I

MAa

b

c

a

ARRA

KCABG

1

ebimcwOacpaa

otsu

0d

Computers and Electronics in Agriculture 66 (2009) 140–146

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journa l homepage: www.e lsev ier .com/ locate /compag

ntelligent electronic nose system for basal stem rot disease detection

.A. Markoma,∗, A.Y. Md Shakaff a, A.H. Adomb, M.N. Ahmadc, Wahyu Hidayata,.H. Abdullahb, N. Ahmad Fikrib

School of Computer and Communication Engineering, Universiti Malaysia Perlis, P.O. Box 77, Pejabat Pos Besar Kangar, 01007, Perlis, MalaysiaSchool of Mechatronic Engineering, Universiti Malaysia Perlis, P.O. Box 77, Pejabat Pos Besar Kangar, 01007, Perlis, MalaysiaSchool of Material Engineering, Universiti Malaysia Perlis, P.O. Box 77, Pejabat Pos Besar Kangar, 01007, Perlis, Malaysia

r t i c l e i n f o

rticle history:eceived 18 July 2008eceived in revised form 30 December 2008ccepted 10 January 2009

eywords:

a b s t r a c t

The agricultural industry has been, for a long time, dependent upon human expertise in using odour forclassification, grading, differentiating and discriminating different types of produce. Odour as a parameterof differentiation can also be used to determine the state of health of crops, although this is not favourablewhen dealing with detecting plant disease that may pose health threats to human beings. In addition tothese, human experts may take years of training and can be inconsistent, as well as prone to fatigue. This

ommercial electronic noseNNasal stem rot diseaseanoderma boninense

paper presents a work conducted on utilising an electronic nose incorporating artificial intelligence todetect plant disease, specifically basal stem rot (BSR) disease that is caused by Ganoderma boninense fun-gus affecting oil palm plantations in South East Asia. This study used a commercially available electronicnose, Cyranose 320, as the front end sensors and artificial neural networks for pattern recognition. Theodour samples were captured on site at Besout oil palm plantation, Perak, Malaysia, and the classificationperformed on a PC. The results showed that the system was able to differentiate healthy and infected oil

odou

palm tree using different

. Introduction

In recent years, the development of innovative devices such aslectronic nose (e-nose) has been investigated and implementedy many researchers. The electronic nose is often referred to as anntelligent device, able to mimic the human olfaction functions and

ay be used for detection, recognition and classification of volatileompounds and odours. This type of electronic olfactory systemas introduced in 1982 by Dodd and Persaud from the Warwicklfaction Research Group, UK (Troy Nagle et al., 1998). Basically,n e-nose consists of sensing elements (an array of sensors), signalollecting unit(s) and suitable pattern recognition algorithm. Theossible implementations of these devices are wide ranging, fromgricultural applications to solving environmental issues (Dutta etl., 2005).

The agricultural industry can benefit tremendously from the usef e-noses. An e-nose may provide rapid, automated as well as realime classification and detection. Also, it can be easily re-trained touit different applications to detect a variety of odours. Hence these of this system may replace traditional methods that are labour

∗ Corresponding author. Tel.: +60 4 9798931; fax: +60 4 9798099.E-mail address: azira putra@yahoo.com (M.A. Markom).

168-1699/$ – see front matter © 2009 Elsevier B.V. All rights reserved.oi:10.1016/j.compag.2009.01.006

r parameters with a high rate of accuracy.© 2009 Elsevier B.V. All rights reserved.

intensive, inaccurate, impractical and time consuming (Brezman etal., 2005).

In terms of agricultural applications, the e-nose systems havebeen used successfully to assist product quality monitoring, fruitripeness determination as well as inspection of fish (Keller et al.,1995). One of the most successful experimental applications of thee-nose system is fruit ripeness classification. Basically, the mainobjective of this research is to avoid large quantities of fruit beingharvested too soon or too late and in turn, reaching consumer mar-kets in poor condition (Berna et al., 2003).

The e-nose has also been used to detect the causal agent orelement of plant disease such as fungi and bacteria. Some workhas been reported on using the e-nose to identify Penicillium,Aspergillus and Fusarium (Falasconi et al., 2005; Balasubramanianet al., 2007). These fungi produce mycotoxins on agricultural com-modities during plant growth or after harvest during storage andshipment. It could be detrimental to living organism when ingested(Balasubramanian et al., 2007). Also, the growth of the moulds ingrains reduces their nutritional quality and subsequent decrease ingermination results in economic loss.

This paper presents a feasibility study of a novel use of an e-nose

to detect plant disease, specifically the basal stem rot (BSR) dis-ease, within its environment. The disease is a threat to the oil palmindustry causing an estimated annual lost of RM80 million (Adomet al., 2006). This study utilises a commercial e-nose, Cyranose 320incorporated artificial neural networks.

lectronics in Agriculture 66 (2009) 140–146 141

2

ofihnr

BaaatUcidt

wbmfIi

sabtxal

14Tpaso

at

M.A. Markom et al. / Computers and E

. Basal stem rot disease

The oil palm (Elaeis guineensis Jacq.) tree is a leading sourcef edible vegetable oil production in the world, with productiongures of more than 32 million tonnes in 2003. In Malaysia, theierarchy of palm oil as a leading cash crop has exceeded that ofatural rubber, and its importance has been further boosted by theecent introduction of bio-diesel (Adom et al., 2006).

Its cultivation, in much of South-East Asia, is threatened by theSR caused by G. boninense, where losses can reach as much as 80%fter repeated planting cycles. BSR has been causing serious dam-ge to the oil palm plantations in Malaysia for more than 50 yearsnd is currently the most important disease of economical impor-ance causing large amount of losses in revenue (Miller et al., 1994).nder severe infestation situations, more than 50% of oil palm standan be lost to the disease. The Federal Land Development Author-ty (FELDA) plantations, for example, recorded high incidence of theisease in the Peninsular Malaysia, ranging to about 50%, from 1994o 2005.

The disease does not show any indications of early infectionhereby it progresses through the palm from the base. The visi-

le disease symptoms appear at a very late stage of infection whenore than half of the root tissues have decayed, leaving no chance

or the grower to cure the infected palms (Karthikeyan et al., 2006).t causes root and stem rot in many tropical perennial crops, includ-ng coconut, betel palm, rubber, tea, coffee, cocoa and forest trees.

The causal agent, G. boninense, a white-rot basidiomycete, is aaprophyte or weak parasite that can infect a living palm if therere massive inoculums as shown in Fig. 1 (Shu’ud et al., 2007). The G.oninense produces enzymes that will degrade the palm tissue. Ashe fungus destroys the internal palm tissues, it will affect the palmylem, thus causing serious problems to the distribution of waternd other nutrients to the top of the palm tree, which eventually,eads to the death of the oil palm.

It has been demonstrated that the fungus can set in as early as2–24 months after planting but more usually when the oil palm is–5 years old, particularly in replanted areas (Shu’ud et al., 2007).he incidence increases rapidly to the extent that by the time thealms were 15 years old, the disease levels can reach between 40nd 50 percent of the palm. In severe cases, up to 85 percent of thetanding palms succumb to BSR by the time the palms are 25 years

ld.

Until today, a few methods based on biochemistry processesre used to detect ganoderma infection, which can be divided intowo categories. The first is culture, such as Ganoderma Selective

Fig. 1. Ganoderma boninense fruiting body.

Fig. 2. Cyranose 320 (Shu’ud et al., 2007).

Medium (GSM). The second category is molecular DNA, such as apolymerase chain reaction (PCR). However, these methods requirethe stem collection for further tests in the laboratories. There has yetto be a method which is able to give results in real time or directlyon-site (Shu’ud et al., 2007).

3. System descriptions

This section will describe the system used in this study whichconsists of a commercial e-nose, Cyranose 320 (refer to Fig. 2(Cyranose, 2008)), and artificial neural networks.

3.1. Cyranose 320

The Cyranose 320 (C-320) is a handheld e-nose instrument madeby Cyrano Sciences Inc. (USA). It is widely used in quality con-trol, process measurement, hazardous material identification andbiomedical sample discrimination (Boilot et al., 2002). The maincomponents of C-320 include sampling, sensory and signal pro-cessing components as well as pattern recognition algorithm.

The sensing component consists of an array of sensors with32 elements. All sensors are conducting-based polymer and eachsensor has different response specificity to a broad range of com-pounds. Therefore, this kind of e-nose is able to produce limitlessnumber of output signals, which is referred to as the signal patternor fingerprint (Boilot et al., 2002).

The embedded pattern recognition algorithms offered by C-320were not employed because initial tests show that these were notable to produce high enough accuracy rates compared to the ANNused.

3.2. Pattern recognition

Apart from the statistical approach, artificial neural networks(ANNs) have been widely used as a pattern recognition tool. It has

been reported to produce good performance which showed promis-ing results in chemical vapour recognition (Scott et al., 2007).

It is an interconnected group of artificial neurons that usesa mathematical or computational model for data processing. Inpractical terms, neural networks are non-linear statistical data

142 M.A. Markom et al. / Computers and Electronics in Agriculture 66 (2009) 140–146

collec

mr(

tatglrth

4

isi

4

tFafiodi

atiddes

dsl

Fig. 3. Data

odelling tools. They offer black-box approach to model complexelationships between inputs and outputs or to find patterns in dataHines et al., 1999).

In this study, the Multi-Layer Perceptron (MLP) which wasrained using the Levenberg–Marquardt algorithm. This traininglgorithm was chosen because it offers the best performance duringraining in terms of speed and efficiency. It has a faster conver-ence rate compared to the conjugate gradient algorithm and henceeading to faster optimisation (Emilia Noorsal, 2005). Although itequired large storage and intensive computation, it is still effec-ive for moderate-sized networks, which may contain up to a fewundred weights.

. Methodology

The work conducted was divided into five main stages whichnvolved data and sample collection, consistency analysis, featureelection, normalisation and ANN training as well as network test-ng.

.1. Data and sample collection

The ganoderma odour including that of its environment, such ashe palm oil tree trunk and soil (data collection), was performed atELDA Besout 7 oil palm plantation, Sungkai, Perak, Malaysia. Therea of interest in the plantation was divided into two sections; therst part would be the normal or healthy plant area while the sec-nd was the infected area. The plantation management has clearlyefined areas of infection, and the chosen distances between the

nfected and healthy areas are around 200 m.Three types of odour samples were selected for both healthy

nd infected oil palm trees, namely odour of the air surrounding theree, odour of bored tree trunk and as well as odour of soil surround-ng the base of the tree trunks. Each of these odours represents aifferent parameter of the tree odour. Three trees were chosen ran-omly from each section. Also, for each tree sample, three points forach odour parameter were marked as the data collection points as

hown in Fig. 3.

The odours of the three different parameters representing theifferent parts of the tree were collected using C-320 on-site. Theseamples were also collected and placed in sealed plastic bags foraboratory analysis. The C-320 readings of these samples were per-

tion point.

formed in the laboratory. This will allow the comparison of theodour readings taken on site and at the laboratory.

The comparison between on-site and laboratory readings wereconducted to verify if there exists any physiological change of thesamples that may alter its odour profile. Also, the results of this testwill dictate whether the odour readings of samples taken on siteand in the laboratory are interchangeable. If the odour profiles ofall parameters are consistent whether they are taken on-site or inlaboratory, the researchers do not need to spend much time in plan-tation. They may focus on collecting odour sample without takenany data of sample at the site. The data can be collected after thesample was brought back to the laboratory.

The comparison of on-site and laboratory odour data was per-formed by plotting the raw sensor readings against the numberof sensors. The odour profile analyses were presented to verify theconsistency of the data recorded regardless of the point from whichthe readings were performed.

4.2. Dimension reduction

Dimension reduction is a process of identifying the most effec-tive subset of the original features to use in the classification processthat should lead to a higher classification accuracy. In this study,Principle Component Analysis (PCA) was used as the dimensionreduction technique as it is the most widely used method in e-nosesignal processing (Emilia Noorsal, 2005). PCA is a linear transfor-mation technique that aims to find a new set of orthogonal axis(principle components) for the input space given by odour pattern(Perera et al., 2002).

Reducing the number of sensors as the input to the ANN froma very large feature set can improve the system performance intwo ways. The first is to find a small set of features that can rep-resent efficiently the odour under analysis, hence disregarding theinsignificant sensor responses. Second is to reduce the number offeatures from response signal is necessary in order to reduce thecomputational complexity (Rodriguez, 2005). This will result in amore accurate and efficient ANN model. Including all sensor out-puts into the ANN without regard of its ability to discriminate the

odour samples will lead to decrease in ANN accuracy.

Only eight of high or significant sensor responses from the 32sensors in C-320 will be considered to be ANN’s inputs. This reduc-tion will make the execution time of training faster than using theinitial 32 inputs.

lectronics in Agriculture 66 (2009) 140–146 143

4

ttNv

aniito

Table 1ANN target.

Parameter Target

Healthy Infected

M.A. Markom et al. / Computers and E

.3. Normalisation and ANN analysis

The collected data are divided into two sets, one for training andhe other for validation/testing. These data sets were normalisedo ensure no dominance of any specific sensor to the ANN output.ormalisation adopted here changes the data mean to zero andariance of one.

Different network sizes were tested, with the number of inputsnd outputs set to eight and one, respectively. This is because the

umber inputs and outputs of the system are determined by the

mplementation. Since there are eight sensors selected to be thenput to the ANN after feature selection process, the network hashe same number of inputs. Also, since the sample implementationnly involve determining the healthy or infected sample, only a

Fig. 4. Onsite and laborato

Odour of surrounding tree 0 1Odour of bored trunk 0 1Odour of soil 0 1

single output is sufficient. ANN output of 1 will indicate positiverecognition while 0 being negative. Table 1 shows the output targetof the ANN.

For this project, the ANN implemented uses sigmoidal activationfunction, and thus limiting its output range from zero to one. Since

ry data comparison.

144 M.A. Markom et al. / Computers and Electronics in Agriculture 66 (2009) 140–146

) are t

ttlsdb0d

totdTaetn

Fig. 5. Data consistency for tree 1 to tree 3. Graph (a), (c), (e

he output of the ANN is analogue, and can vary anywhere from zeroo one, thresholding of the output was performed. If the output isess than 0.4, then it will be defined as zero that will indicate healthyample. If the output of the ANN is greater than 0.6, then, it will beefined as one to indicate the infected sample. Else, the output wille unclassified. The buffer range, in this case ranges from 0.4 to.6, can be varied according to the desired accuracy of the systemesigned.

At the beginning of the training, all of the weights were arbi-rarily chosen in the range of 0.25–0.75, which represents 50%f the output range of the nodes. Through learning iteration ofhe network, the weights are updated until they generate theesired output by comparing the output with the desired target.

his process determines a neural model. The training uses iter-tive weight update approach, and the network validated aftervery epoch. The sum of squared error (SSE) observed duringhis validation are used to determine the convergence of theetwork.

he healthy data and graph (b), (d), (f) are the infected data.

4.4. Network testing

The final stage was the network testing. The testing was per-formed using the testing data set. This will test the accuracy ofthe trained ANN model in discriminating and hence classifying thesamples.

5. Results and discussions

This section presents the results of the experiments performedfor this project. These include the comparison of laboratory and on-site odour data, the consistency of the odour data profile, the PCAas dimension reduction, ANN training as well as network testingresult.

5.1. The comparison on-site and laboratory data

The best graphical method to show this comparison is a 2Dline graph. This graph is obtained by plotting the mean of sensors

M.A. Markom et al. / Computers and Electronics in Agriculture 66 (2009) 140–146 145

F ) Profii

rrF

tbto

5

uF

3rpt

irsh

ig. 6. (a) Profile of healthy and infected data for odour surrounding the trunk. (bnfected for odour of soil.

esponse against the number of sensors. The lines in this graph rep-esent the data taken at laboratory as well as on-site as shown inig. 4.

From the plots obtained, the graphs do not show any similarity inhe laboratory and on-site e-nose data. This means that there muste some physiological changes on the samples during transit andhese caused different odour profiles to be recorded by the sensorsf C-320. Hence, only the data samples recorded on site were used.

.2. The data profile

The graphic representation of the data consistency is obtainedsing the 2D-line graph in Microsoft Excel software as shown inig. 5. Lines in the graph represent sensor outputs.

Fig. 5 shows that T1, T2 and T3 which represents Tree 1, 2 and, respectively have the same sensor responses. This means thategardless of the tree where the odour was captured, the odourrofile of the same parameter remains consistent. This indicateshat the data collected are reliable to be used for the ANN training.

The best graphical method to present the profile or fingerprint,nstead of a bar chart, is a radar plot as shown in Fig. 6. This rep-esentation can give better indication for comparing the responseignature. Fig. 6(a–c) shows that the healthy and infected samplesave their own odour profile.

le of healthy and infected data for odour of bored trunk. (c) Profile of healthy and

5.3. Dimension reduction

PCA method consists of expressing the response vectors in termsof a linear combination of orthogonal vectors that account for acertain amount of variance in the data. Fig. 7 shows the result ofapplying PCA to an array of 32 sensors when applied to three typesof BSR disease environment parameters. Two principle componentswere kept, which accounted for 99.32% of the variance in the dataset (PC1 and PC2 accounted for 74.67% and 24.65% of the variance,respectively).

5.4. Sensors selection

The selection of sensors is taken from the eight high values ofPCA coefficient. The calculation result of PCA values are shown inTable 2. The sensors that have been chosen are sensor 5, 6, 10, 12, 19,23, 28 and 31. These sensors were chosen because they presentedhigh responses to the samples for all parameters. Only the data fromthese sensors will be used as the input to the ANN.

5.5. Results from the ANN training and testing

The training of the ANN showed that the e-nose was able todiscriminate healthy and infected tree trunk samples. After the

146 M.A. Markom et al. / Computers and Electron

Fig. 7. PCA result.

Table 2Sensors selection.

Sensors Coefficient

PCA results5 −0.03836

23 −0.1151731 −0.1181810 −0.132536 −0.14794

28 −0.1505112 −0.1560819 −0.157281 −0.15735

22 −0.161262 −0.169158 −0.17595

20 −0.1788824 −0.18034

9 −0.181057 −0.18846

27 −0.189223 −0.18962

29 −0.1898525 −0.1912113 −0.1937115 −0.1940926 −0.1965121 −0.196716 −0.196832 −0.1972414 −0.197794 −0.20052

3

cw

6

nwGpTeof

Microchim. Acta 156, 183–207.

0 −0.2011711 −0.2016117 −0.2024418 −0.20255

ompletion of the network training phase, 100 data of each sampleere tested using the resulting ANN model with 100% accuracy.

. Conclusions

This paper presented a successful application of an electronicose, Cyranose 320, as front end sensors with artificial neural net-orks for the discrimination of oil palm trees which are infected by. boninense in Besout oil palm plantation. This is the first step in

roving the feasibility of using an e-nose for plant disease detection.he next phase of the research will study the ability of the proposed-nose setup to differentiate the different levels of infection stagesf the fungus. This will lead to the implementation of the systemor early infection detection application.

ics in Agriculture 66 (2009) 140–146

The odour profiles recorded are consistent for each tree, and the32 different sensors in Cyranose 320 were able to give individualodour fingerprints of healthy and infected trees. The PCA used asfeature selection was successful to reduce the time execution ofANN training as well as improve the accuracy of classification.

These results are valuable as they proved the feasibility ofusing an electronic nose with artificial intelligence to discriminatehealthy and infected plants, hence the detection of plant diseases.The system can be easily adopted for different disease, and thisapproach holds promise for a better plant disease detection andmonitoring.

As an additional analysis to support the research, the analyticalresult of e-nose may be interpreted on the basis of Gas Chromatog-raphy Mass Spectrometer (GC–MS) data. This can be interesting tostatistically correlate the sensor responses and GC–MS profiles.

Acknowledgements

This work was supported by YAYASAN FELDA. The authors wishto thank all members from Sensors and Application Group Research,Universiti Malaysia Perlis, for their insights and useful discussion.The authors also wish to thank Mr. Hambali B Kamsan, FELDABesout 7 estate manager.

References

Adom, A.H., Shakaff, A.Y.Md., Ahmad, M.N., Abdullah, A.H., Daud, M.Z., Markom, M.A.,2006. Electronic olfactory prototype of an automated plant disease detectionsystem for the agricultural industry. In: Proceedings of the IEEE Systems, Manand Cybernetics Society, 5th Conference on AICS 2006, September 7–8, Sheffield,UK, pp. 9–13.

Balasubramanian, S., Panigrahi, S., Kottapalli, B., Wolf-Hall, C.E., 2007. Evaluation ofan artificial olfactory system for grain quality discrimination. LWT—Food Sci.Technol. 40 (10), 1815–1825.

Berna, A.Z., Lammertyn, J., Saevels, S., Natale, C.Di., Nicolai, B.M., 2003. Electronicnose systems to study shelf life and cultivar effect on tomato aroma profile.Sens. Actuators B: Chem., 324–333.

Boilot, P., Hines, E.L., Gardner, J.W., Pitt, R., John, S., Mitchell, J., Morgan, D.W., 2002.Classification of bacteria responsible for ENT and eye infections using the Cyra-nose system. IEEE Sens. J. 2 (3), 247–253.

Brezman, J., Ma. Fructuosa, L.L., Llobet, E., Vilanova, X., Recasens, I., Orts, J., Saiz, G.,Correig, X., 2005. Evaluation of electronic nose to assess fruit ripeness. IEEE Sens.J. 5 (1), 97–108.

Cyranose 320—IDEO. Retrieved April 15, 2008, from http://www.ideo.com/work/item/cyranose-320.

Dutta, R., Morgan, D., Baker, N., Gardner, J.W., Hines, E.L., 2005. Identification ofStaphylococcus aureus infections in hospital environment: electronic nose basedapproach. Sens. Actuators B, Chem., 355–362.

Emilia Noorsal, Development of electronic nose system using quartz crystalmicrobalance odour sensor array, MSc. Thesis, Universiti Sains Malaysia, 2005.

Falasconi, M., Gobbi, E., Pardo, M., Torre, M.D., Bresciani, A., Sbeveglieri, G., 2005.Detection of toxigenic strains of fusarium verticillioides in corn by electronicolfactory system. Sens. Actuators B 108, 250–257.

Hines, E.L., Llobet, E., Gardner, J.W., 1999. Electronic nose: a review of signal process-ing techniques. IEE Proc. Circuits Devices Syst. 146 (6), 297–310.

Karthikeyan, M., Radhika, K., Bhaskaran, R., Mathiyazhagan, S., Samiyapan, R.,Velazhahan, R., 2006. Rapid detection of Ganoderma disease of coconut andassessment of inhibition effect of various control measures by immunoassayand PCR. Plant Disease Sci. J. 42 (2), 49–57.

Keller, P.E., Kangas, L.J., Liden, L.H., Hashem, S., Kouzes, R.T., 1995. ElectronicNoses and Their Applications, IEEE Northcorn/Technical Applications Confer-ence, October 12.

Miller, R.N.G., Holderness, M., Bridge, P.D., Paterson, R.R.M., Sarah, M., Hussin, M.Z.,1994. Understanding ganoderma populations in oil palm. In: Proceedings ofInternational Workshop on Perennial Crop Diseases caused by Ganoderma,December 1–3, UPM, Serdang, Selangor, Malaysia.

Perera, A., Gomez Baena, A., Sundic, T., Pardo, T., Marco, S., 2002. Machine olfac-tion: pattern recognition for the identification of aromas. IEEE Sens. J., 410–423.

Rafael Felipe castro Rodriguez, A low cost electronic nose for multimedia applica-tions, M.Sc. Thesis, University of Alberta, 2005.

Scott, S.M., James, D., Ali, Z., 2007. Review: data analysis for electronic nose systems.

Shu’ud, M.Md., Loonis, P., Seman, I.A., 2007. Towards automatic recognition andgrading of ganoderma infection pattern using fuzzy systems, transactions onengineering. Comput. Technol. 19 (January), 1305–5313, ISSN.

Troy Nagle, H., Schiffman, S.S., Gutierrez-Osuma, R., 1998. The how and why ofelectronic nose. IEEE Spectrum 35 (9 (September)), 22–34.