images.
64
Integration of SAR interferometric coherence data with polarimetric images gave significant
increase in classification performance. For example, the combinations of five-date PALSAR
HH polarised images with four and ten coherence images resulted in a rise of 12.92% and
7.39% in the overall accuracy of the SVM classification, respectively, as compared to the
result of using only the multi-date PALSAR HH images. The increase of classification
accuracy was 9.52% and 6.6% for the ANN classifier, respectively. The integration of multi-
date PALSAR cross- and dual-polarisation images with coherence data also provided
remarkable improvements in accuracy in the range of 4.39% to 14.87%.
3.1.3.3. Combination of SAR multi-date polarisation, interferometric coherence and optical
images
Results of classification using different combinations of SAR and optical images is
summarised in Table 3.6.
Table 3.6: Land cover classification accuracy of different combinations of SPOT 2 XS and
PALSAR polarised images.
Datasets
ID
SVM accuracy ANN accuracy ML accuracy
Overall (%) Kappa Overall (%) Kappa Overall (%) Kappa 10 69.43 0.63 70.08 0.64 64.68 0.58
11 86.55 0.84 86.38 0.84 80.19 0.76
12 89.27 0.87 88.67 0.86 82.51 0.79
13 75.31 0.70 72.92 0.68 66.72 0.60
14 88.31 0.86 89.16 0.87 82.59 0.79
15 86.10 0.83 88.17 0.86 82.55 0.79
16 78.18 0.74 79.77 0.76 70.96 0.65
17 86.37 0.84 84.66 0.82 82.57 0.79
18 81.74 0.78 81.52 0.78 78.28 0.74
19 76.63 0.72 77.35 0.73 67.75 0.61
20 84.61 0.82 83.72 0.80 79.21 0.75
The SPOT multi-spectral image gave an acceptable classification accuracy using the ANN
and SVM classifier. However, as can be seen in Figure 3.4 and Table 3.7, for SPOT 2 XS
classification, the urban areas were confused with bare ground, and it was very hard to
distinguish between Dense High Urban (DHU), Low Flat Urban (LFU) and Bare Land (BL)
65
classes due to their similar spectral properties. Consequently, the accuracy of these features,
particularly DHU and LFU, was poor.
Figure 3.4: Part of the classification results using SVM and ANN classifiers on a SPOT 2 XS
image and the combination of SPOT 2 XS and multi-date ALOS/PALSAR HH polarised
images.
A) the SPOT 2 XS image, B) SVM classification of the SPOT 2 XS image, C) SVM
classification of a combination of SPOT 2 XS + five-date PALSAR HH polarised images, D)
the ALOS/PALSAR HH polarised image, E) ANN classification of the SPOT 2 XS image, C)
ANN classification of a combination of SPOT 2 XS+ five-date PALSAR HH polarised
images
66
Table 3.7: Producer and user accuracy for ANN and SVM classifier applied to SPOT 2 XS
multispectral images.
Land cover classes ANN SVM Producer
accuracy (%) User
accuracy (%) Producer
accuracy (%) User
accuracy (%) Water Surface (WF) 98.05 90.40 96.95 96.35
Bare Land (BL) 62.15 94.41 80.86 80.36
Dense High Urban structures (DHU) 2.47 56.76 44.17 38.11
Low Flat Urban structures (LFU) 86.05 46.47 26.94 35.68
High Dense Vegetation (HDV) 61.35 98.52 76.69 92.25
Low Sparse Vegetation (LSV) 95.83 60.70 86.89 65.47
The combination of SPOT 2 multi-spectral and multi-date SAR images gave the most
significant increase in classification accuracy, using either ANN or SVM classifiers. As can
be seen in Figures 3.4 and 3.5, and Table 3.8, confusion between BL, DHU and LFU was
significantly reduced by integrating SPOT 2 and PALSAR images because of differences in
the backscatter patterns of these features in SAR images. The highest overall accuracy was
89.27% for classifying the combined dataset of SPOT 2 + five-date PALSAR like- (HH)
polarised images using the SVM classifier. The improvements compared to the case of using
only SPOT 2 or five-date PALSAR HH images were 19.84% and 31.15%, respectively.
Table 3.8: Producer and user accuracy for ANN and SVM classifier applied to combination
of SPOT 2 XS multi-spectral and five-date PALSAR HH polarised images.
Land cover classes ANN SVM Producer
accuracy (%)
User
accuracy (%)
Producer
accuracy (%) User
accuracy (%)Water Surface (WF) 90.81 96.63 94.31 99.66
Bare Land (BL) 96.32 92.81 98.32 94.83
Dense High Urban structures (DHU) 84.45 96.89 78.45 98.09
Low Flat Urban structures (LFU) 94.50 87.91 95.45 82.12
High Dense Vegetation (HDV) 74.34 90.88 74.13 95.39
Low Sparse Vegetation (LSV) 90.51 70.45 92.36 71.76
Although the multi-date SAR, dual-polarised images (HH+HV) using SVM or ANN
classifiers provided significant improvements for classifying land cover features compared to
the single-polarised images, its combination with the optical data (SPOT 2 XS) did not yield
noticeably higher accuracy than the combination of multi-date PALSAR HH polarisation and
67
SPOT 2 XS images. Actually, this combination reduced the overall accuracy by 0.96% for
the SVM classification. However, it gave a slight increase in overall accuracy of 0.49% for
the ANN classification.
In a similar fashion, adding the interferometric coherence data to the combination of SPOT 2
XS + SAR multi-date polarised images did not further enhance the classification performance
as expected. Only the integration of interferometric coherence with the combination of SPOT
2 XS+ five-date, cross- (HV) polarised images gave a significant increase of overall
classification accuracy by 2.87% and 6.85% for SVM and ANN classifiers, respectively. In
the cases of SPOT 2 XS +like- (HH) or dual- (HH+HV) polarisation, the integration with
coherence data even noticeably decreased the classification accuracy for both SVM and ANN
classifiers.
It is also important to note that in all cases the integration of four short-term interferometric
coherence images produced better results than using all ten interferometric coherence images.
These results demonstrate that incorporation of more input data does not necessarily increase
the classification performance. It is crucial to apply appropriate techniques to select the
relevant features to be included in the classification datasets.
68
Figure 3.5: Part of classification results using SVM and ANN classifiers on a SPOT 2 XS
image and combination of SPOT 2 XS and multi-date PALSAR HH, HV polarised datasets.
A) SPOT 2 XS false colour image, B) PALSAR HH polarised image, C) PALSAR HV
polarised image, D) SVM classification of a SPOT 2 XS image, E) SVM classification of
SPOT 2 XS+ five-date PALSAR HH polarised image, F) SVM classification of SPOT 2 XS
+ five-date PALSAR HV polarised images, G) ANN classification of a SPOT 2 XS image, E)
ANN classification of SPOT 2 XS + five-date PALSAR HH polarised image, F) ANN
classification of SPOT 2 XS + five-date PALSAR HV polarised images
69
3.1.3.4. Integration of textural data
Results of classification from the best combinations of multi-date PALSAR images with
textural data are given in Table 3.9. Sets of textures which gave the best classification
accuracy for each classifier when combined with multi-date PALSAR images are as follows:
Multi-date PALSAR
datasets
Best incorporated textural feature
SVM ANN
Five-date HH images Homogeneity + Entropy
Window size: 7x7
Variance + Entropy + Correlation
Window size: 11x11
Five-date HV images Variance + Entropy
Window size: 11x11
Variance + Entropy + Correlation
Window size: 9x9
The integration of multi-date, single-polarised SAR data with its selected best texture
measures gave a noticeable increase in the classification accuracy. While the combination of
five-date, cross-polarised images with their textures gave an increase of 4.29% and 7.51% in
classification accuracy for both ANN and SVM classifiers, the improvements from the
integration of five-date, like-polarisation images with their textures were rather more
significant, 6.95% and 8.91% for the SVM and ANN techniques, respectively. However, the
contributions of these texture measures were negligible in the case of the five-date, dual-
polarised images, and it even reduced the accuracy (0.41%) in the case of the ANN classifier.
Table 3.9: Comparison of land cover classification with multi-date SAR images including
like-, cross- and dual-polarised data and combination of these images with their best textural
features.
Datasets Accuracy (%)
Datasets Accuracy (%)
SVM ANN SVM ANN
Five-date HH 58.12 56.34 Five-date HH +
Best textural features 65.07 65.25
Five-date HV images 54.00 50.89 Five-date HV +
Best textural features 58.29 58.40
Five-date dual
(HH+HV) 65.51 68.16
Five-date dual
(HH+HV) + Best textural features
66.99 67.75
70
3.1.3.5. Comparison of SVM, ANN and ML classification performance
The comparison of SVM, ANN and ML classification performance is illustrated in Figure
3.6. and 3.7. The SVM produced the highest classification accuracy of 89.27% (Kappa=0.87)
for the combination of SPOT 2 XS and five-date PALSAR HH polarised images. The ANN
produced the highest classification accuracy of 89.16% (Kappa=0.87) for the combination of
SPOT 2 XS and five-date PALSAR dual (HH+HV) polarised images. The highest
classification accuracy achieved by ML was 82.59% (Kappa=0.79) also with the combination
of SPOT 2 XS and five-date PALSAR dual (HH+HV) polarised images. It is clear that the
non-parametric classifiers (SVM, ANN) have achieved significantly better performance than
the parametric (ML) ones, particularly for more complex datasets. In most cases the SVM
and ANN algorithms produced higher classification accuracy than the ML method. Among
the classification of 20 combined datasets the SVM classifier gave the best accuracy for 13
cases, the ANN classifier provided the highest accuracy for 7 cases. The ML classifier did not
provide the highest classification accuracy for any cases. The ML classifier always gave
lower accuracy than the SVM classifier and it performed better than the ANN classifier for
only two datasets.
Results of this study have shown that the SVM is a very efficient method for classifying
multisource remote sensing data. The SVM algorithm clearly outperformed the ML, and
generally gave better accuracy than the ANN algorithm. The analyses of classification Kappa
coefficients confirmed the same trends in classifier performance as using the overall
accuracy.
71
Figure 3.6: Comparison of classification accuracy of SVM, ANN and ML classifiers.
Figure 3.7: Comparison of Kappa coefficients generated by SVM, ANN and ML classifiers.
72
3.1.3.6. Relationship between separability indices and classification performance
The results of classification and an analysis of the separability indices have shown that, in
general, the TD and JM indices gave good indications of classification performance. Low
values of the TD or JM indices are synonymous with poor classification accuracy, while the
high values of TD or JM typically imply good classification performance. However, when the
separability values are high, these indices cannot measure the quality of classification
accurately. Many datasets, which have lower separability values, gave better classification
accuracy than those with higher TD or JM values. The other limitation of these separability
indices is that they tend to increase (even to the maximum level) as more features are
integrated into the datasets. The performance of the two indices was similar, however, the JM
was slightly better since TD index is more easily saturated than the JM index.
3.1.4. SUMMARY REMARKS
The evaluation in this study has revealed the advantages of the integrated approach,
combining multi-date, like- and cross-polarised and interferometric coherence SAR data with
optical images for mapping land cover features using non-parametric classification
algorithms. Joint application of like- and cross-polarised (or dual-polarised) SAR images
enhanced the performance of land cover mapping with respect to the case of using single-
polarised images. The dual-polarised dataset gave 65.51% and 68.16% of overall
classification accuracy for SVM and ANN classifiers, respectively, while the figure was less
than 59% for single-polarised datasets.
The combination of optical multi-spectral data and either single- or dual-polarised, as well as
multi-date, SAR images resulted in a significant improvement in classification accuracy, up
to 19.84% compared to the use of only a SPOT 2 XS image, and 31.15% compared to five-
date, single-polarised dataset. The contribution of interferometric coherence data is also
significant. Its integration with polarimetric SAR images increased the classification accuracy
dramatically. The highest classification accuracies were 75.43% (SVM) and 74.03% (ANN)
for the combination of multi-date, dual-polarised SAR images and the multi-date, short-term
coherence images. However, incorporation of these kinds of data with the combination of
SPOT 2 XS data and polarised SAR images did not produce a significant improvement
except for the case of SPOT 2 XS+ five-date, cross (HV) polarised images. Incorporation of
73
textural information with PALSAR images resulted in a considerable increase in
classification accuracy, particularly for the like-polarised images, with an improvement up to
8.91% (ANN).
The non-parametric classifiers (SVM and ANN) were successfully applied for classifying
multisource remote sensing data. The non-parametric classifiers clearly outperformed the
parametric classifiers (ML) in most cases, particularly for complex datasets. The
performances of the two non-parametric classifiers were rather comparable. However, the
SVM classifier was marginally better and in 13 out of 20 cases provided the highest accuracy.
It is also important to note that in many cases the integration of more input features did not
enhance the classification performance, occasionally reduced the classification accuracy. This
problem will be analysed and resolved in the next chapter of the thesis.
3.2. COMBINATIONS OF L- AND C-BAND SAR AND OPTICAL SATELLITE IMAGES FOR LAND COVER CLASSIFICATION IN THE RICE PRODUCTION AREA
The C-band SAR data mainly provides information on the upper layer of land cover, while
the properties of the lower surfaces can be captured using L-band SAR because of its greater
penetration capability. Hence, integration of L- and C-band SAR could enhance the capability
of land cover mapping. It is also expected that the classification performance could be further
improved by integrating multi-frequency SAR data with optical images, using an appropriate
classification algorithm such as non-parametric classifiers.
This study investigated the performance of non-parametric classifiers, including SVM and
ANN algorithms, to classify various land cover features in the south of Vietnam using a
multi-temporal/dual-frequency combination of ALOS/ PALSAR (L-band) radar,
ENVISAT/ASAR (C-band) radar and SPOT 4 multi-spectral optical satellite data. The
classification processes were carried out for selected single, and combined datasets. The
classification results of the SVM and ANN algorithms were compared with results of the
traditional parametric ML classifier.
74
3.2.1. STUDY AREA AND DATA USED
The study area is located within the borders between Tien Giang and Long An provinces and
Ho Chi Minh City in the South of Vietnam. The centred coordinate is 106o40’ E and 10o 25’
N. This area is typical for the rural area in the Mekong Delta region which is characterised by
very flat terrain and high fertility. It is the most important rice production region of the
country. The main land use/cover types are rice, mangrove, residential areas and water
surfaces.
Four ALOS/PALSAR (L-band) HH polarisation images acquired between June 2007 and
June 2008, and four ENVISAT/ASAR (C – band) HH polarisation images acquired for a
similar period (from May 2007 to July 2008) were used for this study (Table 3.10).
A SPOT 4 multispectral (SPOT 4 XS) image has similar properties as previous generations
(SPOT 1, 2 and 3) with a spatial resolution of 20m. However, instead of having only three
spectral bands – Green (G), Red (R) and Near-Infrared (NIR) – it has a fourth band obtained
in the shortwave infrared region (SWIR) of the spectrum. The SPOT 4 XS image for this
study was acquired on December 09, 2007 (Figure 3.8).
Table 3.10: ALOS/PALSAR and ENVISAT/ASAR images for the study area.
Sensors Date of Acquisition Orbit Wavelengths Polarisation
ALOS/PALSAR
08 Jun 2007 Ascending L HH 08 Sep 2007 Ascending L HH 09 Dec 2007 Ascending L HH 10 Jun 2008 Ascending L HH
ENVISAT/ASAR
30 May 2007 Ascending C HH 12 Sep 2007 Ascending C HH 26 Dec 2007 Ascending C HH 23 Jul 2008 Ascending C HH
75
Figure 3.8: SPOT 4 multispectral false colour image (left) and ALOS/PALSAR HH
polarisation image (right), both images acquired on December 09, 2007.
3.2.2. METHODOLOGY
The geo-rectification of ALOS/PALSAR and SPOT 4 XS images was carried out similarly to
the process described early in the section 3.1.2. The ENVISAT/ASAR images were firstly
geo-rectified using the Nest 4.0 software and then co-registered to the ALOS/PALSAR
images using the same procedure. All data were re-sampled to a pixel size of 25m.
ALOS/PALSAR and ENVISAT/ASAR backscatter data were converted to decibel (dB) and
filtered using the Enhanced Lee filter with 5x5 window size.
Although rice is largely predominant in the study area, the cultivated cycles vary depending
on each land holder. Thus, rice fields appear differently in any single-date image. As in the
case of the SPOT 4 multispectral image, the rice areas were currently growing rice, harvested
(temporally dry bare ground), or filled with water. Hence, the use of multi-date data is more
effective than single-date data for rice area mapping. For this reason this study employed
multi-temporal SAR images together with optical images to map land cover features. The
capability of single-date SAR image for classifying land cover classes was also evaluated,
based on a SAR image acquired on a similar date to the SPOT 4 images in December, 2007.
Grey Level Co-occurrence Matrix (GLCM) textures were generated for each data type with
different window sizes (3x3, 5x5, 7x7 and 9x9). Four texture measures, namely variance,
76
homogeneity, entropy and correlation, were selected because of their low correlation with
respect to each other.
Since the rice fields appear differently in images acquired at different dates, the
classifications were carried out in two steps. At the beginning, seven land cover classes were
defined for classification – Dense Residence (DR), Medium Sparse Residence (MR), Water
Body (WB), Mangrove (MG), Current Rice Field (CR), Rice Temporal Wet (RW) and Rice
Temporal Bare Ground (RG). After the classification process, Current Rice Field, Rice
Temporal Wet, Rice Temporal Bare Ground were merged together to form only one class:
Rice Field (RF). The spectral and backscatter properties of these classes are presented in
Figure 3.9 and 3.10.
The classification process was implemented on different datasets, including single- and multi-
date SAR images, SPOT 4 images, combination of multi-date ASAR+PALSAR data,
combination of SPOT 4 and multi-date SAR images. Similar to the previous section, SVM
using the Radical Basic Function (RBF) Kernel and ANN using the MLP-BP algorithm were
employed for classifying land cover features. The cross-validation technique was applied to
optimise parameters for the SVM classifier. Classifications using ML classifier were also
carried out for comparison. Training and testing datasets were collected based on manual
interpretation with reference to the available land use map created in 2005 over the study area
and from a study of multi-date Google Earth images as described earlier in the section 3.1.2.
77
3.2.3. RESULTS AND DISCUSIONS
Figure 3.9: Spectral properties of different land cover classes in the SPOT 4 XS image.
Figure 3.10: Backscatter properties of different land cover classes in the multi-date L- and C-
band SAR images (Lx_HH and Cx_HH represent L and C band HH polarised images,
respectively).
78
Overall classification accuracies of different datasets are summarised in Table 3.11.
However, the discussion below mainly focuses on the result of classification using SVM and
ANN techniques after merging the Rice classes.
Table 3.11: Overall classification accuracies (%) for different datasets.
AVar3, AVar5, AVar9 and PVar3, PVar9 are GLCM Variance texture measures derived from the ASAR and PALSAR images with window size of 3x3, 5x5 and 9x9 respectively.
ID Datasets Before merging
Rice classes
After merging
Rice classes
ML ANN SVM ML ANN SVM
1 Single-date ASAR (Dec 26, 2007) - - 38.26 - - 48.17
2 Single-date ASAR + AVar5 40.96 35.55 43.97 46.72 44.92 52.98
3 Single-date PALSAR (Dec 09, 2007) - - 36.8 - - 48.17
4 Single-date PALSAR + PVar9 34.55 36.55 40.21 42.31 44.22 51.28
5 Single-date ASAR +PALSAR 43.92 42.01 44.11 49.98 52.73 53.83
6 Single-date ASAR +PALSAR + AVar5+PVar9 45.27 46.42 48.87 53.17 52.13 57.98
7 Four-date ASAR 57.74 58.24 59.84 68.30 68.02 69.65
8 Four-date ASAR + AVar9 58.44 54.68 59.19 70.61 64.45 71.95
9 Four-date PALSAR 51.93 53.08 52.32 65.34 67.05 65.85
10 Four-date PALSAR + PVar3 51.87 53.88 54.83 64.95 71.91 70.0
11 Four-date ASAR + PALSAR 65.2 66.45 66.1 66.1 77.47 78.87
12 Four-date ASAR + PALSAR + AVar9 + PVar3 66.1 65.30 63.35 76.76 78.62 75.06
13 SPOT 4 multispectral 84.63 87.28 83.63 86.53 88.13 84.98
14 SPOT 4+ Four - date ASAR 86.58 87.83 88.98 87.68 89.18 91.43
15 SPOT 4+ Four - date ASAR + AVar9 82.87 88.28 84.63 84.88 89.83 86.88
16 SPOT 4+four - date PALSAR 87.03 90.34 93.49 89.28 93.89 94.99
17 SPOT 4+four - date PALSAR + Pvar3 86.33 89.58 91.59 88.43 90.74 94.09
18 SPOT 4+ Four - date ASAR + Four date PALSAR 85.48 90.26 86.23 87.48 93.69 88.98
19 SPOT 4+ Four - date ASAR + Four-date
PALSAR+ Avart9 + Pvar3 82.42 90.54 91.29 84.78 92.74 93.54
The single-date SAR images had very poor classification accuracy (~ 48%). Combination of
two single-date L- and C-band SAR images gave approximately a 5% increase in accuracy.
Multi-temporal SAR data provided a dramatic increase in classification results with
improvements of 21.48% and 17.68% for the SVM classifiers using ASAR and PALSAR
data, respectively. However, the overall accuracy was still rather low – 69.65% and 65.85%
for four-date ASAR and PALSAR images, respectively. The performance of the ANN
79
classifier was similar with the SVM classifier for multi-date SAR data with an overall
accuracy of 68.00% for ENVISAT/ASAR and 67.05% for ALOS/PALSAR images. The
confusion matrices, which showed classification accuracy of the SVM and ANN classifiers
on four-date ENVISAT/ASAR and four-date ALOS/PALSAR image, for each land cover
class, are listed in Tables 3.12 – 3.15.
Table 3.12: Confusion matrix for SVM classification using four-date ENVISAT/ASAR HH
polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 715 27 59 77 9 80.61
WB 33 299 0 0 0 90.06
MG 64 0 130 76 13 45.94
MR 103 0 59 143 78 37.34
DR 0 0 0 8 104 92.86
Producer’s accuracy (%) 78.14 91.72 52.42 47.04 50.98 Overall accuracy 69.65
Table 3.13: Confusion matrix for ANN classification using four-date ENVISAT/ASAR HH
polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 643 9 35 49 1 87.25
WB 34 317 0 0 0 90.31
MG 111 0 132 74 13 40.00
MR 127 0 81 175 99 36.31
DR 0 0 0 6 91 93.81
Producer’s accuracy (%) 70.27 97.24 53.23 57.57 44.61 Overall accuracy 68.002
Table 3.14: Confusion matrix for SVM classification using four-date ALOS/PALSAR HH
polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 716 37 56 17 8 85.85
WB 83 275 0 0 0 76.82
MG 51 0 30 41 25 20.41
MR 17 0 162 243 78 44.83
DR 48 14 0 3 51 43.97
Producer’s accuracy (%) 78.25 84.36 12.10 79.93 25.00 Overall accuracy 65.85
80
Table 3.15: Confusion matrix for ANN classification using four-date ALOS/PALSAR HH
polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 678 24 29 4 3 91.87
WB 105 302 0 0 0 74.20
MG 117 0 62 69 32 22.14
MR 14 0 143 228 100 47.01
DR 1 0 14 3 69 79.31
Producer’s accuracy (%) 74.10 92.64 25.00 75.00 33.82 Overall accuracy 67.05
The producer’s and user’s accuracy of SVM and ANN classifications on the combination of
four-date ENVISAT/ASAR and four-date ALOS/PALSAR image are given in Tables 3.16 –
3.17. It was revealed that the combination of multi-temporal ASAR (C-band) and PALSAR
(L-band) images gave very significant improvement for land cover classification accuracy for
both SVM and ANN algorithms, with an increase of up to 13% of overall accuracy compared
to the cases of using only four-date ASAR or four-date PALSAR images. In particular, the
accuracy for classifying the RF class reached more than 90% for the SVM method for both
producer’s and user’s accuracy. The results clearly highlighted the complementary
characteristics of L- and C-band data. Separability between classes was remarkably
increased. As can be seen in Figure 3.11 and 3.12 the misclassification between WB and DR,
RF and MR as well as MG and RF is greatly reduced.
Figure 3.11: SVM Classification using four-date PALSAR images (left) and the combination
of four-date ASAR and four-date PALSAR images (right).
81
Figure 3.12: ANN Classification using four-date PALSAR images (left) and the combination
of four-date ASAR and four-date PALSAR images (right).
Table 3.16: Confusion matrix for SVM classification using four-date ALOS/PALSAR HH &
four-date ENVISAT/ASAR polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 847 17 36 27 11 90.30
WB 34 309 0 0 0 90.09
MG 9 0 95 35 14 62.09
MR 4 0 116 235 90 52.81
DR 21 0 1 7 89 75.42
Producer’s accuracy (%) 92.57 94.79 38.31 77.30 43.63 Overall accuracy 78.87
Table 3.17: Confusion matrix for ANN classification using four-date ALOS/PALSAR HH &
four-date ENVISAT/ASAR polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 852 62 44 36 13 84.61
WB 20 264 0 0 0 92.96
MG 31 0 71 20 8 54.62
MR 12 0 133 239 62 53.59
DR 0 0 0 9 121 93.08
Producer’s accuracy (%) 93.11 80.98 28.63 78.62 59.31 Overall accuracy 77.47
82
Results of classifications using SVM and ANN classifiers on the SPOT 4 XS image and its
combinations with four-date ALOS/PALSAR images, which produced highest classification
accuracy, are presented in Figures 3.13, 3.14 and Tables 3.18 - 3.21.
Figure 3.13: ANN classification using the SPOT 4 multi-spectral image (left) and
combination of SPOT 4 and four-date PALSAR images (right).
Figure 3.14: SVM classification using the SPOT 4 multi-spectral image (left) and
combination of SPOT 4 and four-date PALSAR images (right).
83
Table 3.18: Confusion matrix for ANN classification using SPOT 4 XS images.
Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 840 44 16 65 15 85.71
WB 16 282 0 0 0 94.63
MG 0 0 232 0 0 100.00
MR 57 0 0 217 0 79.20
DR 2 0 0 22 189 88.73
Producer’s accuracy (%) 91.80 86.50 93.55 71.38 92.65 Overall accuracy 88.13
Table 3.19: Confusion matrix for SVM classification using SPOT 4 XS images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 757 48 1 27 12 89.59
WB 16 278 0 0 0 94.56
MG 0 0 231 0 0 100.00
MR 105 0 0 240 1 69.36
DR 37 0 16 37 191 67.97
Producer’s accuracy (%) 82.73 85.28 93.15 78.95 93.63 Overall accuracy 84.98
Table 3.20: Confusion matrix for ANN classification using SPOT 4 XS images and four-date
ALOS/PALSAR polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 866 32 0 11 1 95.16
WB 15 294 0 0 0 95.15
MG 0 0 248 0 0 100.00
MR 26 0 0 270 6 89.40
DR 8 0 0 23 197 86.40
Producer’s accuracy (%) 94.64 90.18 100.00 88.82 96.57 Overall accuracy 93.89
Table 3.21: Confusion matrix for SVM classification using SPOT 4 XS images and four-date
ALOS/PALSAR polarised images. Land cover classes RF WB MG MR DR User’s accuracy (%)
RF 894 30 10 9 4 94.40
WB 14 296 0 0 0 95.48
MG 0 0 237 0 0 100.00
MR 4 0 1 288 18 92.60
DR 3 0 0 7 182 94.79
Producer’s accuracy (%) 97.70 90.80 95.56 94.74 89.22 Overall accuracy 94.99
84
Both ANN and SVM classifiers performed well with the SPOT 4 XS image, the overall
accuracy being 87.83% and 84.98% using ANN and SVM classifiers, respectively. However,
they produced a large amount of confusion between MR and RF due to the mixed of houses
and vegetation in the MR class.
The combination of optical and multi-temporal SAR images produced the most significant
improvement for classification accuracy. For example, the SVM classifier resulted in
increases in classification accuracy of 6.45% and 23.13% (SPOT 4 XS + ASAR) and 10.01%
and 29.4% (SPOT 4 XS + PALSAR) in comparison to the cases of using only SPOT 4 XS,
ASAR or PALSAR images. The integration of optical data with SAR images, no matter
whether C- or L-band, effectively reduced confusion between classes. The separability
between RF and MR increased noticeably for these combined dataset. It is interesting to note
that, while the classification accuracy of four-date PALSAR image was about 3% less than
the case of using four-date ASAR images, its integration with SPOT 4 images gave a little
higher accuracy for the combination of the four-date ASAR and the SPOT 4 images. The
possible reason is that the ASAR data with shorter wavelength appeared to be more highly
correlated with SPOT 4 XS data. The integration of SPOT 4 and both L- and C-band SAR
images also resulted in significantly increased classification accuracy compared to any single
kind of data. However, it provided less accuracy than from the integration of the SPOT 4 XS
image with either L- or C-band SAR data.
Integration of textural information
The integration of SAR backscattered data with their derived textural information was carried
out by gradually adding textural measures to the original datasets for classification. Results of
tests for different texture measures showed that the performance of variance was better than
other measures. However, the optimum window sizes were different for each combination.
The best window size to extract variance for the ASAR image on December 26, 2007 was
5x5, while it was 9x9 for the December 09, 2007 PALSAR image, and 3x3 for the four-date
PALSAR dataset. The tests on combinations of all of four texture measures did not give any
further improvement on the classification accuracy compared to a case of using only
variance. Texture information improved classification accuracy by 3 ~5% for the single SAR
images, four-date ASAR and four-date PALSAR images. However, no significant
85
improvement was found for the combination of multi-date ASAR + PALSAR or SPOT 4 XS
+ SAR datasets.
Performance of non-parametric and parametric algorithm
Results of classification using non-parametric (SVM, ANN) and parametric (ML) techniques
clearly showed that in most cases the SVM and ANN classifiers resulted in better
classification accuracy than the ML classifier (Figure 3.15). The SVM classifiers gave the
highest classification accuracy for 13 out of 19 cases, while the ANN classifier provided the
best accuracy for the six remaining cases. There are no cases for which the ML classifier
gave the highest accuracy. The highest classification accuracy was 94.99% for the SVM
classifier using the combination of SPOT 4 XS data + four-date PALSAR HH images. The
highest classification accuracy for ANN and ML classifier were 93.89% and 89.28%,
respectively. These results demonstrate that the SVM and ANN classifiers are very suitable
for classifying remote sensing data, particularly complex integration datasets.
Figure 3.15: Comparison of classification accuracy of SVM, ANN and ML classifiers after
merging rice classes (only combined datasets with at least two input features are presented).
42
52
62
72
82
92
4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Overall accuracy
Copmbined datasets
Comparison of performances of SVM, ANN and ML classifiers
SVM_Mg
ANN_Mg
ML_Mg
86
3.2.4. SUMMARY REMARK
The investigation has confirmed the advantages of combinations of multi-date, dual-
frequency SAR data as well as multi-date, dual-frequency SAR and optical data, for land
cover mapping in a complex rural area such as in the south of Vietnam. The integrated
approach always produced substantial increases in classification accuracy with respect to
cases of single datasets. The performance of non-parametric classifiers (SVM and ANN) is
better than the traditional ML classifier. The highest classification accuracy of 94.99% was
achieved using the SVM algorithm and the combination of SPOT 4 XS data and multi-date
PALSAR HH images. Therefore, theses non-parametric classifiers are considered more
appropriate for handling multisource remote sensing data.
However, as already mentioned in the previous section, in many cases the integration of more
features did not improve (and in some cases actually reduced) the classification accuracy.
This suggests the need for further studies to resolve this problem.
3.3. USES OF MULTI-TEMPORAL SAR AND INTERFEROMETRIC COHERENCE DATA FOR MONITORING THE 2009 VICTORIAN BUSHFIRES
In this section the application of multisource remote sensing data for a special case of land
cover mapping was investigated – the mapping of a burned area after bushfire. The
multisource data used for this study were multi-temporal SAR backscatter imagery and their
derived interferometric coherence data.
3.3.1. INTRODUCTION
In 2009 Australia experienced one of its historically worst natural disasters – The Victorian
Bushfires. The impact of this disaster was catastrophic. A total of 173 lives were lost, many
properties were damaged, and large areas of natural forests were severely destroyed. These
facts had emphasised needs for bushfire detection and monitoring. Optical satellite imagery
such as Landsat TM, SPOT, MODIS has been used extensively for forest fire mapping and
often give good results. However, the optical system can only operate during the day and
sensors may not “see” through clouds and smoke. Consequently it may not be possible to
observe the ground surface. In contrast, SAR systems operating in a microwave part of the
87
electromagnetic spectrum can image Earth’s surface at anytime with minimal impact from
clouds and smoke. Hence SAR imaging is appropriate for forest fire detection and the
mapping of burnt areas.
Use of SAR data for forest fire detection and burnt area mapping has been reported by a
number of investigators (Gimeno et al. 2004, Siegert and Hoffmann 2000, Siegert and
Ruecker 2000, Takeuchi and Yamada 2002, Almeida-Filho et al. 2007, Almeida-Filho et al.
2009). These studies mainly relied on differences in backscatter intensities between before
and after fire images to detect and map burnt areas. In Siegert and Rueker (2000), multi-
temporal ERS-2 SAR images were employed to identify burnt scars over a study site in east
Kalimantan, Indonesia. It was found that backscatter decreased by 2-5 dB in fire impacted
areas, while there was only a slight change in backscatter (less than 0.5 dB) for non-fire
impacted areas. Gimeno et al. (2004) reported that the changes of backscatter between
forested and fire-disturbed zones were up to +8 dB in ERS time series data for Mediterranean
forest regions. However, results of work carried out by Takeuchi and Yamada (2002) showed
a slight decrease in backscatter of JERS-1 SAR data after fire, but the difference was not
large enough to identify fire affected areas. Siegert and Hoffmann (2000) suggested that two
types of fire damage areas can be detected, severe damage with complete burning of
vegetation and medium damage with partly burnt vegetation. The first case related to a strong
decrease in backscatter while the latter related to areas with little change in backscatter.
Takeuhi and Yamada (2002) employed interferometric coherence SAR data for forest fire
damage monitoring. They reported a significant increase in coherence in damaged areas after
the fire allowing better interpretation and extraction of burnt areas compared to using
backscatter data. Antikidis et al. (1998) evaluated the capabilities of interferometric Tandem
ERS SAR data for mapping deforestation areas in Indonesia. It was found that there were
clear changes in coherences between images acquired before and after a fire event. The
objective of this study is to analyse the capabilities of multi-temporal SAR data, particularly
ALOS/PALSAR images, to identify burnt areas resulting from the Victorian Bushfires in
February 2009. The combination of SAR backscatter intensity and interferometric coherence
data is also a goal of the research.
88
3.3.2. STUDY AREAS AND DATA USED
Two test sites have been selected for this study. Site 1 is located in the far north-east of the
city of Melbourne at a distance of 200 km. The study area is covered by one full scene of
ALOS/PALSAR image (track 378, frame 643) with several suburbs, such as Bright, Mt
Beauty and Omeo. The main land cover types are forests, grass and bare ground. The second
test site covers a small part of the city of Melbourne and its northern regions, including the
Kinglake, Marysviles, Broadford, Yea, and Kilmore suburbs. The extent of the study area is
approximately a full scene of a ALOS/PALSAR image (track 383, frame 643). The main land
cover classes are forests, grass and urban residences. Specifications of ALOS/PALSAR data
over two study areas are given in Tables 3.22 – 3.25.
Table 3.22: ALOS/PALSAR images for the 1st study area.
Satellite/Sensor Track - Frame Acquisition dates Polarisation Orbit
ALOS/PALSAR 378_643
30/12/2008 HH Descending
14/02/2009 HH Descending
02/07/2009 HH Descending
Table 3.23: Interferometric pairs over 1st study area.
Satellite/Sensor Track-
Frame Date: Master Date: Slave
Baseline
length (m)
Observation
interval (days)
ALOS/PALSAR 378_643
30/12/2008 14/02/2009 270 46
30/12/2008 02/07/2009 947 186
14/02/2009 02/07/2009 680 138
Table 3.24: ALOS/PALSAR images for the 2nd study area.
Satellite/Sensor Track - Frame Acquisition dates Polarisation Orbit
ALOS/PALSAR 382_643
06/12/2008 HH Descending
21/01/2009 HH Descending
08/03/2009 HH Descending
24/07/2009 HH Descending
89
Table 3.25: Interferometric pairs over 2nd study area.
Satellite/Sensor Track-
Frame Date: Master Date: Slave
Baseline
length (m)
Observation
interval(days)
ALOS/PALSAR 378_643
06/12/2008 21/01/2009 304 46
06/12/2008 08/03/2009 878 92
06/12/2008 24/07/2009 1510 230
21/01/2009 08/03/2009 579 46
21/01/2009 24/07/2009 1196 184
08/03/2009 24/07/2009 626 138
Site 1: Landsat 5 TM acquired on 24/01/2009 and 25/02/2009.
Site 2: Landsat 5 TM acquired on 31/01/2009 and 16/12/2009.
These images were acquired before and just after the bushfire.
3.3.3. METHODOLOGY
All ALOS/PALSAR satellite images, including intensities and coherence data, were
registered to the map coordinate system (UTM projection, WGS84 datum) with 30m spatial
resolution. Speckle noise in the PALSAR images was filtered out using the Enhanced Frost
Filter with a 5x5 window size. SAR backscatter values are converted to decibel (dB) using
Equation (3.1). The Temporal Backscatter Changed images were generated using the
following relation (Chu and Ge 2010a):
SARCH= Max(Db1, Db2,.., Dbn) – Min(Db1, Db2 …, Dbn) (3.13)
where SARCH is the Temporal Backscatter Change images, Db1, Db2 …Dbn are pixel
backscatter values in SAR images, Max and Min are the functions to obtain the largest and
smallest pixel’s values within all images.
The SARCH image is in fact a modification of the intensity differencing technique. It allows
selection of the largest changes of pixel values between multi-temporal images. This image is
therefore very useful for change detection. As mentioned earlier, interferometric coherence
between pairs of SAR images is valuable data for burnt forest mapping/monitoring. The
coherence images measure the correlation between two images. In the coherence images
stable objects appear bright with high coherence value, while changed objects appear darker
with low coherence value. Hence it is expected that the coherence of pixels in fire areas will
90
be low or reduced compared to areas that have not changed. Moreover, the coherence of
pixels in a burnt area in a pair of across fire SAR images would be lower than the coherence
in a pair of before fire or after fires images. In this study the coherence data over two test
sites were analysed in conjunction with backscatter intensity data. Utilisation of the Principal
Component Analysis (PCA) for forest fire monitoring has been reported by researchers
(Siegert and Hoffmann 2000, Gimeno et al. 2004). It is revealed that PCA products generated
from multi-temporal SAR images could be applied for change detection, particularly for
bushfire mapping. The First Component PC1, and sometime even the Second Component
PC2, often represent non-changed features. Other components contain variations due to
changes in the ground surface.
Visual interpretation is an important step for detecting and collecting relevant information
about fire affected areas and other land cover features. Earlier studies have shown that the use
of multi-temporal SAR data improved feature discrimination in comparison to using single-
scene/date data (Gimeno et al. 2004). It is also revealed that more features can be
distinguished in the RGB composites than in a single image. In this study, burnt areas and
other land cover classes were identified by visual interpretation of RGB composites, in which
Red, Green and Blue colour channels were assigned to different dated SAR images or their
derived products (SARCH, PC components or average of multi-date SAR images). Then,
properties of burnt areas as well as other features were extracted for analysis and
classification.
Classification
Three classification techniques, namely the Maximum Likelihood (ML), Artificial Neural
Network (ANN) and Support Vector Machine (SVM), were applied for burn scar
identification. The classification results were compared with the burnt areas derived from
Landsat TM data.
Validation
Due to a lack of ground truth data, the Landsat TM images acquired before and after the fire
event was used for validation at both study sites. It is fortunate that burnt areas appear very
clearly in these optical images. Firstly, the Normalised Difference Vegetation Index (NDVI)
91
was computed for each Landsat image. Then pre- and post-fire NDVIs were subtracted from
each other to create the NDVI difference images. Finally, the NDVI difference image was
segmented to extract burnt areas, which were then used for testing the results of burn area
mapping with ALOS/PALSAR backscatter images and its coherence data.
3.3.4. RESULTS AND DISCUSSION
In both test sites uses of the RGB colour composite generated from original SAR backscatter
images (Figure 3.16) did not provide good separation between fire damaged areas and other
features. In these colour composite images burnt areas appeared in light purple or pink
colour, however the difference was relatively small and it was easy to confuse the similar
colours.
Figure 3.16: A) and C) showed fire affected areas are in purple to pink colour in multi-date
colour composites in 1st and 2nd study sites. B) and D) showed burnt area (dark colour) in
corresponding false colour Landsat TM images.
92
The temporal backscatter change images SARCH over the two study areas were generated
from multi-date SAR images. These images showed some burnt areas more clearly in very
bright tones, significantly different from surrounding features. However, other changed
features which were not subjected to fires also appear bright and it is difficult to differentiate
these changed features from the burnt areas.
An evaluation on all of interferometric coherence images revealed that, for both study sites
the coherence images generated from pairs of images acquired just before and after the
bushfire were useful. Burnt areas appear very dark due to low coherence between pre- and
post-fire events and are rather well separated from surrounding areas. Nevertheless, other
features such as unburnt forest may also exhibit low coherence due to natural de-correlation
factors (winds, changing of leaf position, etc.) and consequently appear similar in the
coherence data. Therefore the combination of backscatter intensity and interferometric
coherence data was applied and gave much better results compared to using any single
dataset. Based on this strategy the most useful RGB colour composite was found, where: Red
was assigned to Average SAR images (or the PC1 component), Green was assigned to
Temporal Backscatter Change images (SARCH) and Blue was assigned to interferometric
coherence data generated from an across-fire image pair. This RGB composite (Figure 3.17)
showed burnt areas in Yellow to Little Red colour due to strong backscatter (Red), large
changes (Green) and low coherence (Blue).
93
Figure 3.17: A) and C) are RGB colour composites of Average SAR image (R), Temporal
Backscatter Change (G) and across-fire coherence data (B) in 1st and 2nd study sites. Burnt
areas appear Yellow to Reddish in colour in A) and C) RGB composites.
PCA images were generated from multi-date SAR backscatter images. The PC1 components
were very similar with the average of all SAR images, and the PC3 and PC2 components
contained changed information useful for the first and second sites, respectively. However,
information on changed features can be readily obtained from the temporal backscatter
change image without concerns about noise as in the case of the PC2, PC3 components.
Based on visual interpretation several land cover classes were developed for each test site.
Site 1: Highly_burnt forest, Shadow_burnt, Un_burnt forest, Grass and Bare_ground.
94
Site 2: Highly_burnt forest, Medium_burnt forest, Shadow_burnt, Un_burnt forest, Urban
residence and Grass.
Details of their backscatters and coherence characteristics are shown in Figures 3.18 - 3.21.
Figure 3.18: Backscatter values for burnt and unburnt classes over the first test site.
Figure 3.19: Coherence values for burnt and unburnt classes over the first test site.
‐30
‐25
‐20
‐15
‐10
‐5
0
30/12/2008 14/02/2009 2/07/2009
Backscatter (dB
)
Mean Backscatter
Higly_burnt forest
Shadow_burnt
Un_burnt forest
Grass
Bare_ground
0
0.1
0.2
0.3
0.4
0.5
0.6
30/12/08 & 14/02/09
30/12/08 & 02/07/09
14/02/09 & 02/07/09
cohe
renc
e va
lue
Mean coherence values
Higly_burnt forest
Shadow_burnt
Un_burnt forest
Grass
Bare_ground
95
Figure 3.20: Backscatter values for burnt and unburnt classes over the second test site.
Figure 3.21: Coherence values for burnt and unburnt classes over the second test site.
It is interesting to note that the backscatter increased after the bushfires over both study sites.
The reason is that in the case of the 2009 Victorian Bushfires for many regions only leaves or
parts of leaves burned while tree trunks remained. Consequently, there are strong corner
‐25
‐20
‐15
‐10
‐5
0
6/12/2008 12/01/2009 8/03/2009 24/07/2009
Back
scat
ter (
dB)
Mean Backscatter
Higly_burnt forest
Medium_burnt forest
Shadow_burnt
Un_burnt forest
Urban Residene
Grass
96
reflector effects in the SAR images resulting from an increase in backscatter. For the second
test site, as Highly-burnt forest and Medium-burnt forest were identified based on the level of
severity of burning and can be detected. The most challenging task was to identify the burnt
areas which were in shadow areas. To map this class it was necessary to analyse the terrain. It
is clear that coherence values of fire affected areas were very low across the fire image pairs
(30/12/2008 and 14/02/2009 in first test site, and 12/01/2009 and 08/03/2009 in the second
test site) making them distinguishable from other features.
The classification process was performed over the two test sites with classes mentioned
previously. At the last steps, classification results were merged into only two classes, namely
Burnt and Unburnt classes. This final result was then compared with the burnt areas extracted
from the Landsat TM images. Since it was cloudy in the lower part of the Landsat TM image
(acquired on 25/02/2009) over the first site, the Landsat TM data for this site was only used
for general comparison. Quantitative assessment was carried out only over the second test
site. The following datasets were selected for classification:
Case 1: All PALSAR backscatter data
Case 2: Combination of Average SAR image + Backscatter Change Image + across-fire
coherence images
Case 3: All PALSAR backscatter data + selected coherence images (before + across + after
fire data)
Results of burnt areas classification are given in Table 3.26 and Figure 3.22.
Table 3.26: Overall classification accuracy assessment for various combined datasets and
classifiers over the second test site. Classifiers/ Datasets ML ANN SVM
Case 1 81.11 % 80.49 % 81.26 %
Case 2 84.50 % 81.25 % 84.90 %
Case 3 85.42 % 80.47 % 84.35 %
It is clear that in general the results of Burnt/Unburnt mapping using multi-temporal SAR
images agreed well with the results derived from Landsat TM images (over 80% agreement).
The use of combined SAR backscatter and coherence data improved the classification
accuracy by more than 4% for the ML and SVM classifiers. Although it is expected that the
97
ANN and SVM classifiers would give higher accuracy than the traditional ML classifier, it is
not a case for this study. The classification accuracy for ML and SVM classifiers was very
similar, but it was noticeably decreased for the ANN classifier. The reason could be the
nature of the validation data generated from the Landsat TM images, which may not
represent burnt areas accurately. Availability of ground truth data would have allowed better
accuracy assessment.
Figure 3.22: A) Burnt scars extraction from Landsat TM images, B) Burnt scars extraction
based on SVM classification (case 3), the Red areas represent burnt scars.
3.3.5. SUMMARY REMARKS
The integration of multi-temporal ALOS/PALSAR backscatter intensity images and
interferometric coherence data has been proposed and applied successfully to identify and
map the 2009 Victorian Bushfires burnt areas. Results indicate that this multisource dataset is
very effective for bushfire monitoring.
Neither backscatter intensity nor interferometric coherence data gives satisfactory separation
between Burnt and Unburnt areas. However, the synergistic use of these data allows
significant improvement in the detections and mapping of fire damaged areas. It was found
that the RGB colour composite (of average multi-temporal SAR images), the temporal
backscatter images and the across-fire interferometric coherence data were the most efficient
datasets.
98
Three classification techniques were implemented. Burnt areas extracted from Landsat TM
images acquired before and after the bushfire were used as validation data. The ML and SVM
algorithms gave similar accuracy. The ANN classifier produced the lowest classification
accuracy. However, it is likely that with more reliable ground truth data the classification
accuracy assessment will be more reliable.
3.4. CONCLUDING REMARKS
In this chapter the performance of non-parametric classifiers, SVM and ANN, for classifying
different combined datasets has been investigated. The advantages of using multisource data
for land cover classification have been demonstrated. It is shown that integration of multiple
kinds of remote sensing data give significant improvement in classification performance. The
contributions of different kinds of data to the improvements of the final results are presented.
The integrated approach does not only enhance the separability, but also improves the
interpretability of different land cover features. Although the improvements in classification
results were not only based on the use of non-parametric classifiers, these classifiers have
proven to be more effective than the traditional parametric classifier. In most cases the non-
parametric classifiers resulted in significantly higher classification accuracy than the
parametric classifier, particularly for the highly complex datasets with more incorporated
input features. Although for the case study of mapping burnt area after a bushfire the ANN
classifier did not perform better than the ML classifier, the reason may be the unreliable
validation dataset. The investigation confirmed that the non-parametric classifiers are not
only suitable for classifying remote sensing data in general, but are even far more appropriate
for classifying multisource combination datasets.
The experiments also revealed that in many cases using more input features does not
necessarily improve the classification performance. On the contrary, there may even be a
noticeable decrease in classification accuracy. Hence it is critical to select only relevant
features to be included in the combined datasets in order to enhance the quality of
classification. The other challenging task is to further improve the classification performance
of multisource remote sensing data – which methods should be used and how much
improvement can be obtained. These problems will be addressed in the following chapters of
this thesis.
99
1. CHAPTER 4
APPLICATION OF FEATURE SELECTION
TECHNIQUES FOR MULTISOURCE REMOTE
SENSING DATA CLASSIFICATION 4.1. SIGNIFICANCE OF DATA REDUCTION AND FEATURE SELECTION FOR CLASSIFICATION OF REMOTE SENSING DATA
The availability of a large variety of remote sensing data permits investigators to combine
various sources of digital data for land cover classification. Such integrated approaches have
the potential to increase the classification performance since inclusion of more relevant
information may reduce the confusion between classes (Tso and Mather 2009). However,
adding too many input data features will also increase the data dimensionality, with more
redundancy due to correlation among features and increase in uncertainty within input
datasets. These factor, might lead to a decrease of classification accuracy as well as an
increase in processing costs (Kavzoglu and Mather 2001, Richard and Jia 2006). Pal (2006)
suggested that the classification process should include only features that add an independent
source of information to the total dataset. Features which do not make any, or very little,
contribution to the separability of land cover classes should be removed (Richard and Jia
2006).
The classification process requires a sufficient number of training pixels in order to reliably
determine the class signature. As the number of input features increase, the required number
of training pixels also increases (Richard and Jia 2006). This requirement becomes very
challenging in practice for high dimensionality data such as multisource remote sensing
datasets. Richard and Jia (2006) recommended that it is essential to use as small a number of
input features as possible in a classification procedure in order to achieve reliable results from
a given set of training pixels.
100
Another important issue regarding the use of high dimensionality data for classification is the
so-called Hughes phenomenon (Figure 4.1), or ‘curse of dimmensity’ (Hughes 1968).
According to Hughes (1968), with a limited training sample size, the mean accuracy will
increase until it reached a peaked value, beyond which no significant improvement will be
achieved with additional measurements. Hence, the difficult task is to identify the subset of
features which provide better, or at least similar, accuracies as compared to when all the
features are used in a classification task (Anthony and Ruther 2007).
Figure 4.1: An example of the Hughes phenomenon with increasing of input data
dimensionality (source: Richard and Jia 2006).
The experiments carried out in the previous chapter have also revealed that in many cases
adding more input features did not improve the classification performance, and in fact could
decrease the classification accuracy significantly. Hence, the application of feature reduction
techniques, which reduce the data dimensionality while preserve informative features, is very
important for classification of remote sensing data, particularly high dimensional and
complex datasets such as multisource remote sensing data.
101
Feature reduction is categorised as “feature selection” and “feature extraction” (or
transformation). Feature extraction (FE) transforms the original data to a new feature space in
which inter-class separability of the transformed feature subset is generally higher than in
original datasets. The most commonly used FE method is based on Principal Component
Analysis (PCA). Unlike FE, the goal of feature selection (FS) is to select a subset from the
original dataset that preserves relevant characteristics of the dataset while removing irrelevant
and redundant information from data for an application such as classification (Chova et. al.
2006, Anthony and Ruther 2007, Pal 2006). The major advantage of FS as compared to FE is
that while FS maintains the physical meaning of the selected features the transformation by
FE changes the physical nature of the original data and leads to difficulties in interpretation
and evaluation of the final results.
Kavzoglu and Mather (2001) pointed out three major advantages in using FS techniques.
Firstly, the performance of a classification might be improved by selecting a subset of a
smaller number of informative and less correlated features. This issue is important for
improving the generalisation capability of the training dataset. Secondly, the processing time
will be reduced with the use of a smaller number of input features. Thirdly, the smaller
datasets would be more appropriate in cases where the number of training pixels is limited
because of the direct relationship between the dimension of the data and the size of the
sample set. Thus, FS is crucial for classification of remote sensing data.
4.2. FEATURE SELECTION TECHNIQUES USED FOR CLASSIFCATION OF REMOTE SENSING DATA
In the classification context, the FS challenge can be defined as: given a set X of N features,
find the best subset of m features (m<N) to be input into the classification process. FS
requires a search strategy and an objective function (Osuma 2005, Pal 2006, Bruzzone and
Persello 2009). The search strategy is an algorithm designed to find candidate subsets of
features from the original dataset, while the objective function will evaluate the effectiveness
of these feature subsets. The output of the FS algorithm is the optimal (or nearly optimal)
subset obtained for the classification (Maghsoudi et al. 2011, Pal 2006). FS techniques can be
categorised according to the search strategies or the objective function.
102
Search Strategy
Exhaustive Search Method
The simplest and most straightforward procedure is an Exhaustive Search by which all
possible combined subsets of features are considered and evaluated. Although this is the only
method that can theoretically guarantee the optimal solution, it may require considerable
computational power, particularly for large datasets, and is often not applied in practical
situations (Pal 2006). For example, in a case of a dataset with d features, there are 2d – 1
possible combined subsets to be considered. Anthony and Ruther (2007) suggested that this
method is practicable if the number of features is less than 10, and use of 10 or more features
would be too expensive in terms of computational resources.
Sequential Feature Selection Methods
The most popular group of FS methods make use of a sequential algorithm. In these methods
features are added or removed sequentially based on a certain criteria (the objective function).
Typically there are sequential forward selection (SFS) and sequential backward selection
(SBS) methods. A SFS algorithm begins with an empty set of features and sequentially adds
new features that maximise an evaluation function when combined with the already selected
features (Osuma 2005). A SBS algorithm, on the other hand, commences with a complete set
of all features and sequentially removes a feature, at each step checking the evaluation
function. The main disadvantages of these two classes of algorithms are that the already
selected features cannot be removed after the addition of other features (SFS) and the already
discarded features cannot be reselected (SBS). Moreover, they might be influenced by the
initial condition. The more advanced alternative is the sequential forward floating search
algorithms (SFFS), as proposed by Pudil et al. (1994). In this approach, after each forward
step with an addition of one feature the backward step is implemented to remove one other
feature as long as the evaluated function was improved. The sequential backward floating
search (SBFS) is the backward version equivalent. The SFFS and SBFS algorithms flexibly
adjust the selected feature subset to approximate the optimal solution as closely as possible
(Kavzoglu and Mather 2002).
The steepest ascent (SA) search algorithm was proposed by Serpico and Bruzzone (2001) for
FS in the case of hyperspectral data. In the SA algorithm the solution is represented by a
discrete binary space, consisting of a binary string of ‘1’ and ‘0’ values. The constrained local
maxima in a subspace is searched based on a local maximum value of an evaluated function.
103
The authors also proposed the fast constrained search (FCS) algorithm, which is a
modification of the SA algorithm, with higher computational efficiency. In the study of
Serpico et. al. (2001), it is shown that the FCS always runs faster than or equal to the SA
algorithm. Moreover, the authors claim that the SA and FCS algorithms provided greater
improvements than the SFFS algorithm. Nevertheless, SFFS (or SBFS) algorithm(s) is(are)
still considered to be very efficient and are the most commonly used FS algorithms (Somol et
al. 2006).
Objective Functions
FS for classification can be categorised into either “filter” and “wrapper” methods (Figure
4.2) according to their objective function. The filter method is independent from the
classifiers to be employed and utilises statistical parameters derived from the training
datasets, such as separability indices, as objective functions to evaluate the feature subsets.
The separability indices measure how well the classes separate from each other. The most
widely used separability indices are the Bhattacharyya distance, Transformed Divergence,
and the Jeffries–Matusita distance. The wrapper method, on the other hand, uses the
classification accuracy as an objective function to evaluate the data subsets. In this approach,
in general, the subset of features yielding the best classification accuracy with a pre-defined
classifier will be selected.
The main advantages of the filter method are fast processing with low computational cost,
and high level of generality – as they do not have to be linked with any particular classifier.
However, the major limitations of this approach are that the filter tends to select the whole
dataset as an optimal solution (Osuma 2005), and in general it does not guarantee the highest
classification accuracy (Anthony and Ruther 2007).
The major advantages of the wrapper method include: providing better classification
accuracy than the filter method since they are tailored to the specific classifier, which
interacts directly with the candidate dataset and avoiding over-fitting by employing the cross-
validation techniques for accuracy anticipation. The main disadvantage of the wrapper
approach is slow processing with high computational cost since it has to train a classifier for
each feature subset. Nevertheless, to date, with the availability of powerful computer systems,
the wrapper approach has become more attractive to investigators. The other limitation of this
104
method is its lack of generality since it has to fit a specific classifier (Osuma 2005, Tso and
Mather 2009).
Figure 4.2: Feature selection with filter (left) and wrapper (right) approach (adapted from
Osuma 2005).
Genetic Algorithm
The genetic algorithm (GA) is a robust technique which has been increasingly used for
solving optimisation problems (Tso and Mather 2009). The central concept of the GA
depends upon the biological principle of natural selection and evolution (Anthony and Ruther
2007, Zhoe et al. 2008). In order to obtain the optimal solution the GA carries out a series of
search processes on a set of solutions which is referred as the population. These searching
processes are based on the notion of ”survival of the fittest” as expounded by Darwinian
theory. At the beginning an initial population consisting of various subsets of features is
chosen at random. Each individual in the population will be tested using either the filter or
wrapper approaches. Performance of individuals is evaluated based on the fitness functions
(i.e., objective functions) as specified by users. The individuals which have better fitness
values are then selected as parents to produce children for the next generation. Over
successive generations the GA modifies the population toward an optimal solution based on
105
fitness functions, taking into account operations such as crossover and mutation (Figure 4.3).
The evolution process will be implemented until the fitness threshold or a pre-defined
number of iterations is reached. The crossover operation produces children for the next
generation by combining pairs of parents in the current population. The mutation operation
creates children by employing random changes to an individual parent. Both operations play
important roles in the GA. While the crossover enables the algorithm to explore the new
searched regions, the mutation increases the diversity of a population and consequently
enhances the possibilities that the algorithm converges to the global optimal state (Matlab
Global Optimization Toolbox User’s Guide, 2010).
The feature subset (individual) in the population is often represented by a chromosome. The
common form of chromosome is a binary string of ‘0’ and ‘1’, in which ‘0’ means discarded
features and ‘1’ means selected features.
Figure 4.3: Illustration of the crossover and mutation operators (Huang & Wang, 2006).
The GA can work effectively with a large number of features and has more chance to avoid a
local minimum solution than other techniques. As noted by Goldberg (1989) and Tso and
Mather (2009), the GA can outperform other traditional search techniques since it has some
significant differences from other algorithms, including use of multiple start points for the
search process and ability to deal with special trial structures, its search process is driven by
an objective function(s) and is independent of auxiliary knowledge, and the evolution
between generations is not based on deterministic rules.
FS is effectively used in remote sensing application, including image classification (Kavzoglu
and Mather 2002, Gomez-Chova et al. 2003, Pal 2006, Anthony and Ruther 2007, Zhou
2010, Maghsoudi et al. 2011) to the reduced size and to optimise input datasets and classifier
106
parameters. In the study by Kavzoglu and Mather (2002), the GA and SFS techniques were
applied to select optimal input data features for Artificial Neural Network (ANN)
classification over two test sites in England using multi-date Landsat TM and SPOT HRV
images (site 1) and SIR-C SAR polarised and SPOT HRV images (site 2). The filter approach
was employed using various statistical separability measures and separability indices such as
Wilks’A, Transformed Divergence, and Jefferies-Matusita indices. Although only eight out
of 24 and 23 bands had been selected for the first and second datasets, respectively, the
overall classification accuracy was about 90%, which is close to the overall accuracy of 92%
when all bands were used. The authors claimed that the GA gave better solutions than the
SFS algorithm based on the value of separability indices. However, it is not guaranteed that
the classification based on these solutions would give better results. In other words, the
values of the separability measures from different datasets do not necessarily reflect the
accuracy of the ANN classifier.
Gomez-Chova et al. (2003) evaluated the utility of the SFFS algorithm for crop classification
using HyMap hyperspectral data in the study area of Barrax, Spain. The Bhattacharrya
distance was used as an objective function. The Maximum Likelihood (ML) algorithm was
selected for the classification. The results show that without FS the classification accuracy
did not improve when the number of input bands was greater than 20, and it decreased when
more than 100 bands were used. In a case in which the SFFS algorithm was applied only ten
bands had been selected and gave a similar accuracy as the best results from a non-feature
selection approach. The authors conclude that FS with the SFFS algorithm increased the
classification performance by reducing the Hughes phenomenon.
Pal (2006) investigated the capability of the Support Vector Machine (SVM) technique and
the GA feature selection using DAIS hyperspectral data. The SVM classification accuracy
achieved using a full set of 65 features was 91.76%. The SVM/GA approach obtained
classification accuracies ranging from 91.87% to 92.44%, with 15 features selected –
requiring long computing time. Luo et al. (2008) proposed a combination of GA and Back
Propagation (BP) algorithm for weight training of the ANN-Multi Layer Perception classifier
in classification of land cover features in China using MODIS data. Two other algorithms,
including the Fuzzy ARTMAP ANN and the ML, were also tested. The results show that the
GA/BP integrated method is not only more efficient, but also outperformed the other two
classifiers in terms of classification accuracy. Moreover, the authors claim that the proposed
107
method could help to avoid the risk of premature convergence while the BP network was in
training mode.
Anthony and Ruther (2007) evaluated two FS techniques, including exhaustive search (ES)
and population based incremental learning (PBIL), in conjunction with the SVM algorithm to
classify land cover features in two study areas in Northern Tanzania and Western Cape
Province of South Africa using Landsat TM multi-spectral images. The PBIL is considered a
modified version of the GA in which the probability vector is used instead of the population
for storing possible solutions. The Thornton’s separability index was employed to design the
objective function. The authors claim that both the ES and PBIL are appropriate FS
techniques for SVM classification, though the obtained classification accuracy was not better
than the Non-FS approach. The ES approach required more computing power than the PBIL
method.
Zhou et al. (2010) applied the GA and SVM to classify PHI hyperspectral data over the
Shanghai World Exposition Park, Shanghai, China. Results illustrate that the SVM-GA
method provided better classification accuracy than the ML and ANN-MLP classifiers and
that the integration of SVM and GA could improve the classification accuracy of
hyperspectral remote sensing data, particularly for the classification of vegetation classes.
However, according to the authors, the SVM-GA method was not superior to other methods
for classification of classes consisting of various textured information, such as cement floors
or car roofs.
The above literature review indicates that FS is a very effective technique for improving
classification performance. It has been applied mainly for hyperspectral data where the data
dimensionality is very high (Gomez-Chova et al. 2003, Pal 2006, Luo et al. 2008, Zhou et al.
2010). The usefulness of this approach has also been illustrated in some studies with smaller
datasets. However, in the case of multisource data, where the combined datasets could be as
large as hyperspectral data and with a much higher level of complexity, the application of FS
has not been adequately investigated.
108
4.3. APPLICATION OF GENETIC ALGORITHM AND SUPPORT VECTOR MACHINE IN CLASSIFICATION OF MULTISOURCE REMOTE SENSING 4.3.1. INTRODUCTION
As has been discussed in the previous section, the use of FS techniques for classification of
highly dimensional remote sensing data is an effective approach. On the other hand, the non-
parametric classifiers such as ANN or SVM are robust for classifying large, complex
datasets. Thus, integration of the FS techniques and non-parametric classifiers should be
investigated in order to improve the classification of multisource remote sensing data. In this
section the results of the integration of FS techniques based on GA and SVM are presented.
The SVM algorithm was selected because of its non-parametric nature, its robustness and its
recent development. The literature and investigations described in the previous chapter, have
shown that the SVM often provided better (or at least same level of) classification accuracy
than other algorithms (Waske and Braun 2007). It is also worth noting that although in
general, the SVM algorithm works well with high dimensional data, without reducing the
dimensionality of the feature space because its nature is based on maximising margin
between classes, it has been demonstrated by numerous studies that even this classifier
benefits from the dimensionality reduction of the feature space (Guyon et. al. 2002, Pal
2006). Among many FS techniques that have been used, the GA has proven to be very
effective for handling global optimisation problems with large datasets, and has less chances
of converging to a local optimal solution than other methods (Huang and Wang 2006, Zhuo et
al. 2008, Zhou et al. 2010). Moreover, the accuracy and efficiency of the SVM classifier
depends on both the input datasets and the kernel parameters. While other methods can only
deal with a single issue at a time, the GA techniques can find the optimal feature subset and
kernel parameters at the same time.
In this study the FS with wrapper approach was selected because of its capability in
producing high classification accuracy. The FS wrapper approach based on SVM-GA method
has been applied successfully in many applications, including biology, medical and financial
data analysis (for example, Huang and Wang 2006, Osowski et al. 2009, Zhuo et al. 2008, Pal
2006). In the field of remote sensing a few studies have been conducted for classification of
hyperspectral data (Pal 2006, Zhuo et al. 2008, Zhou et al. 2010). However, the application of
the SVM-GA method for classifying multisource remotely sensed data has not been
109
previously reported in the literature. Therefore, the SVM-GA model was implemented to
classify multiple combined datasets, consisting of Landsat 5 TM, multi-date, dual-
polarisation ALOS/PALSAR images and their multi-scale textural information. The
performance of the proposed method was compared with that of the traditional stack-vector
approach and the FS wrapper approach based on the SVM classifier coupled with the SFFS
algorithm (SVM-SFFS). A large number of different combined datasets were generated and
classified. The objectives of this study were: 1) to evaluate the integration of optical, SAR
satellite images, and their textural information for land cover mapping; and 2) to propose and
implement an approach based on the combination of FS with GA techniques and SVMs for
classifying multisource remotely sensed data.
4.3.2. STUDY AREA AND DATA USED
The study area was located in the state of Western Australia, Australia, with centre coordinate
116o 57’ 45’’ E; 33o 48’ 40’’ S. The site is characterised by relatively flat terrain with
pastures, crops, sparse and dense tree cover. There are also some small rural residential
settlements in the south of the study area. Two kinds of satellite imagery were employed for
this study (Figure 4.4): four ALOS/PALSAR HH/HV dual-polarisation images acquired in
2010 (Table 4.1), and a Landsat 5 TM image acquired on 07/10/2010 with spatial resolution
of 30m – in this study five spectral bands were used.
Table 4.1: ALOS/PALSAR images of the Western Australia study area.
Satellite/Sensor Path Acquisition dates Polarisation Orbit
ALOS/PALSAR 433
20/07/2010 HH/HV Ascending
04/09/2010 HH/HV Ascending
20/10/2010 HH/HV Ascending
05/12/2010 HH/HV Ascending
110
Figure 4.4: Landsat 5 TM (left) and ALOS/PALSAR HH (right) over the study area acquired
on 07/10/2010 and 20/10/2010, respectively.
4.3.3. METHODOLOGY
4.3.3.1. Data processing
Landsat 5 TM images had been already processed to the Standard Terrain Correction (Level
1T) product by the United State Geological Survey (USGS). At this processing level, the
systematic radiometric and geometric corrections were undertaken using the platform
geometric model which employs both GCPs and DEM for removing the topographic
displacements. In this process, the Landsat 5 TM images were geo-rectified to the map
coordinate (UTM projection, WGS84 datum). The geo-rectification of ALOS/PALSAR
images and the co-registration of optical and SAR images were carried out as described in the
section 3.1.2. ALOS/PALSAR and Landsat 5 TM images were resampled to 10m pixel size.
Speckle noise in the PALSAR images was filtered using the Enhanced Lee Filter (Lopes et al.
1990) with a 5x5 window size. SAR backscatter values were converted to decibel using
Equation (3.1).
111
Besides the original ALOS/PALSAR images, several derivative images were also generated
and integrated for classification, including the Temporal Backscattered Change, average and
dual-polarised SAR index images. The Temporal Backscattered Change (SARCH) image was
generated using all four different SAR images (Equation (3.13)). This image highlights
differences between the stable or non-changing features (such as urban areas, permanent
vegetation, or still water) and temporally changing features (such as annual crops). The SAR
polarised index was computed:
(4.1)
where SARind is the SAR polarised index image derived from the corresponding SAR dual
HH/HV polarised image.
Two groups of texture measures were extracted and employed for classification, namely “first
order” and “second order” texture measures. The first order texture measures involve
statistics computed directly from the original image and do not model relationships with
neighbouring pixels. On the other hand, the second order texture measures consider the
mutual dependence of sets of surrounding pixels (Coburn and Roberts 2004, Hall-Beyer
2007). The most widely used second order textural data is the Grey Level Co-occurrence
matrix (GLCM), which measures relationships between pairs of pixels within a
neighborhood. The First Principal Components (PC1) images generated from each of the four
SAR and Landsat 5 TM images were used to derive textural information. In order to reduce
correlation within datasets it was necessary to select only components which are less
correlated to each other. Hence only three first order texture measures, namely mean,
variance and data range, and four GLCM texture measures, variance, homogeneity, entropy
and correlation, were employed. Since there is no preferred direction, the GLCM texture
measures were computed as an average of texture measures generated for the eight different
directions 0o, 45o, 90o, 135o, 180o, 225o, 270o, 315o. Texture measures were calculated using
the relations below. As the multi-scale texture approach was adopted, textural data were
generated from eight window sizes, including 3x3, 5x5, 7x7, 9x9, 11x11, 13x13, 15x15,
17x17, and used simultaneously.
First order texture measures (F_OR) were generated using the following relations:
HVHHHVHH
+−
=indSAR
112
∑=
×=n
kifi
WMean
0
1 (4.2)
( ) i
n
ifMean
WVariance ×−= ∑
=
2
011 (4.3)
( ) ( )iMiniMaxrangeData −=_ (4.4)
where fi is frequency of pixel’s value i appearing in a moving window and W is the total
number of pixels in a moving window, n is the quantisation level of the digital image (Jensen
2004). The second order GLCM texture measures were computed based on Equations (3.2 –
3.5).
4.3.3.2. Integration, feature selection, parameter optimisation and classification
SVM Classifier
In this study the SVM classifier with the Radial Basic Function (RBF) kernel was applied to
classify land cover features. Two parameters were optimally specified in order to ensure the
best accuracy: the penalty parameter C and the width of the kernel function γ. The common
means of determining the optimal C and γ is using a grid search algorithm (Kavzoglu and
Colkesen 2009).
Genetic Algorithm
The GA model for FS and parameter optimisation involves designing chromosomes,
definition of the fitness function, and specification of the system architecture.
Chromosome Design
The two parameters C and γ need to be defined. The GA-based model will try to optimise
both input features and the SVM’s parameters. Thus, the chromosome consists of three parts,
representing selected features, C and γ. A binary coding technique was used to define the
chromosome (Huang and Wang 2006). In Figure 4.5, Ib1 to IbNf represent input features,
Ibi=1 means a corresponding feature is selected, Ibi=0 means a feature is not selected. Cb1 ~
CbNc represents the value of C and γb1 ~ γbNγ represents the value of γ.
113
Figure 4.5: The binary coding of the chromosome.
Fitness Function
The fitness function is designed to test whether an individual is ‘fit’ for reproduction. The
chromosomes that have the higher fitness value will have more chance to be selected as
parents or selected for recombination in the next generation. Two criteria which are often
used for designing the fitness functions are classification accuracy and the number of selected
features (Huang and Wang 2006, Zhuo et al. 2008, Zhou et al. 2010, Pal 2006). In this study,
a modified design of the fitness function with an additional criterion, namely, average
correlation is proposed:
(4.5)
where OASVM is the overall classification accuracy (%), WOA represents the weight for the
classification accuracy, Wf represents the weight for the number of selected features, NS is the
number of selected features, and N is the total number of input features. Cor is the average
correlation coefficient of selected bands. The values of WOA and Wf can be set by the user.
The design of this fitness function aims to select combined datasets which can provide high
classification accuracy, and consists of a small number of features with a low level of
correlation within the dataset.
The major steps for SVM-GA feature and parameter selection are (Figure 4.6):
1) Randomly create chromosomes of the initial population.
2) Calculate the fitness value of each individual in the population. This step involves
converting the binary code of the chromosomes to identify C, γ and the selected features. The
SVM classifier will implement the classification based on these values and the training
datasets. The fitness value of the individual is calculated by Equation (4.5).
3) In the reproduction step a number of individuals with a high fitness value will be selected
and kept for the next generation. The other individuals will be used for the crossover and
mutation process to generate new children for the next generation.
CorNNW
OAWFitness ××+×=
s100f
SVMOA
114
4) If the stopping condition is satisfied the evolution terminates, and the optimal result
represented by the best individual is returned, otherwise the evolution will continue. The
stopping criterion is usually a predefined change of the fitness values or the maximum
number of desired generations.
Figure 4.6: Major steps of feature selection using the Genetic Algorithm.
Different combined datasets were generated, including Landsat 5 TM + its textures, SAR
single- and dual-polarised images + their textures, SAR dual-polarised images + intermediate
derived images, Landsat 5 TM + SAR single-/dual-images, Landsat 5 TM + SAR single-
/dual-images + textures and intermediate images. These integrated datasets were classified
using the SVM classifier with the RBF kernel.
Non-feature selection (Non-FS) and FS approaches were carried out and the results were
compared. For the Non-FS, the conventional stack-vector approach was applied. The stack-
vector approach is the most straightforward approach, where the data are added together as
input in the classification process. The stack-vector approach was applied for all datasets
including the original images and combinations. In the FS approach, data were selected to
form datasets which give the best solution based on the GA. The FS approach based on the
SFFS strategy was also carried out for comparison purposes. The FS approach was only
applied for the complex datasets (more than 12 input features), where the GA techniques
were employed in order to optimise input data and the SVM’s parameters at the same time.
115
The chromosome and the fitness function for the GA were designed as shown in Figure 4.5,
and Equation (4.5). In this study the weight for classification accuracy (WOA) and the weight
for the number of selected features (Wf) were set in the range 0.65-0.8 and 0.2-0.35,
respectively. The other parameters for the GA were:
Population size = 20-40; Number of generations = 200; Crossover rate: 0.8; Elite count:
3-6; Mutation rate: 0.05.
The 5-fold cross validation techniques were used to estimate the accuracy of each classifier.
The grid search algorithm was applied in the stack-vector approach to search for the best
parameters (C, γ) for the SVM classifiers. The GA was implemented using the Global
Optimization toolbox in Matlab 7.11.1(2010b). The SFFS algorithm was also developed in
Matlab 7.11.1. For the implementation of the SVM classifiers the well-known LIBSVM 3.1
toolkit with Matlab interface was employed (Chang and Lin 2011).
Five land cover classes were identified for classification: Crop (CR), Permanent Pastures
(PA), Dense Forest (DF), Sparse Forest (SF) and Residential Area (RE). The spectral
properties and backscatter signatures of the five land cover features in the Landsat 5 TM
multi-spectral and multi-date PALSAR images are shown in Figure 4.7.
Figure 4.7: Land cover feature characteristics in Landsat 5 TM (left) and multi-date PALSAR
(right) images.
The land cover data used for training and validation were derived from visual interpretation
with the aid of aerial photography, Google Earth images and ground truth data collected from
two field surveys conducted on September 2, 2010 and August 23, 2010. The training and test
datasets were selected randomly and independently using the Region of Interest (ROI) tool of
the ENVI 4.6 software. The samples selection has been carried out based on the procedure
116
described earlier in the section 3.1.2. The sizes of the training and testing datasets are listed in
Table 4.2.
Table 4.2: Contents of training and testing datasets.
Classes Training data
(number of pixels)
Testing data
(number of pixels)
Crop 954 1076
Pasture 956 1107
Spare Forest 961 1080
Residential 313 198
Dense Forest 632 569
4.3.4. RESULTS AND DISCUSSION
Overall classification accuracies of datasets (consisted of less than 12 features) using the
Non-FS (stack-vector approach) and SVM classifiers are summarised in Table 4.3. Results of
classification using the proposed SVM-GA FS technique, SVM-SFFS and the Non-FS
approach for the larger datasets are given in Table 4.4.
The classification results demonstrate the complimentary characteristics and efficiencies from
the integration of optical and SAR images. All combined datasets generated from both kinds
of data, no matter whether the Non-FS or FS approach was used, produced significant
improvements in classification accuracy compared to the original single datasets. In the case
of the stack-vector approach, although the Landsat 5 TM gave a high accuracy of 85.21%, it
exhibited extensive confusion between residential and vegetation classes (commission errors
were 56.39%). The combination of the original Landsat 5 TM and PALSAR images gave
remarkable increases of accuracy, in which the combined use of Landsat TM and four-date
PALSAR HH images resulted in a classification accuracy of 91.46%, with improvements of
6.25% and 25.60% for Landsat 5 TM and PALSAR HH data, respectively, while the
commission errors for the residential class were 29.29%. The integration of Landsat 5 TM
image with four-date PALSAR dual HH/HV images gave an overall accuracy of 88.93% with
an increase of 3.72% and 16.18% over Landsat 5 TM and PALSAR dual HH/HV data,
respectively, while the commission errors of the residential class were 39.94%. However, the
improvements are more obvious when using the GA-based FS approach. This integration
resulted in an overall accuracy of 92.26% with an increase of 7.05% and 19.51% compared to
the single-type datasets, while the residential commission errors were reduced to 19.83%.
117
Table 4.3: Classification accuracy of different datasets using SVM classifiers with the traditional stack-vector approach.
Datasets Overall accuracy (%)
Four-date HH images 65.86
Four-date HH images + HHAVE+HHCH 67.92
Four-date HV images 66.58
Four-date HV images + HVAVE+HVCH 68.44
Four-date HH/HV images 72.75
Four-date HH/HV images + HH/HVAVE + HH/HVCH 73.18
Four-date HH/HV images +SARind 72.73
Four-date HH images 65.86
Four-date HH/HV images + HH/HVAVE+HH/HVCH +SARind 73.20
L5 TM 85.21
(note: L5 TM = Landsat 5 TM; Four-date HH image = Four-date PALSAR HH polarised image; Four-date HV image = Four-date PALSAR HV polarised image; Four-date HH/HV
image = Four-date PALSAR dual HH/HV polarised image).
The complimentary properties of like- (PALSAR HH) and cross- (PALSAR HV) polarisation
images were also clearly highlighted. As can be seen in Table 4.4, except for the crop where
the classification accuracy of PALSAR HH and HV images are rather similar, the like-
polarisation gave better accuracy for the residential class (which is dependent on surface
scattering), while the cross-polarisation, due to its sensitivity to the volume scattering,
provided higher accuracy for vegetation classes Pasture, Sparse Forest and Dense Forest.
Utilisation of both SAR like- and cross-polarised data resulted in noticeable improvements in
overall classification accuracy, particularly for the Residential, Sparse Forest and Dense
Forest features.
Table 4.4: Producer and user accuracy (%) of four-date PALSAR HH, HV and dual-polarised
images.
Land cover classes Four-date HH Four-date HV Four-date HH/HV
Producer User Producer User Producer User
Crop (CR) 80.11 65.80 78.07 69.19 77.79 69.35
Permanent Pasture (PA) 62.87 76.48 66.03 78.69 66.58 75.51
Sparse Forest (SF) 59.63 70.46 70.74 67.02 78.24 77.17
Residential (RE) 68.18 75.84 16.16 12.90 77.78 76.24
Dense Forest (DF) 55.71 44.15 55.54 63.33 63.09 65.27
118
Incorporation of the PALSAR original images with their additional derived data, such as
Average and Temporal Backscattered Change (SARCH), gave an increase in classification
accuracy. The improvements were 2.06% and 1.86% for the case of four-date PALSAR HH
and HV polarised images, respectively. However, in the case of PALSAR dual HH/HV
polarised images, there was only a very slight increase in accuracy of 0.43%. On the other
hand, the SAR polarised index images do not give any improvement compared to the
classification of the original PALSAR dual HH/HV polarised images.
Integration of optical and multi-date SAR data with their textural information gave a
noticeable increase in classification accuracy. As for the Non-FS approach, the combination
of four-date HH images with their textures gave very slight increases in accuracy of 0.81%,
0.92% and 1.34% using the first order, GLCM and both groups of texture measures,
respectively. On the other hand, the increase of 4.41%, 0.79%, and 6.79% in classification
accuracy was obtained for the four-date HV images. The improvements for the four-date
PALSAR dual HH/HV polarised images were 4.42%, 1.86% and 6.05%. The application of
textural information was also effective for optical images. The integration of Landsat 5 TM
images with their textures resulted in increases of 2.48%, 4.37% and 2.95% using the first
order, GLCM and both textural groups, respectively.
The efficiency of using textural information was even more impressive when the SVM-GA
methods were exploited. For example, the improvements for the four-date PALSAR dual
HH/HV polarisation increased by 6.98%, 2.95% and 6.93%. The combination of Landsat 5
TM images with their corresponding texture measures also gave significant improvements of
5.53%, 6.00% and 6.23%. In particular, the integration of four-date PALSAR HV polarised
images with both types of textures resulted in an increase of overall classification accuracy of
up to 12.95%.
It is worth mentioning that although the four-date PALSAR dual HH/HV images resulted in
much lower classification accuracy, 72.75% compared to 85.21% for the Landsat 5 TM
image, its combination with textural data and additional images using the SVM-GA method
gave an overall classification accuracy of 81.19%, which is closer to the accuracy obtained
using optical images. A comparison of classification improvement by incorporation of
textural information using stack-vector and FS-GA approach is shown in Figure 4.8.
119
Figure 4.8: Improvement of accuracy by incorporating textural information with original
datasets using stack-vector and FS-GA approach.
(F_OR(ST), GLCM(ST), F_OR+GLCM(ST) and F_OR(GA), GLCM(GA), F_OR+GLCM(GA) represent integration of textures using stack-vector and FS-GA approach,
respectively). The FS-GA approach clearly outperforms the commonly used Non-FS approach (Table 4.5
and Figure 4.10). In all cases the FS-GA approach gave better results than the Non-FS
approach. The increase of overall classification accuracy ranged from 0.87% (four-date
PALSAR HH/HV images + its first and second order textures) to 7.57% (four-date PALSAR
HV images + its first order textures). It is also important to emphasise that the FS-GA used
less input features than the stack-vector approach, see Figure 4.11.
While in many cases the Hughes phenomenon appeared in the stack-vector method, where
the classification accuracy decreased with an increase in the number of input measurements,
it is not the case for the SVM-GA method. For example, the integration of the Landsat 5 TM
image with PALSAR HH images gave an overall accuracy of 91.46%, while the integration
of the Landsat 5 TM image with PALSAR dual HH/HV images resulted in 88.93% accuracy,
a decrease of 2.53%. However, using the SVM-GA method the same integration gave an
overall accuracy of 92.26% with only six out of a total 13 features used (Figure 4.9).
Similarly, the combination of Landsat 5 TM image with its GLCM textures gave an accuracy
of 89.58%, but when the Landsat 5 TM image was combined with both first order and GLCM
texture measures the accuracy dropped to 88.16%. Nevertheless, when applying GA
techniques the classification accuracy still increased slightly to 91.44%.
120
The highest classification accuracy achieved with the Non-FS methods was 93.47% with 181
data input features, while the GA achieved the best accuracy of 96.47% with only 81 selected
features.
Figure 4.9: Classification of the Landsat 5 TM image (left) and the integration of Landsat 5
TM and four-date PALSAR HH/HV images (right) using the GA technique.
The FS approach based on the SVM-SFFS method can be considered more efficient than the
Non-FS approach. This approach provided higher classification accuracy than the stack-
vector approach in 14 cases, while the stack-vector approach gave better accuracy in the
remaining six cases. However, among of these six cases there are only two cases where the
SVM-SFFS produced decreases of more than 1% in classification accuracy compared to the
Non-FS approach. Moreover, the SVM-SFFS method used far less input features than either
the stack-vector or SVM-GA methods.
The comparison between two FS approaches using SVM-GA and SVM-SFFS methods shows
that the SVM-GA method also performed better than the SVM-SFFS method. The SVM-GA
method provided significantly higher classification accuracy than the SVM-SFFS method in
the classification of 19 out of 20 combined datasets. There was only one case (four-date
PALSAR dual polarised images + its first and second order textural information) where the
SVM-SFFS method gave a slightly higher accuracy of 0.47% compared to the SVM-GA
method. The highest classification accuracy achieved by the SVM-SFFS method was 93.0%,
121
while it was 96.47% for the SVM-GA method. However, the SVM-SFFS method selected
only five input features, and the number of input feature was 81 in the case of the SVM-GA
method. It is also revealed that the problem of Hughes’s phenomenon was not fully resolved
by the SVM-SFFS method. In many cases that the higher dimensional datasets did not result
in increases in classification accuracy.
Table 4.5: Comparison of land cover classification performance between SVM-GA, SVM-SFFS and stack-vector approach; nf = number of selected features.
ID DATASETS SVM OVERALL ACCURACY (%)
Stack Vector SFFS GA
NF Accuracy NF Accuracy NF Accuracy
1 Four-date HH images + First- order textures 28 66.67 10 66.67 8 68.24
2 Four-date HH images +GLCM textures 36 66.77 10 68.06 19 69.08
3 Four-date HH +First – order & GLCM textures 60 67.20 12 64.99 21 70.59
4 Four-date HV + First-order textures 28 70.99 8 71.59 8 78.56
5 Four-date HV +GLCM textures 36 67.37 12 68.14 16 73.60
6 Four-date HV +First - order & GLCM textures 60 73.37 6 78.66 25 79.53
7 Four-date HH/HV images + First- order textures 56 77.17 5 78.01 23 79.73
8 Four-date HH/HV images +GLCM textures 72 74.61 12 73.13 34 75.70
9 Four-date HH/HV +First -order & GLCM textures 120 78.81 6 80.05 51 79.68
10 Four-date HH/HV images + HH/HVAVE + +HH/HVCH
+ SARind + First-order textures 64 77.15 11 78.21 24 80.23
11 Four-date HH/HV images + HH/HVAVE + +HH/HVCH +
SARind + GLCM textures 80 74.17 11 74.14 36 75.65
12 Four-date HH/HV images + HH/HVAVE + HH/HVCH +
SARind + First-order & GLCM textures 128 78.86 7 81.02 49 81.19
13 L5 TM + First-order textures 29 87.69 4 87.07 9 90.74
14 L5 TM + GLCM textures 37 89.58 5 90.02 13 91.21
15 L5 TM + First-order & GLCM textures 61 88.16 6 87.94 20 91.44
16 L5 TM + Four-date HH/HV 13 88.93 4 90.05 6 92.26
17 L5 TM + Four-date HH/HV + HH/HVAVE + +HH/HVCH
+ SARind 21 88.39 7 90.89 8 93.10
18 L5 TM + All of LS5_textures + Four-date HH/HV +
HH/HVAVE+HH/HVCH + SARind 77 92.21 5 92.48 32 95.24
19 L5 TM + All of L5 TM textures + Four-date HH/HV +
All of SAR textures 181 93.47 6 92.51 81 96.20
20 L5 TM + All L5 TM textures + Four-date HH/HV +
HH/HVAVE + HH/HVCH + SARind + All of SAR textures 189 92.98 5 93.00 81 96.47
122
Figure 4.10: Classification accuracy using FS (SFFS and GA) compared to Non-FS (stack-
vector) approaches.
Figure 4.11: Number of input features for the Non-FS (stack-vector) and FS (SFFS and GA)
approaches.
123
The proposed fitness function with an additional parameter of average correlation between
selected features has improved the overall classification accuracy for 18 out of 20 combined
datasets compared to using the commonly designed fitness function – for the remaining two
cases the classification accuracy decreased very slightly (about 0.1%). Impact of the proposed
fitness function compared to the commonly used fitness function is illustrated in Figure 4.12.
Figure 4.12: Impact of the proposed fitness function on overall classification accuracy
compared to the commonly used fitness function.
4.3.5. CONCLUSIONS
A feature selection approach based on the SVM-GA method has been proposed and
compared with non-feature selection approach and the SVM-SFFS method for classification
of multisource remotely sensed data, including optical, multi-date SAR and textural
information. Results of classification of different combined datasets (more than 30 for non-
feature selection and 20 for the feature selection approach) revealed advantages of
multisource remotely sensed data and SVM-based algorithms for land cover classification.
Incorporation of textural information with either optical or SAR data also resulted in an
improvement in accuracy. Feature selection using the SVM-GA method clearly outperformed
the classical stack-vector approach and the SVM-SFFS methods. The feature selection
124
approach based on SVM-GA method always resulted in better classification accuracy with
more significant improvement, and used less data input features, compared to the non-feature
selection approach. The SVM-GA method also provided better classification accuracy than
the SVM-SFFS method for 19 out of 20 combined multisource datasets. The highest overall
classification accuracy of 96.47% was obtained using the SVM-GA method for classifying
the combined dataset of original Landsat 5 TM, four-date PALSAR dual HH/HV polarised
images, all of their textures and additionally derived images with 81 out of 189 data input
features selected. Moreover, classification results also indicated that the proposed fitness
function in this study is more reliable than the commonly used version. Although the SVM-
SFFS method did not provide as high classification accuracy as the SVM-GA method it was
also considered a viable alternative which performed noticeably better than the traditional
stack-vector approach for 14 out of 20 experiment cases. However, the highest classification
accuracy achieved by the stack-vector approach (93.47%) was slightly better than the SVM-
SFFS method (93.0%). The most distinct advantage of the SVM-SFFS method is that it used
a far smaller number of input features than either the SVM-GA method and non-feature
selection approach.
125
CHAPTER 5
APPLICATION OF MULTIPLE CLASSIFIER SYSTEM
FOR CLASSIFYING MULTISOURCE REMOTE
SENSING DATA
The effectiveness of non-parametric classifiers such as Support Vector Machine (SVM) and
Artificial Neural Network (ANN) algorithms, and the advantages of using Feature Selection
(FS) techniques based on the Genetic Algorithm (GA) for classifying multisource remote
sensing data has been demonstrated in the previous chapters. Recently, a Multiple Classifier
System (MCS) has been increasingly used for remote sensing, particularly for handling large
volume and complex datasets (Yan and Shaker 2011). The goal of this chapter is to further
enhance the performance of land cover classification using multisource data by exploring the
capabilities of MCS, which also has the potential to improve the quality of image
classification. The main contribution of this chapter is to propose and to evaluate an
integration of MCS and FS based on GA techniques to resolve the challenge of classifying
highly dimensional combined datasets.
5.1. MCS IN REMOTE SENSING MCS or classifier ensemble has been proposed as a robust and effective method to increase
the classification accuracy (Briem et al. 2002, Bruzzone et al. 2004, Du et al. 2009b, Yan and
Shaker 2011). The MCS can be defined as a system that obtains the final classification results
by merging the outputs of multiple classifiers based on a certain combination rule (Xie et al,
2006, Du et al. 2009a). In general there is no perfect classifier which can give the best
performance in all cases. The classifier which is very effective for one dataset might not be
appropriate for other datasets, and therefore the MCS can provide the additional information
about classification patterns. Consequently, the MCS may outperform an individual classifier
by integrating the advantages of various classifiers. It is also worth mentioning that the
classifier ensemble certainly minimises the risk of poor selection (Polikar, 2006). It is also
126
claimed that the MCS can take advantage of complementary characteristics of multi-sensor
data for classification (Benediktsson et al. 2007).
5.1.1. CREATION OF MCS OR CLASSIFIER ENSEMBLE
In the MCS it is essential to maintain the diversity between the member (or base) classifiers
since if they are identical the combination will exhibit no improvement. It is suggested that
the member classifiers have different decision boundaries (Polikar, 2006). The diversity in
the ensembles is introduced to the system by using different classifiers, or by training
classifier with different parameters and different training datasets. The ensemble of classifiers
can be generated in different ways, including combination of different classifiers or
combination of the same classifiers with various versions of input training data. The
commonly used techniques of “boosting” and “bagging” are based on manipulating input
training sample data (Briem et al. 2002).
“Bagging” is an abbreviation of Bootstrap Aggregating which was introduced by Breiman
(1994). In this approach many training sample subsets (or “bags”) are generated by
employing the random and uniform selection with replacement of m samples from a training
sample set of size m (the same size). The classification process is then carried out based on
training data on each bag using a base classifier, resulting in various outputs. The final result
is constructed by combining all individual outputs. The bagging algorithm can noticeably
improve the classification accuracy with unstable base classifier. However, the classification
accuracy may be decreased in the case of stable classifiers since less training data are used for
each classifier (Tzeng et al. 2006, Briem et al. 2002, Briem et al. 2001).
The “boosting” algorithm was proposed by Schapire (1990) in order to increase the
classification accuracy of a weak classifier. At the beginning, the classifier is trained by
samples which have the same weight. Then, during the training stage, the boosting algorithm
successively reweights the training samples so that the misclassified samples are weighted
higher than the correctly classified samples. The next classifier in the ensemble is trained by
the newly modified samples. As the number of iterations increase, the weights of correctly
classified samples are decreased. Thus, the classifier is driven to concentrate on “difficult”
samples (Briem et al. 2002), and the bias and variance of the classification can be reduced
(Waske and Braun 2009). It is worth noting that, while the bagging can be operated in
127
parallel with all of the bags, boosting is implemented in an iterative procedure on the samples
with the newest weights. As a result the boosting algorithm requires more computational
power and longer processing time. The commonly used boosting algorithms in the field of
pattern recognition and remote sensing are AdaBoost.M1 and M2 methods (Freund and
Schapire, 1996).
5.1.2. COMBINATION RULES
The performance of the ensemble classification is also a function of the methods used to
combine the classification outputs. A range of methods are available depending on the type of
information derived from each classifier. There are three levels of outputs from the
classifiers: abstract level, rank level and measurement level (Foddy et al. 2007, Maghsoudi et
al. 2006). At the abstract level, the output from each classifier is a label, and the final labels
are decided by combining output labels of multiple classifiers. The most commonly used
technique for combining outputs of classifiers is by majority voting. At the rank level, the
output of a classifier is the possibility of a pixel belonging to each class given in a rank. The
rank level combination algorithm is used to determine the final class for each pixel. At the
measurement level, the output of the member classifier can be a quantitative value (e.g.,
separability, distance) or the probability of a pixel belonging to a certain class. In this case,
the quantitative index is employed to combine multiple classifiers, as a measurement level
combination.
The most widely used method to integrate the measurement level outputs is using Bayesian
decision rules, which linearly combine outputs of the individual classifiers. The main
Bayesian decision rules are Sum, Product, Maximum and Minimum:
Sum rule
)|(1
j
n
jii xwpSumP ∑
=
= i = 1, 2, 3,..., k. (5.1)
Product rule
)|(Pr1∏=
=n
jjii xwpoP i = 1, 2, 3,..., k. (5.2)
Max rule
)|(1 jinji xwpMaxMaxP == i = 1, 2, 3,..., k. (5.3)
128
Min rule
)|(1 jinji xwpMinMinP == i=1, 2, 3, …, k. (5.4)
where p(wi|xj) are the posteriori probabilities which represent the probability that the jth
classifier assigns the pixel x to the class wi. In each combination rule the winner is the class
with the largest value according to this rule.
Non-linear combination rules, such as evidence reasoning based on the Dempster-Shafer
theory of evidence and fuzzy integral, were also applied and provided very positive results
(Du et al. 2009a, Trinder et al. 2010, Salah et al. 2010). In the evidence reasoning algorithm,
the single data source is weighted in accordance with their importance or reliability. The
combination rules are generated upon the concepts of belief and plausibility in order to assign
labels to pixels (Mather 2010). Tso and Mather (2009), and Trinder et al. (2010) describe
uses of the evidence reasoning method for the classification of remote sensing data.
Some basics concepts, recent developments and application of MCS in remote sensing have
been reviewed by Benediksson et al. (2007). Briem et al. (2002) evaluated the performance of
bagging, boosting and Bayesian maximum rule on weighted linear and weighted products of
a posteriori probability combination for classifying several multisource remote sensing
datasets. Five classifiers, namely Minimum Eucledian Distance (MED), Maximum
Likelihood (ML), ANN-MLP with conjugated-gradient algorithm, Decision Table, Decision
Tree and simple 1R classifier (using only one feature when determining a class) were
employed. Results illustrated that all multiple classification strategies gave better
classification accuracy than a single classifier. The boosting approach with the AdaBoost.M1
method appeared to be the most accurate classifier for both training and testing datasets. The
authors claim that multiple classification methods can be considered as a valuable alternative
for the classification of multisource data. Tzeng et al. (2006) applied the MCS with the
bagging and boosting algorithms to classify land cover features in Taiwan using multispectral
data similar to SPOT satellite images. The ANN-MLP classifier was used as a base classifier.
In this study the original bagging and boosting methods had been modified by introducing the
concept of confidence index with a view to reduce the confusion among classes. The results
show that both the original and modified MCS significantly improved classification accuracy
compared to the single classifier. Foody et al. (2007) integrated five classifiers based on the
majority voting rule for mapping fernland in East Anglia, UK. The accuracy achieved by the
129
ensemble approach was 95.6%, which is marginally lower than the most accurate single
classifier (96.8%). However, the authors claim that this reduction in classification accuracy is
insignificant and that it is difficult to identify the best individual classifier at the beginning.
Du et al. (2009b) used a different combination approach, including parallel and hierarchical
classifier systems, training samples manipulation with bagging and boosting techniques for
classifying hyperspectral data. Salah et al. (2010) employed the Fuzzy Majority Voting
techniques to combine classification results of three classifiers over four different study areas
using Lidar and aerial images. The results illustrated that while the overall classification
accuracies were slightly improved, the commission and omission errors were reduced
considerably compared to the best individual classifier. Abe et al. (2010) investigated the
performance of the MCS with a SVM base classifier on Landsat and AVIRIS hyperspectral
data in two test sites in South Africa. It was revealed that: (1) in each instance, the MCS
results were better than the majority of the base classifier’s predictions; (2) classification
accuracy increased as the number of features in the base classifier increased; and (3) there
was no significant benefit in having many base classifiers, and that three base classifiers are
sufficient for ensemble classification. However, in this study the ensemble classification
result was not better than the best result obtained from member classifiers in the ensemble.
This result agreed with the previous studies by Foody et al. (2007). Additionally, Yan and
Shaker (2011) evaluated the effects of MCS on classification of SPOT 4 images using
different Bayesian decision rules. In this study, five base classifiers including Contectual
Classifier, k-nearest Neighbour, Mahalanobis, ML and Minimum Distance classifiers, were
applied to classify six land cover classes in Hong Kong. The authors pointed out that: (1)
significant diversity amongst the base classifiers cannot enhance the performance of classifier
ensembles; (2) accuracy improvement of classifier ensembles is only possible by using base
classifiers with similar or lower accuracy; (3) increasing the number of base classifiers cannot
improve the overall accuracy of the MCS; and (4) none of the Bayesian decision rules
outperforms the others.
The examples mentioned above suggest that although the concept of the MCS is robust, the
improvements in classification performance on remote sensing data, including multisource
data, by applying this technique are still not clear. The recognisable benefits of the MCS are
reducing the impact of poor selection of input parameters and providing a classification result
which is better than the result of the worst constituent classifiers. Moreover, in most of the
previous studies the individual classifiers were applied directly to the input datasets, and then
130
the MCS techniques were used to combine results of these classification processes. Only a
few authors (e.g. Abe et al. 2010) used feature selection (FS) techniques to search for relevant
input features before implementing the classification process. In the study of Abe et al.
(2010), different input datasets for the SVM classifier were selected after an exhaustive
search using a filter approach and the Bhattachryya distance separability index as an
objective function. However, as has been discussed in chapter 4, there are some limitations
that affect the performance of this approach: 1) the exhaustive search is not suitable for
handling high dimensionality datasets; 2) the filter approach does not guarantee a high
classification accuracy; and 3) the separability indices do not directly reflect the classification
accuracy. Therefore, in the next section, a more advanced method for integrating FS wrapper
approach based on the GA technique and MCS for classifying multisource remote sensing
data is proposed.
5.2. USE OF MCS AND A COMBINATION OF GA AND MCS FOR CLASSIFYING MULTISOURCE REMOTE SENSING DATA – A CASE STUDY IN APPIN, NSW, AUSTRALIA.
5.2.1. INTRODUCTION
The first aim of this study is to investigate the capabilities of MCS algorithms for classifying
multisource remote sensing data. Secondly, the combination of FS wrapper approach based
on the GA method and the MCS algorithm was proposed, developed and demonstrated.
Although these techniques have been used in classification of remote sensing data, this study
is one of the first attempts to integrate these techniques for classifying multisource satellite
imagery. It is expected that this new approach can improve the generality and diversity of the
final solution and therefore can increase the classification accuracy.
As mentioned earlier, various FS methods have been used in remote sensing, such as
Exhaustive Search, Forward and Backward Sequential Feature Selection, Simulated
Annealing and GA. Results of numerous studies and the work presented in Chapter 4 have
demonstrated that the GA technique is very efficient for dealing with large datasets and is
better able to avoid a local optimal solution than other methods (Goldberg 1989, Huang and
Wang 2006, Zhou et al. 2010). Another important advantage of the GA techniques is its
capability to search for input features and parameters of classifiers simultaneously. Therefore,
the FS technique based on the GA was employed to address the challenging task of selecting
131
the optimal combined subsets of features. In this study, the synergistic use of the FS wrapper
method with GA and multiple classifiers combination for classifying land cover features in
New South Wales, Australia will be evaluated. This approach is referred to here as the FS-
GA-MCS model.
The non-parametric classification algorithms, such as ANN with a Multilayer Perceptron-
Back Propagation algorithm (MLP-BP) or SVM, have proven to be more effective than the
commonly used ML algorithms for classifying multisource remote sensing data since they are
not constrained by the assumption of normal distribution. In this section, beside the SVM and
MLP-BP network, another non-parametric classifier, namely the Self-Organising Map (SOM)
is introduced and tested.
5.2.2. STUDY AREA AND USED DATA
The study area is located in Appin, in the state of New South Wales, Australia, centred
around coordinate 150o 44’ 30” E; 34o 12’ 30” S. The site is characterised by a diversity of
covered features such as native dense forest, grazing land, urban and rural residential areas,
facilities and water surfaces. Remote sensing data used for this study includes:
SAR: Six ENVISAT/ASAR VV polarisation and six ALOS/PALSAR HH polarisation
images acquired in 2010 (Figure 5.1, Figure 5.2 and Table 5.1).
Optical: Three Landsat 5 TM images acquired on 25/03/2010, 10/9/2010 (Figure 5.3) and
31/12/2010 with seven spectral bands and spatial resolution of 30m. In this study six spectral
bands (except the thermal band) were used.
Table 5.1: ENVISAT/ASAR and ALOS/PALSAR images for the study area.
Satellite/Sensors Date Polarisation Mode 04/01/2010 HH Ascending 22/05/2010 HH Ascending
ALOS/PALSAR 07/07/2010 HH Ascending 22/08/2010 HH Ascending 07/10/2010 HH Ascending 22/11/2010 HH Ascending 03/04/2010 VV Descending 24/06/2010 VV Ascending
ENVISAT/ASAR 25/06/2010 VV Descending 27/06/2010 VV Ascending 28/06/2010 VV Descending 25/09/2010 VV Descending
132
Figure 5.1: ENVISAT/ASAR VV polarised image acquired on 25/09/2010.
Figure 5.2: ALOS/PALSAR HH polarised image acquired on 07/10/2010.
133
Figure 5.3: Landsat 5 TM images (false colour) acquired on 10/09/2010.
5.2.3. METHODOLOGY
5.2.3.1. Data processing
Both SAR (ALOS/PALSAR & ENVISAT/ASAR) and Landsat 5 TM images were registered
to the same map coordinate system (UTM projection, WGS84 datum) and resampled to 15m
pixel size based on the procedure described previously in section 3.1.2, 3.2.2 and 4.3.3. The
DEM was used in the geometric correction process to remove the relief displacements and
adjusted SAR backscatter coefficients. The Enhanced Lee Filter with a 5x5 window size was
applied to filter speckle noise in the SAR images. SAR backscatter values were converted to
decibel (dB) using Equation (3.1). Pixel’s values in the Landsat 5 TM images were converted
to reflectance using Equation (2.23).
In this study, first and second order texture measures (Grey Level Co-occurrence matrix, or
GLCM) were extracted for classification. The First Principal Components (PC1) images,
which were computed from each of multi-date ALOS/PALSAR, ENVISAT/ASAR and
multi-date Landsat 5 TM image datasets, were used to generate multi-scale textural
134
information. Finally, three first order texture measures, including Mean, Variance and Data
range, and four GLCM texture measures, namely Variance, Homogeneity, Entropy and
Correlation with four window sizes 5x5, 9x9, 13x13 and 17x17 were selected.
Normalised Difference Vegetation Indices (NDVI) images were computed from the Red and
Near-Infrared bands of Landsat 5 TM images:
REDNIRREDNIRNDVI
+−
= (5.5)
where NIR and RED are reflectance values of a pixel at the Near-Infrared and Red bands,
respectively.
Four different combined datasets have been generated and applied for the classification
processes (Table 5.2).
Table 5.2: Combined datasets for land cover classification in the study area.
ID Datasets Number of features
1 Six-date PALSAR + six-date ASAR images 12
2 Three-date Landsat 5 TM images 18
3 Three-date Landsat 5 TM + six-date PALSAR + six-date ASAR images 30
4 Three-date Landsat 5 TM + six-date PALSAR + six-date ASAR +
+ Landsat 5 TM & SAR’s textures + three-date NDVI images 173
5.2.3.2. Classification algorithms
Three non-parametric classifiers were employed for classification processes: MLP-BP, SOM
and the SVM.
Artificial Neural Network (ANN)
The MLP-BP model with three layers (input, hidden and output layer) was employed. The
number of neurons in the input layer is equal to the number of input features, the number of
neurons in the output layer is the number of land cover classes to be classified. The optimal
number of input neurons and the number of neurons in the hidden layer was searched by GA
techniques. The sigmoid function was used as the transfer function. The other important
135
parameters were set as follows: maximum number of iteration: 1000; learning rate: 0.01-0.1;
training momentum: 0.9.
Support Vector Machine (SVM)
The SVM classifier with a Gausian Radical Basis Function (RBF) kernel has been used
because of its highly effective and robust capabilities for handling of remote sensing data
(Kavzoglu and Colkesen 2009, Waske and Benediksson 2007). Two parameters need to be
optimally specified in order to ensure the best accuracy: the penalty parameter C and the
width of the kernel function γ. These values will be determined by the GA while searching
for optimal combined datasets. For other cases a grid search algorithm with multi-fold cross-
validation was used.
Self-Organising Map (SOM)
The SOM is another wellknow neural network classifier. The SOM network has the unique
property that it can automatically detect (“self-organising”) the relationships within the set of
input patterns without using any predefined data models (Salah et al. 2009, Tso and Mather
2009). Previous studies suggest that the SOM is an effective method for classifying remotely
sensed data (Ji 2000, Janwen and Bagan 2005, Hugo et al. 2006, Salah et al. 2009, Lee and
Lathrop 2007). As this algorithm has not been introduced in the previous chapter, the basic
concepts and applications of the SOM will be presented in this section.
The SOM network includes input and output layers. The input layer consists of neurons
which represent input feature vectors – the number of input neurons is equal to the number of
input features. The output layer is generally a two-dimensional array of neurons arranged in a
spatial rectangular or hexagonal lattice. A neuron j in the output layer is connected with
neuron i in the input layer by the synaptic weight Wji. These weights are set randomly at the
beginning and then are gradually modified in the training process to characterise input
patterns. In the final result, the neurons which have similar properties (in terms of weight
value) will be organised close together (Tso and Mather 2009). An example of a SOM
network architecture is shown in Figure 5.4.
136
Figure 5.4: Example of the SOM architecture with three input neurons and 25 (5x5) output
neurons.
Classification using the SOM network is carried out in two stages: coarse tuning and fine
tuning. The coarse tuning involves searching for the winning neuron which has synaptic
weights closest to the input data vector, updating weights of the winning neuron and its
neighbouring neurons, and finally assigning labels to neurons.
The Euclidean distance is often used to determine the wining neuron:
⎟⎠
⎞⎜⎝
⎛−= ∑
=
n
i
tji
tic wxM
1
2)(minarg (5.6)
where xit is the input to neuron i at a training step t, and wt
ji is the synaptic weight
connecting the input neuron i with output neuron j
Weights of the winning neuron and its neighbouring neurons are then adjusted so that the
wining neuron will be more similar to the input vector. The most commonly used functions
for determining the neighbourhood of the wining neuron is a decay function, which makes
the neighbourhood radius shrink over time:
⎟⎠⎞
⎜⎝⎛−=
λσσ 1exp)( 0t (5.7)
where σ(t) is the neighbourhood radius of the winning neuron at a training step t, σ0 is the
initial radius, and the λ is a time constant. A decay function reduces the neighbourhood radius
137
from an initial size that can include all of the neurons to a final one which includes only the
winner.
The weights of the winning neuron and its neighbours within a radius σ(t) are then updated
while those outside are left unchanged:
),(1 tji
ti
ttji
tji WxWW −+=+ α t
cAj∈∀ (5.8)
tc
tji
tji AjWW ∉∀=+ ,1 (5.9)
where Act is the set of neighbourhood neurons of the winning neuron j which fall inside the
radius σ(t), and αt is the learning rate at iteration t.
The learning rates also shrink based on a similar time decay function. During the training
process the winning neuron is gradually moved closer to the input training vector in the input
space. At the end of the training stage, a label for each neuron is determined based on a
majority voting rule. A neuron will be labelled by the class which was assigned to it most
frequently.
In the fine tuning stage the SOM network will be further trained and class labels initially
assigned by the coarse tuning phase will be refined according to Learning Vector
Quantisation (LVQ). The known input patterns are again input to the trained SOM and the
winning neuron is determined by selecting the closest Euclidean distance between the input
pattern and weights. Weights of the winning neuron are adjusted according to Equations
(5.10) and (5.11).
If the label of the winning neuron is the same as the input pattern, the corresponding weights
are adjusted so that they move toward the input pattern (Tso and Mather 2009):
)(11 tci
ttc
tc WxWW −+= ++ δ (5.10)
If the label of the winning neuron is not assigned correctly, then the corresponding weights
are adjusted so that they move away from the input pattern (Tso and Mather 2009):
)(11 tci
ttc
tc WxWW −−= ++ δ (5.11)
where Wc is the weight vector of the winning neuron, and δt is a gain term that will decrease
with time. The gain factor has to be set in the range of (0-1).
138
Ji (2000) evaluated the performance of the Kohonen SOM network for classifying seven land
use/cover classes using a Landsat 5 TM image over the northern sub-urban area of Beijing,
China. The classification by ML and Back Propagation network were also implemented for
comparison. It is revealed that the SOM network outperformed the ML method when four
spectral bands were used. The SOM achieved similar accuracy as the BP network. There are
a number of factors affecting the learning capability of SOM network such as size of
network, training samples and learning rate. The author went on to conclude that the SOM
network is a viable alternative for land use classification with remote sensing data. Jianwen
and Bagan (2005) used the SOM network to classify ASTER images over the Beijing area,
China. The SOM provided an increase in overall accuracy of 7% compared to the traditional
ML algorithm, and in particular a 50% increase for the features ‘residential area’ and ‘road’.
Hugo et al. (2006) carried out classification using the SOM classifier on MERIS full
resolution data acquired in 2004 for Portugual. Results of the classification using the SOM
for 19 land use/cover classes were then compared with the result of k-Nearest Neighbour
(kNN) classification. They found that the SOM was more efficient than the kNN for
classifying land cover features with MERIS data. Although the increase in classification
accuracy was only 3% the authors believe that there are some ways to further improve the
performance of the SOM. The authors suggested that the SOM algorithm is powerful, but that
it should be applied by experienced users since there are several parameters that need to be
tuned. The authors also recommended the use of multiple SOMs as an aspect for future
research. In the study of Salah et al. (2009) the SOM classifier has been applied successfully
for the detection of buildings from Lidar data and multi-spectral aerial images.
In this chapter, the input layer is dependent on different input datasets. The output layer of the
SOM was a two dimensional array of 15x15 of neurons (total 225 neurons). The neurons in
the input layer and output layer are connected by synaptic weights which are randomly
assigned within a range of 0 to 1.
GA techniques
FS based on the GA and wrapper approach was implemented. The application of the GA
requires the design of the chromosomes, the fitness function and the architecture of the
system. The fitness function is designed to take into account the classification accuracy, the
139
number of selected features and average correlation within selected features, and has been
described in Chapter 4 (Equation 4.5).
The weight for the classification accuracy (WOA) and the weight for the number of selected
features (Wfs) were set for the range 0.65-0.8 and 0.2-0.35, respectively. The other parameters
for the GA were:
Population size = 20-40; Number of generations = 200; Crossover rate: 0.8; Elite count:
3-6; Mutation rate: 0.05.
The experiments were carried out as follows. Firstly, the SVM, ANN and SOM classifiers
were used to classify the original combined datasets without searching for relevant features.
Then the FS with GA (FS-GA) was implemented for each combined datasets using the SVM,
ANN and SOM classifiers. These processes will give the classification results of each
classifier with corresponding optimal datasets and parameters. Results are compared with the
conventional non-feature selection approach. Secondly, the MCS techniques were applied to
the above results of the non-feature selection approach. Two MCS approaches were
investigated: including training samples manipulation with bagging and boosting
(Adaboost.M1) and a combination of different type of classifiers. Finally, the proposed model
which integrates the FS-GA and the MCS was implemented, in which results of classification
based on the FS-GA model with SVM, ANN and SOM classifiers were combined using MCS
techniques. For comparison purposes, three combination rules were employed to integrated
the classification results, namely the Majority Voting (MV), the Bayesian Sum (Sum) and the
evidence reasoning based on Dempster – Shafer theory (DS). This integration approach is
illustrated in Figure 5.5. In this study, all experiments were run on the Matlab 7.11.1(2010b)
platform. The ANN classification and the GA were implemented using the ANN toolbox and
the Global Optimization toolbox, respectively. The SVM classification was carried out using
the LIBSVM 3.1 toolkit. The SOM and MCS were also implemented in Matlab 7.11.1.
140
Figure 5.5: The integration of feature selection based on GA and MCS techniques for
classifying multisource data.
Six land cover classes, namely Native Forest (NF), Natural Pastures (NP), Sown Pastures
(SP), Urban Areas (UB), Rural Residential (RU) and Water Surfaces (WS) were identified
for classification. The UB class includes residential, commercial, industrial, educational or
research facilities, tourist developments, gaols, cemeteries or abandoned urban area.
The data used for training and validation were derived from visual interpretation of an old
land use map with the help of Google Earth images. The training and test data were selected
randomly and independently. The samples selection has been carried out based on the
procedure described earlier in the section 3.1.2. The sizes of training and testing datasets are
listed in Table 5.3.
Table 5.3: Contents of training and testing datasets.
Classes Training data
(number of pixels)
Testing data
(number of pixels)
Native Forest (NF) 719 718
Natural Pastures (NP) 616 682
Sow Pastures (SP) 458 466
Urban Areas (UB) 422 425
Rural Residential (RU) 419 458
Water Surfaces (WS) 343 376
141
5.2.4. RESULTS AND DISCUSSION
5.2.4.1. Classification using Non-FS and FS-GA methods
The overall classification accuracy and Kappa coefficients for the SVM, ANN and SOM
classifiers over different datasets without feature selection is summarised in Table 5.4.
Table 5.4: The classification performance of SVM, ANN and SOM algorithms
on different combined multisource datasets.
Datasets
Overall classification accuracy (%) and Kappa coefficients
SVM ANN SOM
Accuracy Kappa Accuracy Kappa Accuracy Kappa
1 59.26 0.50 59.39 0.50 56.06 0.46
2 79.01 0.74 75.49 0.70 79.97 0.75
3 81.47 0.77 80.99 0.76 80.03 0.75
4 82.78 0.79 81.70 0.77 78.84 0.74
The classification results illustrate the efficiencies of the synergistic use of multi-date optical
and SAR images. All of the classifiers offered significant increases in classification accuracy
when the combined datasets (3rd and 4th) were used. The combined multi-date Landsat 5 TM
and SAR data increased overall accuracy by 2.46% and 22.1% for SVM, 5.5% and 21.6% for
ANN and 0.06% and 23.97% for SOM as compared to the cases that used only the multi-date
Landsat 5 TM or multi-date SAR images. The integration of textural information and NDVI
images slightly enhanced the classification results of the SVM and ANN classifiers by 1.31%
and 0.71%, respectively. However this integration reduced the accuracy by 1.19% in the case
of the SOM algorithm. The Hughe’s phenomenon could be the reason for this problem as the
number of input features is 173, which is very large.
A comparison of the performance of the SVM, ANN and SOM classifiers using feature
selection based on the GA (FS-GA) model and the non-feature selection approach is given in
Table 5.5.
142
Table 5.5: Comparison of classification performance between FS-GA approach and
the Non-FS approach.
Datasets
Overall classification accuracy (%)
Non-FS FS-GA
SVM ANN SOM SVM ANN SOM
1 59.26 59.39 56.06 59.94 61.15 57.41
2 79.01 75.49 79.97 81.06 80.19 80.37
3 81.47 80.99 80.03 82.37 81.28 80.74
4 82.78 81.70 78.84 85.22 82.77 81.54
The improvements obtained by using multisource data were more significant when the FS
approach was applied. For instance, with the FS-GA method, the classification of a
combination of original optical and SAR images with their textural and NDVI data (4th
dataset) gave increases of overall accuracy of 2.85%, 1.49% and 0.80% for the SVM, ANN
and SOM classifiers, respectively.
It is clear that the FS-GA approach performed better than the traditional Non-FS approach,
particularly with high dimensional datasets. For all of the datasets and classifiers that have
been evaluated, the FS-GA approach provided significant improvements in classification
accuracy. The increases of overall classification accuracy ranged from 0.29% (ANN classifier
with the 3rd dataset) to 2.70% (SOM classifier with the 4th dataset). The highest accuracy of
85.22% was achieved by the integration of FS approach with the SVM classifier for the 4th
dataset. The FS-GA approach used many less input features than the traditional approach. For
instance, in the case of the SVM classifier and the 4th dataset, only 68 out of 173 features
were selected. Moreover, the problem of Hughe’s phenomenon does not exist when the FS-
GA technique was applied. Unlike the conventional Non-FS approach, in the FS-GA
approach the classification using the SOM classifier on the 4th dataset led to a slight increase
of classification accuracy of 0.80%. Figure 5.6 shows the result of classification using the FS-
GA approach with the SVM classifier which gave the best accuracy for the 4th dataset.
143
Figure 5.6: Results of classification of the 4th dataset using FS-GA with SVM classifier.
The improvements of classification accuracy by using the FS-GA technique for different
classifiers and combined datasets was shown in Figure 5.7.
Figure 5.7: Improvements of accuracy by applying the FA-GA approach for the SVM, ANN
and SOM classifiers.
144
5.2.4.2. Classification using Multiple Classifier Systems (MCS)
Bagging and boosting
Results of the bagging and Adaboost.M1 methods with the SVM, ANN and SOM base
classifiers are presented in Tables 5.6 - 5.8 and Figures 5.8 - 5.10. It can be seen that the
bagging and boosting (Adaboost.M1) algorithm using the SVM base classifier did not
provide significant improvement in the classification of multisource data over the study area
as compared to the original single SVM classifier. In particular, in the case of bagging, the
overall classification accuracies decreased for all four datasets. The SVM-Adaboost.M1
algorithm gave some improvements in the 2nd and 3rd datasets, but had reduced classification
accuracy for the 1st and 4th datasets. This is not surprising since the SVM classifier is
considered to be a stable and accurate algorithm, and consequently the ensemble technique
does not help much to enhance the classification performance.
Unlike the case of SVM-based classifiers, the bagging and boosting algorithms using ANN
and SOM classifiers, in general, gave considerable improvements compared to the
performance of the original classifiers. The ANN-Bagging algorithm provided significant
increases in overall accuracy (up to 4.8%) for all four datasets. The ANN-Adaboost.M1 gave
noticeable improvement in classification accuracy for the 2nd and 3rd datasets, while
marginally reduced accuracy for the 1st and 4th datasets. Similarly, the SOM-Bagging
algorithm also outperformed the original SOM classifier in all of the datasets. The success of
these MCS techniques can be explained by the unstable nature of ANN and SOM algorithms.
However, the SOM-Adaboost.M1 produced noticeable decreases in classification
performance.
The highest classification accuracy obtained by the bagging and boosting techniques was
84.06% for the ANN-Bagging while classifying the 4th dataset. This result is slightly better
than the best result of an individual classifier, with 82.78% obtained by the SVM method
using the same dataset.
145
Table 5.6: Results of classification using single SVM classifier, bagging and Adaboost.M1
techniques based on SVM classifier for different combined datasets.
Datasets
Overall classification accuracy (%) and Kappa coefficients
SVM SVM-Bagging SVM-Adaboost.M1
Accuracy Kappa Accuracy Kappa Accuracy Kappa
1 59.26 0.50 59.20 0.50 58.98 0.50
2 79.01 0.74 74.27 0.68 80.93 0.77
3 81.47 0.77 74.72 0.69 82.37 0.78
4 82.78 0.79 80.38 0.76 80.42 0.76
Table 5.7: Results of classification using single ANN classifier, bagging and
Adaboost.M1 techniques based on ANN classifier for different combined datasets.
Datasets
Overall classification accuracy (%) and Kappa coefficients
ANN ANN-Bagging ANN-Adaboost.M1
Accuracy Kappa Accuracy Kappa Accuracy Kappa
1 59.39 0.50 64.19 0.55 59.04 0.49
2 75.49 0.70 77.09 0.7168 78.56 0.74
3 80.99 0.76 81.02 0.7662 80.35 0.76
4 81.70 0.77 84.06 0.80 82.91 0.79
Table 5.8: Results of classification using single SOM classifier, bagging and
Adaboost.M1 techniques based on SOM classifier for different combined datasets.
Datasets
Overall classification accuracy (%) and Kappa coefficients
SOM SOM-Bagging SOM-Adaboost.M1
Accuracy Kappa Accuracy Kappa Accuracy Kappa
1 56.06 0.46 58.46 0.49 52.90 0.419
2 79.97 0.75 80.86 0.76 78.02 0.7301
3 80.03 0.75 82.34 0.78 79.52 0.7485
4 78.84 0.74 79.30 0.75 77.76 0.7267
146
Figure 5.8: Comparison of SVM classification with SVM-Bagging and SVM-Adaboost.M1
methods.
Figure 5.9: Comparison of ANN classification with ANN-Bagging and ANN-Adaboost.M1
methods.
Figure 5.10: Comparison of SOM classification with SOM-Bagging and SOM-Adaboost.M1
methods.
147
MCS using Majority Voting, Sum and Dempster-Shafer theory combination rules
Results of combination of three classifiers for different datasets are presented in Table 5.9
and Figure 5.11 along with results of best individual classifier (BIC) for each dataset. It is
clear that the MCS approach outperformed the BICs in most cases. The MCS using Majority
Voting and Sum rules produced significantly higher classification accuracy then the best
individual classifiers in all cases, while the MCS based on the Dempster-Shafer rule gave
better performance in three out of four datasets. The highest accuracy of 84.77% (Kappa=
0.81) was achieved by using the MV rule classifying the largest datasets (which includes 173
features). The performance of the three decision rules was rather comparable, however the
MV and SM rules gave marginally better accuracy than the DS method. It is also worth
mentioning that the results of the MCS using a combination of three different classifiers
provided consistently higher classification accuracy than the bagging and boosting methods.
This superior classification performance can be explained by the higher level of diversity of
the different classifiers. The only exception is for the 1st dataset where the ANN-Bagging
approach gave the highest classification accuracy of 64.19%, while the MCS using the
Dempster-Shafer combination rule gave rather low accuracy of 57.76%.
Table 5.9: Comparison between the best classification results obtained by individual
classifiers and MCS based on the decision rules of Majority Voting, Sum and Dempster-
Shafer theory.
Datasets
Overall classification accuracy (%) and Kappa coefficients
BIC MV Sum DS
Accuracy Kappa Accuracy Kappa Accuracy Kappa Accuracy Kappa
1 59.39 0.50 60.32 0.51 59.49 0.50 57.76 0.48
2 79.97 0.75 81.41 0.77 81.06 0.77 80.51 0.76
3 81.47 0.77 83.17 0.79 83.52 0.80 82.98 0.79
4 82.78 0.79 84.77 0.81 84.48 0.81 84.54 0.81
148
Figure 5.11: Comparison between the best classification results obtained by original
classifiers, MCS based and FS-GA approaches.
5.2.4.3. Integration of feature selection techniques and multiple classifier system (FS-GA-
MCS)
The investigation on FS-GA and MCS techniques showed that these methods are very robust
and effective in classifying complex datasets. Thus, it could be an ideal approach to combine
these two methods in order to maximise the benefits of using multisource remote sensing
data.
A comparison of classification results between the BIC, FS-GA, best results of the MCS
approach and best results of integration approach (FS-GA-MCS) of FS-GA with the MCS
technique using MV, Sum and DS rules are given in Table 5.10 and Figure 5.12.
Table 5.10: Comparison of best classification results using single classifier (Non-FS), FS-
GA and FS-GA-MCS classifier combination approaches.
Datasets
Overall classification accuracy (%) and Kappa coefficients
BIC FS-GA Best MCS Best FS-GA-MCS
Accuracy Kappa Accuracy Kappa Accuracy Kappa Accuracy Kappa
1 59.39 0.50 61.15 0.51 60.32 0.51 62.56 0.53
2 79.97 0.75 81.06 0.77 81.41 0.77 82.59 0.78
3 81.47 0.77 82.37 0.78 83.52 0.80 84.35 0.80
4 82.78 0.79 85.22 0.82 84.77 0.81 88.58 0.86
149
Figure 5.12: Comparison between the BIC, FS-GA, the best results MCS, the best (FS-GA-
MCS1) and the poorest (FS-GA-MCS2) results of FS-GA-MCS methods for different
datasets.
Although the FS-GA and MCS approaches have already produced significant increase in
classification accuracy for the evaluated multisource remote sensing datasets, the integration
of the MCS with FS-GA method has further improved the classification performance. The
FS-GA-MCS approach always gave better accuracy than any BIC in all cases. As compared
to the FS-GA approach the FS-GA-MCS approach resulted in increases in classification
accuracy for all four cases, with a range from 1.41% for the 1st dataset to 3.36% for the 4th
dataset. The FS-GA-MCS approach also gave noticeable improvements in classification
accuracy in comparison with the case of the MCS, with a range from 0.83% for the 3rd dataset
to 3.81% for the 4th dataset. Increases in classification are even more significant compared to
the traditional Non-FS approach. As can be seen in Figure 5.12, the FS-GA-MCS approach is
very reliable, and even the poorest results of this method were quite comparable with the FS-
GA and the best results of the MCS method. The highest classification accuracy obtained by
the FS-GA-MCS approach was 88.58% with the largest combined datasets. A comparison of
improvements in classification performance between FS-GA and Non-FS; FS-GA-MCS and
FS-GA and FS-GA-MCS and MCS models is given in Figure 5.13.
One of the possible reasons for the success of the FS-GA-MCS approach is its capability to
integrate various optimal (or nearly optimal) solutions given by the GA for specific classifiers
such as SVM, ANN or SOM, in order to enhance the generality and diversity of the final
solution.
150
Figure 5.13: Improvements of overall classification accuracy achieved by using FS-GA, MCS
and FS-GA-MCS approaches.
All three combination rules, including MV, Sum and DS algorithm, were effective for
combining classification results. In the case of integration of FS-GA and MCS, the
performance of the three rules were similar. This result agrees with that of Yan and Shaker
(2011). The performance of the MV, Sum and DS algorithms are shown in Table 5.11.
Table 5.11: Classification accuracies by applying MV, Sum and DS algorithm for
combination of FS-GA and MCS approaches.
Algorithm
Datasets
1 2 3 4
Accuracy Kappa Accuracy Kappa Accuracy Kappa Accuracy Kappa
DS 62.56 0.53 82.30 0.78 83.81 0.80 88.29 0.86
MV 61.89 0.53 82.30 0.78 83.52 0.80 88.22 0.86
Sum 61.18 0.52 82.59 0.79 84.35 0.81 88.58 0.86
151
5.2.5. CONCLUSIONS
Numerous MCS (classifier ensemble) techniques have been proposed and used to classify
different combined datasets of multi-date ENVISAT/ASAR, ALOS/PALSAR and Landsat 5
TM images. While the improvements in classification performance by employing the bagging
and boosting algorithms, except for the case of ANN-Bagging, are not very clear, the use of
MCS with SVM, ANN and SOM classifiers gave significant increases in classification
accuracy no matter which MV, Sum or DS combination rules were applied. The FS-GA
approach is very effective for handling multisource data, and its combination with any
classifier (ANN-MLP, SVM or SOM) gave considerable improvement in classification
accuracy.
The integration of FS-GA and the MCS techniques, which aims to further enhance the
performance of multisource data classification, has been implemented and evaluated. Results
of the investigation reveal that the proposal method is viable and useful. The FS-GA-MCS
approach outperformed the classical Non-FS, and both the FS-GA and MCS approaches in all
studied cases. The FS-GA-MCS approach always gave significantly higher accuracy than any
single best classifier. The FS-GA-MCS approach produced the highest overall classification
accuracy of 88.58% (Kappa=0.86) for the largest combined dataset consisted of original SAR
and optical images, their textural information and NDVI. The FS-GA and MCS based on
MV, Sum or DS combination rules are also very effective for classification of multisource
remote sensing data, although they are less accurate than the FS-GA-MCS approach. Finally,
the classification results confirm the advantages of multisource remote sensing data,
particularly the integration of SAR and optical data, for land cover mapping.
152
CHAPTER 6
CONCLUSIONS 6.1. SUMMARY AND CONCLUSIONS
The objectives of this thesis were to develop an appropriate methodology for improving land
cover classification using multisource remote sensing data.
To achieve this goal, firstly, an extensive literature review was conducted into the major
image classification techniques, the reported benefits of using multisource remote sensing
data and related work in this area. The review was conducted in chapter 2. It was revealed
that unlike the traditional parametric classifiers, the non-parametric classifiers are robust for
handling complex datasets because their principles are not constrained by the normal
distribution assumption. Hence, these algorithms should be, in principle, better suited for
classifying multisource remote sensing data. However, it was a surprise that in many studies,
which used combinations of different kinds of imagery and spatial data as an input for land
cover classification, the traditional Maximum Likelihood (ML) classifier was still very
popular. The application of non-parametric classifiers was reported in a number of studies but
they were not adequately investigated.
Chapter 3 was dedicated to evaluating the performance of multisource data and the non-
parametric algorithms for land cover classification in three different scenarios. A large
number of combined datasets generated from various types of data, including multi-temporal,
multi-polarisation and dual-frequency SAR images (ALOS/PALSAR and ENVISAT/ASAR),
multi-spectral images (SPOT 2 XS, Landsat TM) as well as derivative products such as
interferometric coherences, indices and textural information, were investigated. The Artificial
Neural Network (ANN) classifier based on the Multilayer Perceptron-Back Propagation
(MLP-BP) algorithm and the Support Vector Machine (SVM) non-parametric classifier were
applied, while the commonly used ML classifier was also implemented for comparison
purposes. The results obtained did confirm that the use of multisource data for land cover
153
classification is an appropriate approach. The classification based on combined datasets
always produced higher classification accuracy than that of single-type images/datasets. The
complimentary characteristics, and the impacts and contributions of each constituent data
type were taken into account and clearly emphasised. The use of multisource remote sensing
data not only provided an increase in the classification accuracy but also improved the
detectability/discernability/separability of land cover features. For instance, the combination
of ALOS/PALSAR backscatter and interferometric coherence data allowed visual detection
of burnt areas in the Victorian Bushfires, which was not possible using any single dataset.
Although these improvements in classification performance were obtained for all the
classifiers applied to multisource data, the MLP-BP and SVM algorithms did, in general,
outperform the ML algorithm, particularly when dealing with high dimensional, complex
datasets. Moreover, the SVM algorithm often provided the highest classification accuracy,
and marginally outperformed the MLP-BP classifier. This confirmed the advantages of non-
parametric classifiers over traditional parametric classifier for multisource remote sensing
data classification.
The integration of a feature selection technique based on the Genetic Algorithm (GA) with
the non-parametric classifiers was proposed as a means of resolving the challenging task of
selecting optimal combined datasets and defining parameters when classifying multisource
data. The GA algorithm has been used in remote sensing when dealing with high
dimensionality datasets such as hyperspectral data. However, the evaluation of this technique
for classification of multisource data has not been well addressed by previous investigators.
The basic concepts of feature selection, the GA algorithm and their application in an
experiment using a combination of SVM and GA algorithms (SVM-GA) for classifying
multisource remote sensing in Western Australia were discussed in chapter 4. In this
experiment the SVM-GA method has been used to classify 20 different combined datasets
generated from Landsat 5 TM multispectral, multi-temporal ALOS/PALSAR dual-
polarisation images and multi-level textural data. An approach based on “non-feature
selection” and the combination of SVM and Sequential Forward Floating Search (SVM-
SFFS) algorithms were also implemented for comparison purposes. Results of the
experiments illustrated the benefit of the proposed feature selection/SVM-GA approach. The
SVM-GA approach produced higher classification accuracy than the non-feature selection
/SVM-SFFS approach (in 19 out of 20 cases). The feature selection/SVM-GA approach also
achieved the highest overall classification accuracy of 96.47% for the largest combination
154
datasets, in which 81 out of 189 features were selected. It is interesting that in all cases that
the SVM-GA method was applied the Hughes phenomenon did not appear. Finally, the
modified objective function led to more reliable classification results than other functions
reported in the literature.
Other advanced techniques which have potential to increase classification accuracy of remote
sensing data, were also applied to multisource data classification. One is the Multiple
Classifier Systems (MCS) or classifier ensemble. The principles of MCS and its application
in remote sensing were reviewed in chapter 5. A new method was proposed and implemented
based on the integration of the GA with the MCS algorithm (FS-GA-MCS). Investigations
have been carried out using different MCS techniques with three non-parametric classifiers,
namely SVM, ANN-MLP and Self-Organising Map (SOM), feature selection based on the
GA (FS-GA) and the proposed approach over a study area in Appin, NSW, Australia. It was
revealed that both the FS-GA and MCS approaches (using classifier combination) produced
noticeable improvements in classification accuracy compared to the results of any individual
classifier. However, the FS-GA-MCS approach clearly outperformed other approaches in all
cases and gave the highest overall classification accuracy of 88.58% for the largest datasets.
This confirmed the value and effectiveness of the proposed approach for classifying
multisource data.
It must be emphasised that experiments in this thesis have been carried out successfully in
various study areas with highly diverse landscape and land cover patterns. These include two
test sites in Vietnam which represent an urban-suburban area in Ho Chi Minh City and the
typical rice production area in Mekong Delta; three test sites in Australia which represent
forest area prone to bushfire in Victoria, grassland (pasture) area in Western Australia and an
area with mixed use forest, pasture and residential in Appin, New South Wales. However,
because of time constraints the combination of the feature selection technique and the MCS
were only investigated for the Australian study areas. Nevertheless, the results of these
studies have illustrated the efficiency and usefulness of the proposed methods for different
types of land surfaces.
155
6.2. MAIN FINDINGS
The work carried out in the thesis has led to the following major findings:
- Multisource data can significantly improve the land cover classification performance,
because the complementary properties of different kinds of data are taken into
account.
- Non-parametric classifiers such as the ANN and SVM are appropriate for exploiting
the benefits of multisource remote sensing data. These algorithms generally produce
higher classification accuracy than a traditional parametric classifier such as the ML
algorithm.
- The integration of the GA with a non-parametric classifier is an efficient method for
handling multisource data. Such a hybrid approach ensures selection of the best
subsets and classifier parameters when applying them to high dimensionality datasets.
This method outperformed both the non-feature selection approach and the integration
of the conventional SFFS algorithm with the non-parametric classifier. In addition,
this method shows great capability to avoid the Hughes phenomenon, which limits the
classification accuracy while dealing with high dimensionality datasets.
- The MCS algorithm based on non-parametric classifiers is also capable of improving
the classification of multisource data, and this approach often resulted in higher
classification accuracy than any best individual classifier.
- The newly developed method of integration of the FS wrapper based on GA and the
MCS techniques is very robust and effective for classifying highly dimensional and
complex datasets. This method can exploit advantages of both FS-GA and MCS
techniques to improve the classification performance. It was found that this integrated
approach outperforms other approaches, including both FS-GA and MCS techniques.
6.3. FUTURE WORK
The proposed methodologies are not confined to the data that had been used for experiments
described in this thesis. To date, considering the large diversity of available satellite imagery
and other digital data sources, the use of these techniques with other kinds of imaging data,
such as high resolution optical and SAR images (Quick Bird, Worldview, TerraSAR-X or
Cosmo-SkyMed), hyperspectral (Chris/Proba, Hyperion) and Lidar data should be
investigated.
156
The textural information, particularly multi-level texture information, is a valuable data
source and has been incorporated into the classification process. However, the relationship
between textural data and particular feature classes has not been evaluated in detail.
Furthermore, some other kinds of texture measures, such as geostatistical and temporal
textures, have not been included in these studies. These issues could be investigated in future
work.
The feature selection technique based on the GA has been proposed and successfully applied
for handling highly complex datasets. However, the wrapper approach is very time
consuming, hence it is necessary to improve the efficiency of this approach. One possible
solution is to combine the filter and wrapper approaches into a single process.
Several MCS techniques have been applied for combining classification outputs and have
demonstrated positive results. In future, the use of other MCS techniques, such as multi-level
classification and fuzzy integral methods, should be studied.
The integration of several MCS and FS techniques based on GA algorithms studied in the
thesis has shown a great capability to improve the classification performance of multisource
remote sensing data. The further research should investigate on different combination of
these techniques to seek for the most efficient solution.
157
REFERENCES
Abe, B., Gidudu, A., and Marwal, T., 2010. Investigating the effects of ensemble
classification on remotely sensed data for land cover mapping, Proceeding of the
International Geoscience and Remote Sensing Symposium (IGARSS), Honolulu HI,
25-30 July, pp. 2382-2835.
Agelis, C.F., Freitas, C.C., Valeriano, D.M. and Dutra, L.V., 2002, Multitemporal analysis of
land use/land cover JERS-1 backscattering in the Brazilian tropical rainforest.
International Journal of Remote Sensing, 23 (7), pp. 1231–1240.
Aitkenhead, M.J., and Aalders, I.H., 2008. Classification of Landsat Thematic Mapper
imagery for land cover using neural networks, International Journal of Remote
Sensing 29 (7), pp. 2075–2084.
Almeida-Filho, R., Rosenqvist, A., Shimabukuro, Y.E., and Silva-Gomez, R., 2007.
Detecting deforestation with multitemporal L-band SAR imagery: a case study in
western Brazilian Amazonia. International Journal of Remote Sensing, 28 (6),
pp.1383-1390.
Almeida-Filho, R., Rosenqvist, A., Shimabukuro, Y.E., and Sanchez, G. A., 2009. Using
dual-polarized ALOS PALSAR data for detecting new fronts of deforestation in the
Brazilian Amazonia. International Journal of Remote Sensing 30 (14), pp. 3735-3743.
Amarsaikhan, D., 2006. Application of multisource data for urban land cover mapping.
Proceeding of ISPRS Commission VII Mid-term Symposium "Remote Sensing: From
Pixels to Processes", Enschede, the Netherlands, 8-11 May.
Anthony, G., and Ruther, H., 2007. Comparison of Feature Selection Techniques for SVM
classification. Proceeding of the ISPRS Working Group VII/1 Workshop
ISPMSRS’07: “ Physical Measurements and Signatures in Remote Sensing”, 12-14
March, Davos, Switzeland.
Antikidis, E., Arino, O., Laur, H., and Arnaud, A., 1998. ERS SAR coherence and ATSR hot
spots: a Synergy for Mapping Deforested Areas. The Special Case of the 1997 Fire
Event in Indonesia. Proceedings of the Retrieval of Bio and Geo-Physical Parameters
from SAR Data for Land Applications Workshop ESTEC, 21–23 October,
Netherlands, ESA SP 441.
Argany, M., Amini, J., and Saradjian, M.R., 2006. Artificial neural networks for
improvement of classification accuracy in Landsat +ETM images, Proceeding of Map
158
Middle East Conference, Dubai, 26-29 March. Available at:
http://www.gisdevelopment.net/proceeding/mapmiddleeast/2006/mm06pos_98.htm.
Avery, T. E., and Berlin, G. E., 1992. Fundamentals of Remote Sensing and Airphoto
Interpretation. Fifth Edition. Prentice Hall.
Ban, Y., and Wu, Q., 2005. RADARSAT SAR data for landuse/land-cover classification in
the rural-urban fringe of the greater Toronto area. Proceedings of 8th AGILE
Conference on Geographic Information Science, 26–28 May 2005, Estoril, Portugal.
Bargiel, D., and Herrmann, S., 2011. Multi-Temporal Land-Cover Classification of
Agricultural Areas in Two European Regions with High Resolution Spotlight
TerraSAR-X Data. Journal of Remote Sensing 3, pp. 859-877.
Baronti, S., Del Frate, F., Ferrazzoli, P., Paloscia, S., Pampaloni, P., and Schiavon, G. 1995.
SAR polarimetric features of agricultural areas. International Journal of Remote
Sensing 16 (14), pp. 2639–2356.
Basili, R., Del Frate, F., Luciani, M., Mesiano, F., Pacifici, F., and Rossi, R., 2008. A
comparative analysis of kernel-based methods for the classification of land cover
maps in satellite imagery. International Geoscience and Remote Sensing Symposium
(IGARSS), 4, 7-11 July, pp. IV - 735 - IV – 738.
Benediktsson, J.A., Chanussot, J., and Fauvel, M., 2007. Multiple classifier systems in
remote sensing: from basics to recent developments. In: M. Haindl, J. Kittler and F.
Roli, eds.Multiple Classifier Systems. Heidelberg, Germany: Springer, pp. 501–512.
Berberoglu, S., Lloyd, C. D., and Atkinson, P. M., Curran, P. J., 2000. The integration of
spectral and textural information using neural networks for land cover mapping in the
Mediterranean. Computer & Geosciences 26, pp.385-396.
Berberoglu, S., Curran, P. J., Lloyd, C. D., and Atkinson, P. M., 2007. Texture classification
of Mediterranean land cover. International Journal of Applied Earth Observation and
Geoinformation 9, pp. 322-334.
Blaes, X., Vanhalle, L., Defourny, P., 2005. Efficientcy of crop identification based on
optical and SAR image time series. Remote Sensing of Environment 96 (3-4), pp.
352-365.
Briem, G. J., Benediktsson, J. A., and Sveinsson, J. R., 2002. Multiple classifier applied to
multisource remote sensing data. IEEE Transactions on Geosciences & Remote
Sensing, 40 (10), pp. 2291-2299.
159
Briem, G. J., Benediktsson, J. A., and Sveinsson, J. R., 2001. Boosting, Bagging, and
Consensus Based Classification of Multisource Remote Sensing Data. Multiple
classifier systems – Lecture notes in Computer Science, 2096/2001, pp. 279-288,
DOI: 10.1007/3-540-48219-9_28.
Brisco, B. and Brown, R.J., 1995, Multidate SAR/TM synergism for crop classification in
western Canada. Photogrammetric Engineering and Remote Sensing 61, pp.1009-
1014.
Bruzzone, L., Marconcini, M., Wegmuller, U., Wiesmann, A., 2004. An advanced system for
the automatic classification of multitemporal SAR images. IEEE Transaction on
Geoscience & Remote Sensing 42 (6), pp. 1321-1334.
Bruzzone, L., and Persello, C., 2009. A Novel Approach to the Selection of Spatially
Invariant Features for the classification of Hyperspectral Images with Improved
Generalization Capability. IEEE Transactions on Geosciences & Remote Sensing 47
(9), pp. 3180-3191.
Charlton, M. B., and White, K., 2006. Sensitivity of radar backscatter to desert surface
roughness. International Journal of Remote Sensing 27 (8), pp. 1641-1659.
Chang, C. H., and Lin, C. J., 2011, LIBSVM : a library for support vector machines”, ACM
Transactions on Intelligent Systems and Technology, 2:27:1--27:27.
Chen, D. M., and Stow, D., 2003. Strategies for Integrating Information from Multiple
Spatial Resolutions into Land-Use/Land-Cover Classification Routines.
Photogrammetric Engineering & Remote Sensing Vol. 69 (11), pp. 1279 - 1287.
Chen, D., Stow, D.A. and Gong, P., 2004. Examining the effect of spatial resolution and
texture window size on classification accuracy: an urban environment case.
International Journal of Remote Sensing 25 (11), pp. 2177–2192.
Chova, L. G., Prieto, D. F., Calpe, J., Soria, E., Vila, J., Valls, G. C., 2006. Urban monitoring
using multi-temporal SAR and multi-spectral data. Journal of Pattern Recognition
Letter 27, pp. 234- 243.
Chu, H.T., and Ge, L., 2010a. Synergistic use of multi-temporal ALOS/PALSAR with SPOT
multispectral satellite imagery for land cover mapping in the Ho Chi Minh city area,
Vietnam. Proceeding of the International Geoscience and Remote Sensing
Symposium (IGARSS), Honolulu HI, 25-30 July, pp. 1465-1468.
Chu, H.T., and Ge, L., 2010b. Land cover classification using combination of L- and C-band
SAR and optical satellite images. Proceeding of 31st Asian conference of Remote
sensing (ACRS), Hanoi, 1-5 November, paper TS39-2, (on CD-ROM).
160
Chust, G., Ducrot, D., and Pretus, J. LL., 2004. Land cover discrimination potential of radar
multitemporal series and optical multispectral images in a Mediterranean cultural
landscape. International Journal of Remote Sensing 25 (17), pp. 3513–3528.
Coburn, C. A., and Roberts, A. C., 2004. A multiscale texture analysis procedure for
improved forest stand classification, International Journal of Remote Sensing 25 (20),
pp. 4287–4308.
Dixon, B., and Candade, N., 2008. Multispectral landuse classification using neural networks
and support vector machines: one or the other, or both. International Journal of
Remote Sensing 29 (4), pp. 1185–1206.
Du, P., Sun, H., and Zhang, W., 2009a. Multiple classifier combination for target
identification from high resolution remote sensing images. ISPRS Archive Volume
XXXVIII-1-4-7/W5, ISPRS Hannover Workshop 2009, High-Resolution Earth
Imaging for Geospatial Information, June 2-5, Hannover, Germany.
Du, P., Zhang, W., and Sun, H., 2009b. Multiple Classifier Combination for Hyperspetral
Remote Sensing Image Classification. MCS 2009, LNCS 5519, Springer-Verlag
Berlin Heidelberg. pp.52-61.
Dutra, L. V., Scofield, G. B., Neta, S. R. A., Negri, R. G., Freitas, C. C., Andrade, D., 2009.
Land Cover Classification in Amazon using Alos Palsar Full Polarimetric Data. Anais
XIV Simpósio Brasileiro de Sensoriamento Remoto, Natal, Brasil, 25-30 April 2009,
INPE, pp. 7259-7264
Engdahl, M. E., Hyyppa, J. M., 2003. Land cover classification Using Multi-temporal ERS -
1/2 InSAR Data. IEEE Transactions on Geosciences & Remote Sensing 41 (7), pp.
1620-1628.
Erasmi, S., and Twele, A., 2009. Regional land over mapping in the humid tropics using
combined optical and SAR satellite data – a case study from Central Sulawesi,
Indonesia. International Journal of Remote Sensing 30 (10), pp. 2465-2478.
Foody, G.M., 1999, Image classification with a neural network: from completely crisp to
fully-fuzzy situations, In P.M. Atkinson and N.J. Tate, Advances in Remote Sensing
and GIS analysis, Chichester: Wiley & Son.
Foody, G.M., 2002, Status of land cover classification accuracy assessment. Remote Sensing
of Environment, 80, pp. 185–201.
Foody, G. M., and A. Mathur. 2004. Toward intelligent training of supervised image
classifications: Directing training data acquisition for SVM classification. Remote
Sensing of Environment 93, pp.107-117
161
Foody, G. M., Boyd, D. S., and Sanchez-Hernandez, C., 2007. Mapping a specific class with
an ensemble of classifiers. International Journal of Remote Sensing 28 (8), pp. 1733-
1746.
Forster, B. C., 1986. Evaluation of Combined Multiple Incident Angle SIR-B Digital Data
and Landsat MSS Data over an Urban Complex. Proceeding of the 7th International
Symposium Remote Sensing for Resources Development and Environment
Management, Enschede, Netherlands, 25-29 August, pp. 813- 816.
Forster, B. C., 1987. Applications in Urban Monitoring and Engineering. Microwave Remote
Sensing, Course Notes, The University of New South Wales, School of Geomatic
Engineering, pp. 50 - 53.
Freund, Y., and Schapire, R. E., 1996. Experiments with a new boosting algorithm.
Proceedings of the Thirteenth International Conference of Machine Learning, pp. 148-
156.
Frost, V. S., Stiles, J. A., Shanmugan, K. S., and Holtzman, J. C., 1982. A Model for Radar
Images and Its Application to Adaptive Digital Filtering of Multiplicative Noise.
IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-4 (2), pp.
157-166.
Gamba, P., Fabio Dell’Acqua, F., Dasarathy, B. V., 2005. Urban remote sensing using
multiple data sets: Past, present, and future. Information Fusion 6, pp.319–326.
Gao, J., 2009, Digital Analysis of Remotely sensed imagery. McGraw-Hill Companies, Inc.
Gimeno, M., San-Miguel-Ayanz, J., and Schmuck, G., 2004, Identification of burnt areas in
Mediterranean forest environments from ERS-2 SAR time series. International
Journal of Remote Sensing 25 (22), pp. 4873-4888.
Gomez-Chova, L., Calpe, J., Camps-Valls, G., Martin, J.D., Soria, E., Vila, J., Alonso-
Chorda, L., Moreno, J., 2003. Feature Selection of Hyperspectral Data Through
Local Correlation and SFFS for Crop Classification. International Geoscience and
Remote Sensing Symposium (IGARSS), 1, pp. 555-557.
Gong, P., and P. J. Howarth, 1990. An Assesement of Some Factors Influencing
Multispectral Land-Cover Classification. Photogrammetric Engineering and Remote
Sensing 56 (1), pp. 597-603.
Goldberg, D., 1989, Genetic algorithm in Search, Optimization in Machine Learning
Reading, MA: Addison Wesley.
162
Guerschman, J.P., Paruelo, J. M., Di Bella, C., Giallorenzi, M. C., and Pacin, F., 2003. Land
cover classification in the Argentine Pampas using multi-temporal Landsat TM data,
International Journal of Remote Sensing 24 (17), pp. 3381-3402.
Guyon, I., Weston, J., Barnhill, S., and Vapnik, V., 2002. Gene selection for cancer
classification using support vector machines. Machine Learning 46, pp. 389–422.
Fundamentals of Remote Sensing, a Canada Centre for Remote Sensing Remote Sensing
Tutorial. Available at: http://www.nrcan.gc.ca/sites/www.nrcan.gc.ca.earth-
sciences/files/pdf/resource/tutor/fundam/pdf/fundamentals_e.pdf.
Haack, B. N., 1984. Multisensor Data Analysis of Urban Environments. Photogrammetric
Engineering and Remote Sensing 50 (10), pp. 1471 - 1477.
Haack, B., and Bechdol, M., 2000. Integrating multisensor data and RADAR texture
measures for land cover mapping. Computers & Geosciences 26, pp. 411- 421.
Haack, B., and Khatiwada, G., 2010. Comparison and integration of optical and
quadpolarization radar imagery for land cover/use delinieation. Journal of Applied
Remote Sensing 4 (1), pp. 1-16.
Hall-Beyer, M., 2007, GLCM Texture Tutorial; Version: 2.10, University of Calgary.
Available from: http://www.fp.ucalgary.ca/mhallbey/tutorial.htm.
Henderson, F.M., and A.J. Lewis (editors), 1998. Principles & Applications of Imaging
Radar, Manual of Remote Sensing, Third edition, Volume 2, American Society for
Photogrammetry and Remote Sensing, John Wiley & Sons, Inc., New York.
Herold, N. D., Haack, B. N., and Solomon, E., 2005. Radar spatial considerations for land
cover extraction. International Journal of Remote Sensing, 26 (7), pp.1383-140.
Hughes, G. F., 1968, On the mean accuracy of statistical pattern recognizers, IEEE
Transactions on Information Theory, 14(1), pp. 55–63.
Huang, C., Davis, L. S., Townshend, J. R. G., 2002. An assessment of support vector
machines for land cover classification. International Journal of Remote Sensing 23
(4), pp. 725–749.
Huang, Z. and Lees, B.G., 2004, Combining non-parametric models for multisource
predictive forest mapping. Photogrammetric Engineering and Remote Sensing, 70,
pp. 415–425.
Huang, C.L., Wang, C.J., 2006. A GA-based feature selection and parameters optimization
for support vector machines. Expert Systems with Applications 31, pp. 231-240.
163
Huang, H., Legarsky, J., and Othman, M., 2007. Land-cover Classification Using Radarsat
and Landsat Imagery for St. Louis, Missouri. Photogrammetric Engineering &
Remote Sensing 73 (1), pp. 037–043.
Hugo, C., Capao, L., Fernando, B. and Mario, C., 2006. Meris Based Land Cover
Classification with Self-Organizing Maps: preliminary results, Proceedings of the 2nd
EARSeL SIG Workshop on Land Use & Land Cover, 28 - 30 September, Bonn,
Germany.
Hussin, Y. A., 1997. Tropical Forest Cover Typees Differntiation using Satellite Optical and
Radar Data:A Case Study From Jambi, Indonesia. Proceeding of Asian Conference of
Remote Sensing (ACRS), Kuala Lumpur, Malaysia, 20-24 November.
Heinl, M., Walde, J., Tappeiner, G., and Tappeiner, U., 2009. Classifiers vs. input variables-
The drivers in image classification for land cover mapping. International journal of
Applied Earth Observation and Geoinformation 11, pp. 423-430.
Jensen, J.R., 2004. Introduction to Digital Image Processing: A remote sensing perspective,
4th edition, Prentice Hall.
Kavzoglu, T., and Mather, P. M., 2002. The role of feature selection in artificial neural
network applications. International journal of Remote sensing 23 (15), pp. 2919-2937.
Kavzoglu, T., and Mather, P. M., 2003. The use of backpropagating artificial neural network
in land cover classification. International journal of Remote sensing 24 (23), pp.
4907-4938.
Kavzoglu, T., 2009. Increasing the accuracy of neural network classification using refined
training data. Environmental Modelling & Software, 24 (7), pp 850-858.
Kavzoglu, T., and Colkesen, I., 2009. A kernel functions analysis for support vector machines
for land cover classification, International Journal of Applied Earth Observation &
Geoinformation 11 (5), pp. 352–359.
Kim, S.H., and Lee, K.S., 2005. Integration of Landsat ETM+ and Radarsat SAR for land
cover classification. Trans Tech Publications, Switzerland,
doi:10.4028/www.scientific.net/KEM.277-279.838.
Lee, J.-S., 1980. Digital Image Enhancement and Noise Filtering by Use of Local Statistics,
IEEE Transactions on Pattern Analysis and Machine Intelligence, .2 (2), pp. 165-168.
Lee, J.-S., 1981. Speckle Analysis and Smoothing of Synthetic Aperture Radar Images,
Computer Graphics and Image Processing 17 (1), pp. 24-32.
J. S. Lee, 1983. A simple speckle smoothing algorithm for synthetic aperture radar images
IEEE Transactions on Systems, Man, and Cybernetic SMC-13 (1), pp. 85–89.
164
Lehmann, E., Caccetta, P., Zhou, Z. –S., Mitchell, A., Tapley, I., Milne, A., Held, A., Lowell,
K., and McNeill, S., 2011. Forest Discrimination Analysis of Combined Landsat and
ALOS-PALSAR Data. Proceeding of the 34th International Symposium on Remote
Sensing of Environment, Sydney, Australia, 10-15 April.
Lemonie, G.G., de Grandi, G. F., and Sieber, A. j., 1994. Polarimetric contrast classification
of agricultureal fields using MASTRO1 AirSAR data, International journal of remote
sensing, Vol. 24 (1), pp. 28-35.
Li, G., Lu, D., Moran, E., and Hetrick, S., 2011, Land-cover classification in a moist tropical
region of Brazil with Landsat TM imagery, International Journal of Remote Sensing,
Vol. 32 (23), pp. 8207-8230.
Lillesand, T. M., and Kieffer, R. V., 2004. Remote sensing and Image Interpretation, Fourth
edition, John Wiley and Sons, New York, 2004.
Lloyd, C. D., Berberoglu, S., Curran, P. J., and Atkinson, P. M., 2004. A comparision of
texture measures for the per-field classification of Mediterranean land cover.
International Journal of Remote Sensing 25 (19), pp. 3943-3965.
Lopes, A., Touzi, R., and Nezry, E., 1990, Adaptive Speckle Filters and Scene Heterogeneity,
IEEE Transactions on Geoscience & Remote Sensing, Vol. 28 (6), pp. 992-1000.
Luo, C., Liu, Z., and Meng, W., 2008. A integrated classifier for large regional-scale land
cover classifying. The International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences. Vol. XXXVII. Part B6b. pp.607-612.
Lu, D., Moran, E., and Batistella, M., 2003, Linear mixture model applied to Amazonian
vegetation classification. Remote Sensing of Environment, Vol. 87, pp. 456-469.
Lu, D., Batistella, M., and Moran, E., 2007. Land-cover classification in the Brazilian
Amazon with the integration of Landsat ETM+ and Radarsat data. International
Journal of Remote Sensing 28 (24), pp. 5447-5459.
Lu, D., Li, G., Moran, E., Dutra, L., and Batistella, M., 2011, A Comparison of Multisensor
Integration Methods for Land Cover Classification in the Brazilian Amazon. GIScience & Remote Sensing 48 (3), pp. 345–370.
Maghsoudi, Y., Alimohammadi, A., Valadan Zoej, M. J., and Mojaradi, B., 2006. Weighted
combination of multiple classifiers for the classification of hyperspectral images using
genetic algorithm. In : ISPRS Archives, Vol XXXVI(part 7)- Proceeding of the
ISPRS Commission VII Symposium ‘Remote Sensing: From Pixels to Processes’,
May 8-11, Enschede, The Netherlands.
165
Maghsoudi, Y., Valadan Zoej, M. J., and Michael Collins, M., 2011. Using class-based
feature selection for the classification of hyperspectral data, International Journal of
Remote Sensing, 32 (15), pp. 4311-4326
Maillard, P., Alencar-Silva, T. and Clausi, D., 2008. An evaluation of Radarsat-1 and ASTER
data for mapping veredas (palm swamps). Sensors, 8 (9), pp.6055-6076.
Marceau, D.J., Howarth, P.J., and D.J. Gratton, D. J., 1994a. Remote sensing and the
measurement of geographical entities in a forest environment 1: The scale and spatial
aggregation problem, Remote Sensing of Environment, Vol. 49, pp. 49:93–104.
Marceau, D.J., Howarth, P.J., and D.J. Gratton, D. J., 1994b. Remote sensing and the
measurement of geographical entities in a forest environment 2: The optimal spatial
resolution, Remote Sensing of Environment, Vol. 49, pp. 105–117.
Markham, B. L., and Townshend, J. R. G., 1981, Land cover classification accuracy as a
function of sensor spatial resolution. Proceedings of the 15th International Symposium
on Remote Sensing of Environment, Ann Arbor, MI, 11–15 May (Ann Arbor: ERIM),
pp. 1075–1090.
Markowetz, F., 2003, Classification by SVM, Practical DNA Microarray Analysis,
Available at: compdiag.molgen.mpg.de/ngfn/docs/2003/mar/SVM.pdf.
McNairna, H., Catherine Champagnea, C., Shanga, J., Holmstromb, D., Reichert, G.,2009.
Integration of optical and Synthetic Aperture Radar (SAR) imagery for delivering
operational annual crop inventories. ISPRS Journal of Photogrammetry & Remote
Sensing 64, pp. 434-449.
Melgani, F., and Bruzzone, L., 2004. Classification of hyperspectral remote sensing images
with support vector machine, IEEE Transactions on Geosciences and Remote Sensing
42 (8), pp. 1778-1790.
Mills, H., Cutller, M.E.J., and Fairbairn, D., 2006. Artificial neural network for mapping
regionalscale upland vegetation from high spatial resolution imagery, International
Journal of Remote Sensing 27 (11), pp 2177-2195.
Mountrakis, G., Im, J. and Ogole, C., 2011, Support vector machines in remote sensing: A
review. ISPRS Journal of Photogrammetry and Remote Sensing Vol. 66, pp. 247–
259.
Narasimha Rao, P.V., Sesha Sai, M.V.R., Sreenivas, K., Krishna Rao, M.V., Rao, B.R.M.,
Dwivedi, R.S. and Venkataratnam, L., 2002. Textural analysis of IRS-1D
panchromatic data for land cover classification. International Journal of Remote
Sensing, 23 (17), pp. 3327–3345.
166
Nizalapur, V., 2008, Land cover classification using multi-source data fusion of Envisat
ASAR and IRS P6 Liss-III Satellite data – a case study over tropical moist deciduos
forested region of Karnataka, India. The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII.
Part B6b. pp. 329-334.
Nizalapur, V., Maduundu, R., and Jha, C. S., 2011. Coherence-based land cover classification
in forested area of Chttigarh, Central India, using environment satellite – advanced
synthetic aperture radar data. Journal of Applied Remote Sensing 5, pp. 059501-1-6.
Nordkvist, K., Granholm, A. H., Holmgren, J., Olsson, H., & Nilsson, M., 2012, Combining
optical satellite data and airborne laser scanner data for vegetation classification,
Remote Sensing Letters 3(5), pp. 393-401.
Nyoungui, A., Tonye, E. and Akono, A., 2002. Evaluation of speckle filtering and texture
analysis methods for land cover classification from SAR images. International Journal
of Remote Sensing 23 (9), pp. 1895–1925.
Osowski, S., Siroic, R., Markiewicz, T., and Siwek, K., 2009. Application of Support Vector
Machine and Genetic Algorithm for improved blood cell recognition, IEEE
Transactions on Instrumentation and Measurement 58 (7), pp.2159 – 2168.
Osuma, R., 2005. Pattern Analysis, Sequential Feature Selection, Lecture note. Texas A&M
University.
Pacifici, P., Chini, M., Emery, J. M, 2009. A neural network approach using multi-scale
textural metrics from very high-resolution panchromatic imagery for urban land-use
classification. Remote Sensing of Environment 113, pp. 1276–1292
Pal, M. and Mather, P.M., 2005. Support vector machines for classification in remote
sensing, International Journal of Remote Sensing 26 (5), pp. 1007–1011.
Pal, M., 2006. Support vector machine-based feature selection for land cover classification: a
case study with DAIS hyperspectral data', International Journal of Remote Sensing,
27 (14), pp. 2877-2894.
Pal, M., 2008, Ensemble of support vector machines for land cover classification,
International Journal of Remote Sensing 29 (10), pp. 3043-3049.
Paola, J.D., and Schowengerdt, R.A., 1995. A review and analysis of back propagation
neural networks for classification of remotely sensed multispectral imagery.
International Journal of Remote Sensing 16 (16), pp. 3033–3058.
167
Park, N. –W., Chi, K. –H., 2008. Integration of multitemporal/polarization C– band SAR data
sets for Land cover Classification. International Journal of Remote Sensing 29 (16),
pp. 4667-4688.
Peddle, D. R., and Ferguson, D. T., 2002, Optimization of multisource data analysis: an
example using evidence reasoning for GIS data classification. Co mp u t e r s
& Geosciences 28, pp. 45-52.
Pierce L.E., Bergen K.M., Dobson M.C., and Ulaby F.T., 1998, Multitemporal land-cover
classification using SIR-C/X-SAR imagery, Remote Sensing of Environment 64 (1),
pp. 20-33.
Polikar, R., 2006. Ensemble Based Systems in Decision Making, IEE Circuits and Systems
Magazine 6 (3), pp. 21-45.
Pouteau, R., Stoll, B., and Chabrier, S., 2010. Multi-source SVM fusion for environmental
monitoring in Marquesas archipelago. Proceeding of the International Geoscience and
Remote Sensing Symposium (IGARSS), pp. 2719-2722. Honolulu, HI, 25-30 July.
Pradhan, R., Pradhan, M. P., Bhusan, A., Pradhan, R. K., and Ghose, M. K., 2010. Land-
cover classification and mapping for Eastern Himalayan State Sikkim. Journal of
Computing 2 (3), pp. 166-170.
Puissant, A., Hirsch, J., and C Weber, C., 2005. The utility of texture analysis to improve per-
pixel classification for high to very high spatial resolution imagery. International
Journal of Remote Sensing 26 (4), pp.733–745.
Quegan, S., Toan, T. L., Yu, J. J., Ribbes F., and Floury, N., 2000, Multitemporal ERS SAR
Analysis Applied to Forest Mapping, IEEE Transactions on Geoscience and Remote
Sensing 38 (2), pp. 741-753.
Qiu, F., Berglund, J., Jensen, J. R., Thakkar, P., and Ren, D., 2004, Speckle Noise Reduction
in SAR Imagery Using a Local Adaptive Median Filter, GIScience and Remote
Sensing 41 (3), pp. 244-266.
Richards,J. A., Jia, X., 2006. Remote Sensing Digital Image Analysis – An introduction.
Springer, Fourth Edition.
Shafri, H. Z. M., and Ramle, F. S. H., 2009. A comparison of Support Vector Machine and
Decision Tree Classifications Using Satellite Data of Langkawi Island, Information
Technology Journal 8 (1), pp. 64-70.
Salah, M., Trinder, J.C., Shaker, A., Hamed, M., and Elsagheer, A., 2010. Integrating
multiple classifiers with fuzzy majority voting for improved land cover classification.
168
ISPRS Int Arch Photogrammetric Engineering and Remote Sensing & SIS 39 (3), Part
A, pp.7-12.
Salah, M., Trinder, J. and Shaker, A., 2009. Evaluation of the self-organizing map classifier
for building detection from lidda data and multispectral aerial images. Journal of
Spatial Science, 54, pp. 15-34.
Santos, J.R., Freitas, C.C., Araujo, L.S., Dutra, L.V., Mura, J.C., Gama, F.F.,Soler, L.S. and
Sant’Anna, S.J.S., 2003, Airborne P-band SAR applied to the aboveground biomass
studies in the Brazilian tropical rainforest. Remote Sensing of Environment 87, pp.
482–493.
Schapire, R. E., 1990. The strength of weak learnability. Machine Learning 5 (2), pp. 197-
227.
Shang, J., McNairm, H., Champagne, C., and Jiao, X., 2009. Appliation of Multi-Frequency
Synthetic Aperture Radar(SAR) in Crop Classification, Chapter 27, Advances in
Geoscence and Remote Sensing, Published: October 01 by InTech.
Shang, J., McNairm, H., Champagne, C., and Jiao, X., 2008. Contribution of multi-
frequency, multi-sensor, and multi-temporal radar data to operational annual crop
mapping. Proceeding of the International Geoscience and Remote Sensing
Symposium, Boston (Massachusetts), 7-11 July, 4 pp. Invited paper.
Shang, J., McNairm, H., Champagne, C., and McNairm, H., 2006. Agriculture land use using
multi-sensor and multi-temporal Earth Observation data. Proceeding of the
MAPPS/ASPRS 2006 Fall Specialty conference, San Antonio, Texas, USA.
Shafri, H. Z. M., and Ramle, F. S. H., 2009. A comparison of Support Vector Machine and
Decision Tree Classifications Using Satellite Data of Langkawi Island, Information
Technology Journal 8 (1), pp. 64-70.
Sheoran, A., Haack, B., and Sawaya, S., 2009. Land cover/use classification using optial and
quad polarization radar imagery. Proceeding of the ASPRS 2009 Annual
Conferencce, Baltimore, Maryland, 9 -13 March.
Shin, J., Yoon, J., Kang, S., and Lee, K., 2007. Comparison of Signal Characteristics between
X-, C-, and L-band SAR Data in Forest Stand. Available at: http://www.a-a-r-
s.org/acrs/proceeding/ACRS2007/Papers/TS30.3.pdf.
Serpico, S. B., and Bruzzone, L., 2001. A new search algorithm for feature selection in
hyperspectral remote sensing images. IEEE Transactions on Geoscience and Remote
Sensing, 39 (7), pp. 1360-1367.
169
Siegert, F., and Hoffmann, A. A., 2000. The 1998 Forest Fires in East Kalimantan
(Indonesia): A Quantitative Evaluation Using High Resolution, Multitemporal ERS-2
SAR images and NOAA-AVHRR Hotspot Data. Remote Sensing of Environment, 72,
pp-64-67.
Siegert, F., and Ruecker, G., 2000. Use of multitemporal ERS-2 SAR images for identication
of burned scars in south-east Asian tropical rainforest. International Journal of
Remote Sensing 21 (4), pp. 831-837.
Skriver, H., 2008. Comparison between multi-temporal and polarimetric SAR data for land
cover classification. Proceeding of the International Geoscience and Remote Sensing
Symposium (IGARSS), Boston (Massachusetts), 7-11 July. pp.558-561.
Somol, P., Novovicova, J., and Pudil, P., 2006. Flexible-Hybrid Sequential Floating Search in
Statistical Feature Selection. Lecture Notes in Computer Science 4109/2006, pp. 632-
639, DOI: 10.1007/11815921_69.
Soria-Ruiz, J., Fernadez-Ordonez, Y., and Woodhouse, I. H., 2010. Land-cover lassification
using radar and optical images: a case study in Central Mexico. International Journal
of Remote Sensing 31 (12), pp. 3291–3305.
Srivastava, H. S., Patel, P., Navalgund, R. R., 2006. Application potentials of synthetic
aperture radar interferometry for land-cover mapping and crop-height estimation.
Current Science 91 (6), pp.783-788.
Steninger, M.K., 2000, Satellite estimation of tropical secondary forest aboveground biomass
data from Brazil and Bolivia. International Journal of Remote Sensing 21 (6),
pp.1139–1157.
Tassetti, A. N., Malinverni, E. N., and Hahn, M., 2010. Texture analysis to improve
supervised classification in IKONOS imagery, ISPRS TC VII Symposium – 100
Years ISPRS, Vienna, Austria, 5-7 July, IAPRS, Vol. XXXVIII, Part 7A, pp. 245-
250.
Takeuchi, S., Suga, Y., and Yoshimura, M., 2001. A comparative study of coherence
information by L band and C band SAR for detecting deforestation in tropical rain
forest. Proceeding of the International Geosciences and Remote Sensing Symposium
IGARSS’01, Sydney, Australia, 9-13 July, Vol. 5, pp. 2259-2261.
Takeuchi, S., and Yamada, S., 2002, Monitoring of Forest Fire Damage by Using JERS-1
InSAR. IEEE International Geosciences and Remote Sensing Symposium, IGARSS
’02, Torongto, Canada, 24-28 June, pp. 3290-3292.
170
Thiel, C. J., Thiel, C., and Schmullius, C. C., 2009. Operational Large-area forest Monitoring
in Siberia Using ALOS PALSAR Summer intensity and winter Coherence, IEEE
Transactions on Geoscience and Remote Sensing 47 (12), pp. 3993-4000.
Trinder, J.C., Salah, M., Shaker, A., Hamed, M., and Elsagheer, A., 2010. Combining
statistical and neural classifier using Dempster-Shafer theory of evidence for
improved building detection. Proceeding of 15th Australasian Remote Sensing &
Photogrammetry Conference (15th ARSPC), Alice Springs, Australia, 13-17
September.
Tzeng, Y. C., Chiu, S. H., and Chen, K. S., 2006. Improvement of Remote Sensing Image
Classification Accuracy by using a Multiple Classifiers System with Modified
Bagging and Boosting Algorithms. Proceeding of the International Geoscience and
Remote Sensing Symposium (IGARSS), Denver, Colorado, USA, 31 July- 04 August,
pp. 2758-2761.
Tso, B., and Mather, P., 2009. Classification Methods for remotely sensed data. Second
edition, CRC Press, Taylo & Francis Group.
Vapnik, V. N., 2000. The Nature of Statistical Learning Theory. Springer, New York,
(chapter 8)
Waske, B., and J. A. Benediktsson, J. A., 2007. Fusion of Support Vector Machines for
classification of multisensor data, IEEE Transactions on Geosciences & Remote
Sensing, 45 (12), pp. 3858-3866.
Waske, B. & Braun, M. (2009): Classifier ensembles for land cover mapping with multi-
spectral SAR imagery. International Journal of Photogrammetry and Remote Sensing
64 (5), pp. 450-457.
Wang, Y. C., Feng, C. C., and Vu, D. H., 2012. Integrating Multi-Sensor Remote Sensing
Data for Land Use/Cover Mapping in a Tropical Mountainous Area in Northern
Thailand. Geographical Research 50 (3), pp. 320-331.
Weydahl, D. J., Becquey, X., and Tollefsen, T., 1995. Combining ERS-1 SAR with Optical
Satellite Data over Urban Areas. Proceedings of International Geoscience and Remote
Sensing (IGARSS) 4, pp. 2161- 2163.
Woodcock, C.E., and Strahler, A.H., 1987. The factor of scale in remote sensing, Remote
Sensing of Environment, Vol. 21(3), pp. 311–332.
Woodcock, C.E., and Harward, V.J., 1992. Nested-hierarchical scene models and image
segmentation, International Journal of Remote Sensing, Vol. 13(16), pp. 3167–3187.
171
Yamaguchi, T., Kishida, K., Nunohiro, E., Park, J. G., Mackin, K. J., Hara, K., Matsushita,
K., and Harada, I., 2010. Artificial Neural networks paddy-field classifier using
spatiotemporal remote sensing data. Artif Life Robotics 15, pp. 221-224.
Yan, W. Y., and Shaker, A., 2011. The effects of combining classifiers with the same training
statistics using Bayesian decision rules, International Journal of Remote Sensing, 32
(13), pp. 3729-3745.
Zhang, Y., Chen, L., and Yu, B., 2007. Multi-scale texture analysis for urban land use/cover
classification using high spatial resolution satellite data, Geoinformatics 2007:
Remotely Sensed Data and Information, Proceeding of the SPIE 6752, pp. 67523G,
doi:10.1117/12.761232.
Zhou, M., Shu, J., Chen, Z., 2010. Classification of hyperspetral remote sensing image based
on Genetic Algorithm and SVM. Remote Sensing & Modeling of Ecosystems for
Sustainability VII, Proceeding of the SPIE 7809, pp.78090A.
Zhou, L., and Yang, X., 2008. Use of Neural networks for land cover classification from
remotely sensed imagery. The International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008, 575-
578.
Zhuo, L., Zheng, J., Wang, F., Li, X., Ai, B., Qian, J., 2008. A Genetic Algorithm based
wrapper feature selection method for classification of hyperspectral images using
support vector machine. The International Archives of the Photogrammetry, Remote
Sensing & Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing, pp.397-402.