www.elsevier.com/locate/rse

Remote Sensing of Environm

Shadow allometry: Estimating tree structural parameters using

hyperspatial image analysis

Jonathan Asher Greenberga,*, Solomon Z. Dobrowskib, Susan L. Ustinb

aNASA Ames Research Center, Moffett Field, CA, United StatesbCalspace, University of California, Davis, United States

Received 19 January 2005; received in revised form 24 February 2005; accepted 26 February 2005

Abstract

We present a novel approach to generating regional scale aboveground biomass estimates for tree species of the Lake Tahoe Basin using

hyperspatial (<1 m2 ground resolution) remote sensing imagery. Tree crown shadows were identified and delineated as individual polygons.

The area of shadowed vegetation for each tree was related to two tree structural parameters, diameter-at-breast height (DBH) and crown area.

We found we could detect DBH and crown area with reasonable accuracy (field measured to image derived cross correlation results were

0.67 and 0.77 for DBH and crown area, respectively). Furthermore, the counts of the delineated polygons in a region generated overstory

stem densities validated to manually photointerpreted stem densities (photointerpreted vs. image-derived stem densities correlation was 0.87).

We demonstrate with accurate classification maps and allometric equations relating DBH or crown area to biomass, that these crown-level

parameters can be used to generate regional scale biomass estimates without the signal saturation common to coarse-scale optical and

RADAR sensors.

D 2005 Elsevier Inc. All rights reserved.

Keywords: Biomass; Allometry; DBH; Crown area; Stem density; Trees; Shadow; Vectorization; Forestry; Hyperspatial imagery; IKONOS; Lake Tahoe Basin

1. Introduction

Minimizing uncertainties in estimating landscape level

carbon sequestration is a critical component to under-

standing many ecosystem processes ranging in scale, from

regional fuel loading for fire risk assessments (Pyne et al.,

1996) to global climate models (Post, 1993). The reason for

much of our uncertainty about biomass levels lays primarily

in the inability to link small-scale, precise carbon estimates

using ground-based sampling and large-scale sampling from

remote sensing technologies. Optical and RADAR sensors

are generally ineffective for detecting the high biomass

levels found in much of the world’s forests (Kasischke et al.,

1997). LIDAR sensors, although proving to be highly

capable at detecting high biomass levels (Hurtt et al., 2004;

0034-4257/$ - see front matter D 2005 Elsevier Inc. All rights reserved.

doi:10.1016/j.rse.2005.02.015

* Corresponding author.

E-mail address: jgreenberg@arc.nasa.gov (J.A. Greenberg).

Riano et al., 2003), suffer from a lack of available orbital

platforms with profiling capabilities, which limits their

usefulness to only highly localized analyses. An alternate

method of detecting biomass could be provided by the new

generation of high resolution, spaceborne satellites such as

IKONOS and Quickbird which have potential to generate

regional-scale biomass estimates, especially through the

fusion of digital image analysis techniques with ground-

based forestry. Our research goal was to determine estimates

of aboveground tree biomass for the Lake Tahoe basin,

located in both California and Nevada. Additionally, we

wish to test whether biomass estimates can be accurately

detected beyond the published upper limits of optical and

RADAR sensors.

Direct measurement of tree biomass can be accom-

plished only by destructive removal and weighing of

entire trees or, minimally, by estimating tree component

volumes (e.g. stem, branches, leaves) and using species-

and component-specific density measurements to extrap-

ent 97 (2005) 15 – 25

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–2516

olate whole-tree biomass. If large-scale field sampling of

tree biomass is required, researchers typically first

generate site and species-specific allometric equations

relating rapidly measurable tree parameters such as

diameter at breast height (DBH) or sapwood area to

destructively sampled biomass. Once these equations are

generated and validated, they are applied to future field-

collected tree parameters.

Although ground-based techniques are accurate for

estimating aboveground biomass, regional scale assess-

ments of biomass can only be approached through remote

sensing technologies, as it is impossible to generate

regional-level estimates of biomass by ground sampling

alone. As such, there has been an emphasis within the

remote sensing community to generate relationships

between spectral signals received by the orbital or airborne

sensor to biomass. These relationships are necessarily

empirical, typically using regression and correlation techni-

ques (Lu et al., 2004).

Hyperspatial sensors, defined as sensors with less than or

equal to 1 m2 ground resolution, open a new possibility for

approaching the problem of estimating aboveground bio-

mass at a landscape scale. Photogrammetrists routinely

estimate crown area (and, subsequently, DBH) on an

individual tree basis through manual photointerpretation

(Bertolette & Spotskey, 1999; Clark et al., 2004). Although

trained analysts can estimate individual tree sizes, to be

usable for landscape-level analyses, this process must be

automated to increase analyst efficiency. Automatic extrac-

tion of crown area, DBH and biomass at the tree level is a

relatively new type of image-analysis capability. All current

techniques rely on delineating a tree crown from back-

ground soil and vegetation, and adjacent trees (Erikson,

2004; Gougeon, 1999; Leckie et al., 2003; Pouliot et al.,

2002). This discrimination produces a GIS layer of vectors

that represent every tree crown resolvable by the particular

vectorization technique. The vectors contain data that can be

used for classification (Brandtberg, 2002; Erikson, 2004;

Leckie et al., 2003) and crown parameter estimation (e.g.

polygon area=crown area) (Maltamo et al., 2003). Although

these techniques are extremely promising, they seem

unlikely to succeed with ‘‘coarser’’ and potentially off-nadir

hyperspatial sensors such as IKONOS (4 m multispectral

which can be pansharpened to 1 m) (Pouliot et al., 2002).

Current research has shown good results with data ranging

between 3 (Erikson, 2004) and 60 cm2 (Gougeon et al.,

1999). However, at 1 m2 resolution, the edges of trees are

difficult to detect even with photointerpretation. We

approach the problem of estimating tree-level biomass by

first estimating tree-level DBH and crown area and, through

field-based allometric relationships, biomass. To do this, we

delineated individual crown shadows, rather than the more

complex task of identifying entire tree crowns. The

shadowed areas are then related to DBH and crown area.

DBH was estimated to fit into existing biomass allometric

equations, but we included crown area here due to its

expected higher accuracy, and better fusion with remote

sensing technologies.

2. Methods

2.1. Site information

The Lake Tahoe Basin area spans the California and

Nevada border between the Carson and Sierra Nevada

mountain ranges. The total land area in the basin is

approximately 82,000 ha, roughly half of which are public

lands, including the Tahoe, Toiyabe, and Eldorado National

Forests. The elevation of the basin ranges between 1900 m

(a.s.l.) at lake level to 3400 m (a.s.l.) at the top of the highest

peaks.

2.2. Image acquisition and preprocessing

Space Imaging’s IKONOS sensor is on a polar orbiting

satellite. The sensor includes four (4) 4 m ground resolution

multispectral bands (blue, green, red, and near-infrared) and

one (1) 1 m panchromatic band. The east–west spatial

extent of a single IKONOS flightline is approximately 12

km. Four (4) IKONOS images were acquired between 19

and 22 July 2002, between 19:03 and 19:11 GMT (12:03–

12:11 pm local). IKONOS has a pointable sensor to provide

more frequent acquisition of local targets, and consequently

our imagery was collected at different sensor zeniths and

azimuth angles. Three images were collected at 15- zenith

and one was collected at nadir. Each image was collected at

a different azimuth: the three off-nadir images were

collected pointing north, south and west. Flightlines were

labeled 1–4 from west to east.

Preprocessing included 6 steps: orthorectification, topo-

graphic shade correction, atmospheric correction, empirical

line calibration to ground collected spectra, image-to-image

radiometric normalization, and pansharpening. Orthorectifi-

cation was provided by Space Imaging. Field analysis with a

Trimble Pro XRS differential GPS showed a relatively high

geographic accuracy in flat terrain (T3–5 m), but more

significant errors in steeper regions. Using ATCOR3

(Richter, 2003) in combination with a Shuttle Radar

Topography Mission (SRTM) 30 m digital elevation model

(DEM), we topographically flattened all 4 and 1 m imagery

using the ray-tracing capability in ATCOR which reduced

shadowing effects from the extremely heterogeneous terrain

present throughout the basin. In addition, ATCOR3 was

used to perform the initial atmospheric correction using the

‘‘us_standard_urban’’ atmosphere profile.

ATCOR3 was not able to precisely estimate the ground

reflectance for all four flightlines, which was confirmed by

examining the spectra of pixels with a primarily Lamber-

tian surface in the overlapping regions between images

(e.g. sand and asphalt). We measured ground spectra using

a GER1500 spectrometer at 21 sites within flightline 3 and

1 2 3 4

740000 750000 760000 770000120˚W

740000 750000 760000 770000

120˚W120˚15'W

4 2900

004 3

00000

4 3100

004 3

20000

4 3300

004 3

40000

4 3500

00

4290000

4300000

4310000

4320000

4330000

4340000

4350000

0 1 2

0 2 4 6 8 10Km

Miles3 4 5 6 7

EW

S

N

Fig. 1. Lake Tahoe Basin with flightlines 1–4 overlaid. UTM Zone 10,

NAD83.

Table 1

TBEVM 3.0 tree species’ names, codes and classification accuracies

Tree species Species code Producer’s (%) User’s (%)

Abies concolor ABCO 69.2 52.4

Abies magnifica ABMA 36.3 29.5

Juniperus occidentalis JUOC 9.6 38.3

Pinus albicaulis PIAL 27.1 54.9

Pinus contorta PICO 32.6 35.6

Pinus jeffreyi PIJE 47.4 48.9

Pinus monticola PIMO 21.0 20.8

Populus tremuloides POTR 44.5 70.4

Tsuga mertensiana TSME 18.2 36.7

Average 34.0 43.1

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–25 17

acquired its geographic position using differential GPS

(ground accuracy was approximately 50–75 cm). These

ground spectra were resampled to the IKONOS band

centers and full width at half maximum (FWHM)

response function using ENVI 3.6 (Research Systems,

2004). Using the empirical line technique (Roberts et al.,

1985), we performed a linear regression between the ground

spectra and the corresponding pixel spectra extracted from

flightline 3.

We performed an image-to-image radiometric normal-

ization using the empirical line technique between flightlines

2 and 4, and flightline 3. This required selecting approx-

imately 50 pixels in the overlap zones between flightlines of

varying brightness. Areas with significant bi-directional

reflectance effects such as forests and shallow water were

avoided in this selection. Using Excel to view the data, any

obvious outliers were removed from the analysis, and the

empirical line was computed using the remaining pixels.

This process was repeated on flightline 1 using the newly

calibrated flightline 2 as the reference image.

The next preprocessing step involved using Principle

Components pansharpening (Welch & Ahlers, 1987) to

sharpen the 4 m multispectral data to 1 m resolution using

the brightness values from the 1 m panchromatic band as a

guide. All four images were then combined into a single

mosaic. Finally, a mask of the Lake Tahoe Basin boundaries

was applied to the mosaic. Fig. 1 shows a color-infrared

(RGB=NIR, Red and Green bands, respectively) of the final

image with the individual image boundaries overlaid.

‘‘Seam lines’’ are still seen within the lake, which are

caused by waves on the water creating specular scattering,

not by poor image-to-image calibration.

2.3. Tahoe Basin Existing Vegetation Map (TBEVM) v 3.0

classification image

We used two products from the pre-existing Tahoe Basin

Existing Vegetation Map v 3.0 (Dobrowski & Greenberg,

2004). The raster classification map was used to determine

the species of vectorized tree crown shadows, used for

calculating the DBH, crown area and stem density estimates.

The mean user’s accuracy for tree species in TBEVM 3.0

was 43.1% (range 20.8–70.4%, see Table 1 for a full listing

of species used in this analysis, their species codes, and a

summary of user’s and producer’s accuracies). In addition,

we estimated landscape level stem density and biomass

based on vector layer generated by segmenting the 4 m

imagery using eCognition 4.0 (Definiens Imaging, 2004).

2.4. Generation of shadowed vegetation raster and GIS

layer

To generate a mask of shadowed vegetation, we created

two separate masks: 1) shadow/not-shadow and 2) vegeta-

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–2518

tion/not-vegetation. We found that basing the shadow purely

on a single-band threshold vastly overestimated the shadow

coverage caused by vegetation, particularly where tree

shadows were projected onto bare soil, or where there was

confusion between shadow and water. To produce the

shadow mask, we randomly selected 100 pixels that were

visually identified as belonging to the shadowed portion of a

tree crown. The mean and standard deviation of the red band

was used to determine the shadowed pixels. We visually

inspected the effects of various shadow/not-shadow thresh-

olds by stepping through mean plus various multiples of the

standard deviation. We found that the best threshold to

capture a tree crown shadow, but not significantly over- or

under-estimate the shadow coverage was the mean red

reflectance+2 standard deviations, or 4.2% reflectance in

the red band.

To generate a mask of vegetated/not-vegetated, we used a

regression of aerial vegetation cover estimates of known

plots vs. the mean normalized difference vegetation index

(NDVI) values for pixels found in these plots. We used

inventory plots acquired from local resource agencies,

including the Natural Resource Conservation Service and

the USDA Forest Service. These plots were collected

following Terrestrial Ecological Unit Inventory (TEUI)

standards (Winthers et al., 2003). Aerial cover is defined

as a ground estimate of the % cover for a given cover class if

viewed from above. Vegetated pixels are defined as having

aerial cover exceeding 1% (Brewer et al., 2004). We found

the relationship to be:

NDVI¼ 0:0033�% aerial cover þ 0:2511; RMSE 0:15:

The threshold, therefore, is classified as vegetation if

NDVI> 0.25. These results were confirmed visually by

Fig. 2. (a) Color infrared image of a forest/wet meadow edge and (b) correspond

stemmed and multi-stemmed shadowed vegetation regions can be seen. (For interp

to the web version of this article.)

inspecting pixels known to be vegetated or non-vegetated

regions.

We combined these two masks to generate a binary mask

in which the pixel value was 1 if the pixel was both shadowed

(red�4.2% reflectance) and vegetated (NDVI>0.25), and 0

otherwise. Fig. 2 shows an example of a shadowed vegetation

mask compared to a CIR image.

We performed a raster-to-vector conversion using ENVI

4.1 on the shadowed vegetation mask. This generated a

polygon coverage of contiguous shadowed vegetation

pixels across the entire landscape. Due to software

limitations, we had to perform this vectorization on spatial

subsets, ranging from 2500 to 10,000 lines. Each

shadowed vegetation polygon was assigned an area

(number of pixels that comprised the contiguous region),

a species (from the TBEVM 3.0 class map, using the most

common pixel class falling in a shadowed vegetation

polygon), a flightline number, and the ID of the TBEVM

polygon in which the particular shadowed vegetation

polygon falls.

2.5. Field and GIS data

We opportunistically collected training samples for each

common tree species found in the basin using differential

GPS, combined with Solo Field software (Tripod Data

Systems, 2004) and the false-color IKONOS imagery as a

background. Common tree species were defined as those

species that comprise prevalent vegetation communities in

the basin as described by regional ecologists. Because of

GPS error and image misregistration, rather than the

traditional method of taking GPS data and using it to re-

register the imagery to fit the GPS coordinates, we used the

GPS in the field to navigate close to a visually identified

ing shadowed vegetation mask (white=shadowed vegetation). Both single

retation of the references to colour in this figure legend, the reader is referred

Table 2

Ninety percent DBH quantile threshold used for single tree/multiple tree

discrimination

Species code 90% DBH quantile (cm)

ABCO 95.9

ABMA 105.7

JUOC 97.4

PICO 73.8

PIJE 104.9

PIMO 124.1

POTR 62.6

TSME 102.3

Table 3

Basin-wide median DBH for trees less than the 90% quantile threshold

Species Median DBH

ABCO 50.9

ABMA 56.5

JUOC 70.2

PICO 46.1

PIJE 58.2

PIMO 69.0

POTR 34.7

TSME 48.2

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–25 19

tree. Then, the tree center was manually selected in the

IKONOS imagery by visually identifying the vegetation.

This technique produced much higher precision in pixel

selection since there was no reliance on high accuracy of the

image orthorectification. Crown width and DBH were

collected for all measured trees. We used a tape measure

at breast height and visual estimation of the edge of the

crown to determine crown radius to the nearest 0.5 m.

Crowns were assumed to be symmetric, so a single radius

was determined in a random direction. Crown radius was

not corrected for the radius of the bole. In cases where trees

appeared as multistemmed individuals (most common in

Juniperus occidentalis and Pinus albicaulis), crown width

was recorded but not DBH. In cases where trees could not

be distinguished as individuals, they were recorded as

clusters of trees that were homogenous in species compo-

sition, patch radius and species, but varied in DBH or crown

width (most common in Populus tremuloides and Abies

concolor). In total, 804 trees (or patches of trees) were

collected.

We created a polygon coverage using the crown radius or

patch radius for each data point to create a circular feature,

multiplying the radius by 75% to guarantee that the true

location of the pixels were within the buffer and solely

within the cover class. Using STARSPAN (Rueda &

Greenberg, 2004), we extracted the 1-m pansharpened

multispectral pixel data with the shadowed vegetation mask

from the region within the buffers. The area of shadowed

vegetation within each crown was determined (number of

pixels of shadow/number of pixels in crown buffer). In

addition, the flightline in which the field data was collected

was determined by performing a spatial join in ArcGIS 9.0

(ESRI, 2004).

2.6. DBH and crown area estimation

We randomly subsetted the field data into two

equivalent sized groups to perform cross-validation

(unscreened n =402 trees per subset). We performed a

natural-log transform on the shadowed vegetation area

(SVA), crown area (CA) and Diameter at Breast Height

(DBH). A multiple regression predicting ln (DBH) and ln

(CA) was performed using ln (SVA) as a continuous

explanatory variable and flightline and species as two

nominal explanatory variables. Outliers from the field data

were identified and removed if: a) the recorded DBH was

greater than the maximum published DBH for that

particular species (indicating an error in data collection),

b) the shadow area was less than 1 m2 (e.g. presence of at

least one shadow pixel), and c) the polygon was not a

patch of trees. Once the outliers and missing data were

removed, the sample size became n =150 and 149 for the

two subsets for DBH. All crown area data collected was

used. One regression was performed for each subset. There

was insufficient data to predict DBH based on shadowed

vegetation area for Pinus albicaulis, so this species was

dropped from further analysis.

Double cross-validation was performed using regression

equations from each of the two subsets to predict DBH and

crown area for the other subset. A Pearson product-moment

correlation coefficient between actual and predicted DBH

and crown area was performed for each subset. High

correlation coefficients (approaching 1.0) would indicate a

good cross-validation.

2.7. Application of regression parameters and multiple

crown correction

We randomly selected one of the two regression

equations generated from the previous step, and applied

them to all shadowed vegetation polygons extracted from

the imagery. Visual inspection revealed that in dense

forests, particularly in off-nadir imagery, crown shadows

often overlapped creating a single large shadow polygon

that was comprised of multiple crowns. To correct for

this error, a histogram of DBH values for each species

was created using the GIS of the entire basin. The 90%

quantile for DBH was used as a threshold to distinguish

single-crown polygons from multiple-crowned polygons.

Table 2 provides the DBH thresholds used by

species.

2.8. Tree density

To validate estimates of tree density, we performed

photointerpretation on 49 randomly chosen 1 ha (100�100

Table 4

DBH (‘‘dbh_pred_cor’’) to aboveground biomass allometric equations

Cover class AGB equation

ABCO1 Exp(4.36982+(2.5043*Log(dbh_pred_cor)))/1000

ABMA1 Exp(2.61856+2.9121*Log(dbh_pred_cor))/1000

JUOC3 Exp((�8.3939)+2.6344*Log(dbh_pred_cor*Pi()))+

Exp((�7.3115)+2.3337*Log(dbh_pred_cor*Pi()))+

Exp((�11.846)+2.8323*Log(dbh_pred_cor*Pi()))+

Exp((�4.243)+1.5606*Log(dbh_pred_cor*Pi()))

PICO1 (�0.463)+0.106*dbh_pred_cor

PIJE2 Exp((�4.2847)+2.718*Log(dbh_pred_cor))+

Exp((�4.2062)+2.2475*Log(dbh_pred_cor))+

Exp((�5.29)+2.6524*Log(dbh_pred_cor))+

Exp((�3.7969)+1.7426*Log(dbh_pred_cor))+

Exp((�3.9739)+2.0039*Log(dbh_pred_cor))

PIMO2 Exp((�4.2847)+2.718*Log(dbh_pred_cor))+

Exp((�4.2062)+2.2475*Log(dbh_pred_cor))+

Exp((�5.29)+2.6524*Log(dbh_pred_cor))+

Exp((�3.7969)+1.7426*Log(dbh_pred_cor))+

Exp((�3.9739)+2.0039*Log(dbh_pred_cor))

POTR1 Exp((�2.3778)+2.4085*Log(dbh_pred_cor))

TSME1 Exp(3.69564+2.37489*Log(dbh_pred_cor))/1000

DBHs are in centimeters and AGB are in kilograms.1From Jenkins et al. (2004); 2P. jeffreyi and P. monticola eqs. were

unavailable; we used generic Pinus spp. from Jenkins et al. (2004); 3 From

Gholz (1980).

Table 5

Predicted parameter values for the random subsets i and ii for a) species and

b) flightline for the DBH to shadowed vegetation area relationship

a.

Species SPi SPii

ABCO �0.323 �0.203

ABMA 0.093 0.127

JUOC 0.367 0.352

PICO �0.205 �0.207

PIJE �0.087 0.027

PIMO 0.285 0.328

POTR �0.398 �0.360

TSME 0.269 �0.065

b.

Flightline FLi FLii

1 0.043 0.012

2 0.047 0.098

3 0.011 0.091

4 �0.101 �0.201

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–2520

pixels) image subsets stratified by flightline, but weighted

according to the area each flightline contributed to the final

mosaic (flightlines 1, 2, 3 and 4 had 2, 14, 15 and 19

random subsets chosen, respectively). A single researcher

performed the photointerpretation, and used both a false-

color infrared and NDVI image to place points on all

identifiable crowns. A polygon layer of the subset bounda-

ries was used to select the shadowed vegetation polygons

falling within the subsets, and then to clip these polygons to

the subset boundaries.

The image-derived stem density was calculated by

adding the number of single crowned polygons to

the corrected multiple crowned polygons. The number

of stems in a multiple crowned polygon was calcu-

lated as the predicted DBH for the polygon divided

by the median DBH for all trees of that species

falling in the 90% quantile range (see Table 3). A

correlation was performed to determine the relation-

ship between photointerpreted and image-derived stem

densities.

2.9. Biomass estimation

We applied published species-specific allometric equa-

tions (Table 4) relating DBH to aboveground biomass of

all tree species in the Lake Tahoe Basin to generate a map

of biomass per hectare (kg/ha) for TBEVM 3.0 polygons.

In the case of multiple-crowned polygons, we used the

predicted biomass given the median DBH per species,

multiplied by the predicted number of stems contained by

the multiple-crowned polygon.

3. Results

3.1. DBH estimation

The regression equations used for the two random

subsets i and ii (n =141 and 139, respectively) to predict

DBH from shadow area, species and flightline was found

to be:

ln DBHið Þ ¼ 4:025þ SPi þ FLi þ 0:145Tln SVAið Þ; andð1Þ

ln DBHiið Þ ¼ 3:808þSPii þ FLii þ 0:207Tln SVAiið Þ: ð2Þ

Species (SPi and SPii) and flightline (FLi and FLii)

parameters are summarized in Table 5a and b, respectively.

Root mean square error (RMSE) for subset i and ii was

and 0.315 and 0.263, respectively. The double cross-

validation results yielded a Pearson product-moment

correlation coefficients between the field-measured DBH

and the DBH predicted from the regression coefficients as

0.69 and 0.66 (mean 0.67; see Fig. 3a and b for the

scatterplots for subsets i and ii). Based on these equations

(using shadow area=1 m2, or 1 pixel), flightline 4 had the

smallest resolvable DBH. The mean minimum DBH for all

species on flightline 4 was 37.9 cm, as contrasted with a

minimum 51.2 cm for flightline 2. Overall mean minimum

resolvable DBH for all species on all flightlines was 46.7

cm. Species appeared to have a relatively small impact on

overall DBH estimation: the coefficients for both random

subsets ranged from �0.398 (0.7 cm DBH) to 0.367 (1.4

cm DBH) or only a 0.7 cm range of values between all

DB

H (

cm),

imag

e de

rived

DB

H (

cm),

imag

e de

rived

DBH (cm), field measured DBH (cm), field measured

a. b.

Fig. 3. Cross-validation results for field measured DBH vs. image derived DBH for subsets i (a) and ii (b). Dashed line is the 1 :1 line, solid line is the least-

squares line fit to the actual results.

Table 7

Predicted parameter values for the random subsets i and ii for a) species and

b) flightline for the crown area to shadowed vegetation area relationship

a.

Species SPi SPii

ABCO 0.342 0.254

ABMA 0.102 0.301

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–25 21

species. Table 6 shows a summary of minimum resolvable

DBH by species and flightline.

3.2. Crown area estimation

The regression equations for crown area for the two

random subsets i and ii (n =378 and 374) were:

ln CAið Þ ¼ 1:960þ SPi þ FLi þ 0:630Tln SVAið Þ; and ð3Þ

ln CAiið Þ ¼ 2:006þ SPii þ FLii þ 0:628Tln SVAiið Þ; ð4Þ

where the parameters SP and FL are described in Table

7a and b, respectively. The RMSE for the two subsets i

and ii was 0.641 and 0.627 respectively. The double

cross-correlation yielded coefficients of 0.71 and 0.83

(mean 0.77; see Fig. 4a and b for the corresponding

scatterplots).

Table 6

Minimum resolvable DBH (cm) by species and flightline

Species Flightline 1 Flightline 2 Flightline 3 Flightline 4

ABCO 37.2 40.6 40.3 30.1

ABMA 51.8 56.5 56.0 41.8

JUOC 64.8 70.7 70.2 52.4

PICO 37.1 40.4 40.1 30.0

PIJE 46.8 51.1 50.7 37.8

PIMO 63.3 69.0 68.5 51.1

POTR 31.8 34.7 34.4 25.7

TSME 42.7 46.6 46.2 34.5

3.3. Tree density

The Pearson product-moment correlation between photo-

interpreted stem density and the image-derived stem density

counting was 0.87. Fig. 5 shows the scatterplot of photo-

interpreted vs. image-derived stem density results. Fig. 6

shows the basin wide distribution of tree densities. Stem

densities for vegetated polygons ranged from 0.1 to 625

stems/ha (99.5% interval 0.4–132.4 stems/ha, mean 47.1

stems/ha). For tree dominated communities (tree cov-

er�25%, (Tart et al., 2004)), the 99.5% stem density range

was 5.0–146.5 stems/ha, mean 72.8 stems/ha.

JUOC 0.127 0.218

PIAL 0.202 �0.129

PICO �0.303 �0.194

PIJE 0.087 �0.070

PIMO 0.177 �0.179

POTR �0.297 �0.109

TSME �0.437 �0.092

b.

Flightline FLi FLii

1 0.786 0.813

2 �0.754 �0.731

3 0.147 �0.018

4 �0.179 �0.065

Fig. 4. Cross-validation results for field measured crown area vs. image derived crown area for subsets i (a) and ii (b). Dashed line is the 1 :1 line, solid line is

the least-squares line fit to the actual results.

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–2522

3.4. Aboveground tree biomass estimation

Fig. 7 displays the basin-wide estimation of biomass (kg/

ha). Biomass for vegetated polygons ranged from 0.0 to

748.0 Mg/ha (99.5% interval 0.0–308.1 Mg/ha, mean 77.5

Mg/ha). For tree dominated communities, the 99.5% range

for biomass/ha range was 2.1–334.9 Mg/ha, and a mean of

120.7 Mg/ha.

Fig. 5. Photointerpreted vs. image derived stem density estimations.

Dashed line is the 1 :1 line, solid line is the least-squares line fit to the

actual results. Fig. 6. Stem density for TBEVM 3.0 management polygons.

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–25 23

4. Conclusions

Identification of vegetation shadows and vectorization of

these small pixel clusters of shadows is a promising technique

for estimating DBH, crown area, and stem density over

landscape scales. Through this process, a wealth of related

biophysical parameters can be accessed by using allometric

equations to relate crown parameters to biomass. These

techniques leverage both the strengths of hyperspatial

imagery and comparatively simple vectorization techniques

(compared to Erikson, 2004; Gougeon, 1999; Leckie et al.,

2003; Pekkarinen, 2002; Pouliot et al., 2002). DBH, crown

area and stem density showed good correlations between

image-derived estimates and field/photointerpreted results,

although to differing degrees of success.

For DBH, crown area and stem density estimation, a

major limiting factor was the vectorization process. We

implemented the single most basic raster-to-vector conver-

sion currently available: treat all connected pixels as

Fig. 7. Estimated aboveground tree biomass per ha (Mg/ha) for TBEVM 3.0

management polygons.

belonging to a single polygon. A more complex vectoriza-

tion could have improved the estimations by potentially

eliminating the need to set somewhat arbitrary thresholds to

distinguish single crowned polygons from multiple crowned

polygons. This became a critical issue in dense forests,

where entire stands of trees might be included in a single

shadow polygon. We plan on investigating other forms of

vectorization for future analyses.

Although promising, DBH estimation was the least

effective relationship generated due, in addition to the

vectorization issues, to a number of complications. The

DBH estimation showed a significant effect of flightline

number, probably a result of differing sensor angles, which

could have been avoided early in the study by acquiring

only nadir imagery. The imagery request was designed to be

acquired over a short time frame to avoid significant

phenological changes that could have led to significant

changes in detected spectral signatures. However, although

phenological changes probably would have resulted in

significant spectral changes in the herbaceous and, poten-

tially, shrub communities, they are unlike to have had a

serious affect on tree signals (with the exception of the

single broadleaf Populus tremuloides). Acquisition of

entirely nadir imagery, even if it took a much longer span

of time to collect, would have vastly simplified the analysis

and would probably have improved our ability to estimate

DBH. In addition to the effects of flightlines, species were

also found to have a significant, but limited, effect on

estimated DBH. The lack of a more pronounced effect of

species was probably the relatively similar architecture of

many of the species used, suggesting this technique may be

useful even in the absence of more accurate species maps.

With that said, the more unique architectures, particularly

Juniperus occidentalis, did show significantly different

coefficients than the other conifer species so we cannot

discount the importance of, minimally, an architecture-

specific classification maps and allometric relationships.

Crown area was much more accurate than DBH, as

shown by the larger correlation coefficients. Like DBH,

complicating factors also reduced the potential accuracy,

which also related to the different sensor angles and

uncertainty in our species assignments. Crown area, unlike

DBH, is directly viewable by the sensor (particularly for

nadir sensors), so the shadow to crown area relationship is

largely geometric, and not complicated by more complex

physiological relationships.

Stem counting by isolating shadow clusters proved to be

an excellent way to assess stem density. However, the stem

counts were based on photointerpretation, and not from

ground-based measurements. Therefore, some uncertainty

exists for the accuracy of the photointerpreted based map. A

qualitative comparison to ground-based records of old-

growth sites in the Lake Tahoe Basin (Barbour et al., 2002)

showed that our stem densities are in the right range in the

context of this particular watershed, but only for large,

overstory trees (>40 cm DBH). Understory tree densities

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–2524

were unrepresented in our analysis since, like all optical and

RADAR sensors, there is little reflected spectral information

from the subcanopy that can be easily decoupled from the

dominant canopy signature. Full waveform LIDAR is

currently the only type of sensor capable of detecting

subcanopy biomass levels, particularly at this spatial scale

(Hurtt et al., 2004). In addition, due to limitations of the

spatial resolution of the IKONOS sensor, the two smallest

Federal Geographic Data Committee (FGDC) subclasses

(0–5� and 5–10� DBH, Brewer et al., 2004) fall below the

minimum resolvable DBH. Only higher spatial resolution

sensors or more empirical approaches to analysis are likely

to improve these estimates.

Biomass estimates could not be directly validated at this

time, but several observations can be made. Like stem

density, at best this approach will provide accurate

estimations of canopy biomass, but not understory biomass,

which will lead to somewhat underestimated aboveground

biomass. However, what is promising from this approach is

that detection of tree size (be it DBH or crown area) does

not have any inherent upper limits, since the only thing that

restrict our estimates are sensor resolution and crown

overlap. Individual trees are known to have biomass levels

far exceeding any optical or RADAR sensor: the world’s

largest white fir, for instance, has an estimated biomass/ha

of 5421.5 Mg/ha (based on a 276� circumference and 39Vcrown spread), compared to the upper biomass limits

demonstrated by RADAR technology which is about 360

Mg/ha, Kasischke et al., 1997).

To fully exploit these techniques, further advances need to

be made in the fields and technology of forestry, image

classification and computer vision. All future field derived

allometric biomass equations should include crown area or

radius in addition to DBH. There is an overemphasis on

relating DBH to biomass, which leads to greater error when

trying to predict tree biomass using remote sensing. Crown

radius can be rapidly acquired in conjunction with DBH, and

the benefits will far outweigh the small additional time it

takes to collect this data as part of an effort to derive biomass

equations and provide a direct link to hyperspatial remote

sensing data. In terms of classification, species mapping on

the scale of hyperspatial data remains error-prone. Advances

in both the techniques of producing classification maps from

existing hyperspatial sensors, and a push towards using

newer technologies such as hyperspectral sensors will likely

improve this input. Finally, individual tree (or relevant tree

component) extraction is a new but growing field of

applications. Pixel-level analyses are no longer relevant at

this scale, since the objects in question are larger than a pixel.

Advances in pattern recognition, image segmentation, and

computer vision will lead to better extraction of tree crowns,

leading to improved estimations of crown area, and poten-

tially to more precise spectral and textural information which

can be used to perform classification. With further advances

along these lines, these techniques should be widely

applicable across a wide range of tree-dominated ecosystems,

and could potentially lead to major improvements in our

understanding of climate change processes at a microscale

across a landscape.

Acknowledgements

This project was funded by the Tahoe Regional Planning

Agency (TRPA) and the Forest Service (USFS). Special

thanks to Mike Vollmer (TRPA), Shane Romsos (USFS),

Hugh Safford (USFS), George Scheer, Lawrence Ross,

Shawn Kefauver, Jaylee Tuil, Missy Voight, Ben Addle-

stone, Jennifer Riddell, Dylan Burge, Shannon Murphy and

Jessica Gorin.

References

Barbour, M., Kelley, E., Maloney, P., Rizzo, D., Royce, E., & Fites-

Kaufmann, J. (2002). Present and past old-growth forests of the Lake

Tahoe Basin, Sierra Nevada, US. Journal of Vegetation Science, 13(4),

461–472.

Bertolette, D. R., & Spotskey, D. B. (1999). Fuel model and forest type

mapping for FARSITE input. In G. E. Gollberg (Ed.), The joint fire

science conference and workshop. Boise, ID’ University of Idaho and

International Association of Wildland Fire.

Brandtberg, T. (2002). Individual tree-based species classification in high

spatial resolution aerial images of forests using fuzzy sets. Fuzzy Sets

and Systems, 132, 371–387.

Brewer, C. K., Schwind, B., Warbington, R. J., Clerke, W., Krosse, P. C.,

Suring, L. H., et al. (2004). Appendix 1B: Overview of the NVCS

physiognomic hierarchy. In R. J. Brohman, & L. D. Bryant (Eds.),

Existing vegetation classification and mapping technical guide.

Washington’ USDA Forest Service.

Clark, D. B., Read, J. M., Clark, M. L., Cruz, A. M., Dotti, M. F., & Clark,

D. A. (2004). Application of 1-m and 4-m resolution satellite data to

ecological studies of tropical rain forests. Ecological Applications,

14(1), 61–74.

Definiens Imaging. (2004). eCognition. Munchen’ Professional.

Dobrowski, S., & Greenberg, J. A. (2004). Tahoe basin existing vegetation

map v 3.0. Tahoe Regional Planning Agency, Stateline.

Erikson, M. (2004). Species classification of individually segmented tree

crowns in high-resolution aerial images using radiometric and

morphologic image measures. Remote Sensing of Environment,

91(3–4), 469–477.

ESRI. (2004). ArcGIS. Redlands’ ESRI.

Gholz, H. L. (1980). Structure and productivity of Juniperus occidentalis in

central Oregon USA. American Midland Naturalist, 103(2), 251–261.

Gougeon, F. A. (1999). Automatic individual tree crown delineation using a

valley-following algorithm and a rule-based system. In D. A. Hill, & D.

G. Leckie (Eds.), Proceedings of the international forum on automated

interpretation of high spatial resolution digital imagery for forestry,

February 10–12, 1998, Victoria, B.C., Canada (pp. 11–23).

Gougeon, F. A., Leckie, D. G., Paradine, D., & Scott, I. (1999). Individual

tree crown species recognition: The Nahmint study. Automated

interpretation of high spatial resolution digital imagery for forestry,

International Forum, PG, 1999.

Hurtt, G. C., Dubayah, R., Drake, J., Moorcroft, P. R., Pacala, S. W., Blair,

J. B., et al. (2004). Beyond potential vegetation: Combining lidar data

and a height-structured model for carbon studies. Ecological Applica-

tions, 14(3), 873–883.

Jenkins, J. C., Chojnacky, D. C., Heath, L. S., & Birdsey, R. A. (2004).

Comprehensive database of diameter-based biomass regressions for

J.A. Greenberg et al. / Remote Sensing of Environment 97 (2005) 15–25 25

North American tree species. U.S. Department of Agriculture. Newtown

Square’ Forest Service, Northeastern Research Station.

Kasischke, E. S., Melack, J. M., & Dobson, M. C. (1997). The use of

imaging radars for ecological applications—a review. Remote Sensing

of Environment, 59(2), 141–156.

Leckie, D. G., Gougeon, F. A., Walsworth, N., & Paradine, D. (2003). Stand

delineation and composition estimation using semi-automated individual

tree crown analysis. Remote Sensing of Environment, 85(3), 355–369.

Lu, D., Mausel, P., Brondizio, E., & Moran, E. (2004). Relationships

between forest stand parameters and Landsat TM spectral responses

in the Brazilian Amazon Basin. Forest Ecology and Management,

198(1–3), 149–167.

Maltamo, M., Tokola, T., & Lehikoinen, M. (2003). Estimating stand

characteristics by combining single tree pattern recognition of digital

video imagery and a theoretical diameter distribution model. Forest

Science, 49(1), 98–109.

Pekkarinen, A. (2002). Image segment-based spectral features in the

estimation of timber volume. Remote Sensing of Environment,

82(2–3), 349–359.

Post, W. M. (1993). Uncertainties in the terrestrial carbon cycle. In A. M.

Solomon, & H. H. Shugart (Eds.), Vegetation dynamics and global

change (pp. 116–132). London’ Chapman and Hall.

Pouliot, D. A., King, D. J., Bell, F. W., & Pitt, D. G. (2002). Automated tree

crown detection and delineation in high-resolution digital camera

imagery of coniferous forest regeneration. Remote Sensing of Environ-

ment, 82(2–3), 322–334.

Pyne, S. J., Andrews, P. L., & Laven, R. D. (1996). Introduction to wildland

fire. New York’ Wiley and Sons.

Research Systems, Inc. (2004). ENVI: The environment for visualizing

images. Boulder’ Research Systems, Inc.

Riano, D., Meier, E., Allgower, B., Chuvieco, E., & Ustin, S. L. (2003).

Modeling airborne laser scanning data for the spatial generation of

critical forest parameters in fire behavior modeling. Remote Sensing of

Environment, 86(2), 177–186.

Richter, R. (2003). Atmospheric/topographic correction for satellite

imagery: ATCOR-2/3 user guide, version 5.5, January 2003. Wessling’

ReSe Applications Schlapfer.

Roberts, D. A., Yamaguchi, Y., & Lyon, R. J. P. (1985). Calibration of

airborne imaging spectrometer data to percent reflectance using field

spectral measurements. 19th international symposium on remote

sensing of environment, Ann Arbor, MI.

Rueda, C., & Greenberg, J. A. (2004). STARSpan. Davis’ University of

California, Davis.

Tart, D., Williams, C. K., Brewer, C. K., DiBenedetto, J. P., & Schwind, B.

(2004). Section 1: Existing vegetation protocol framework. In R. J.

Brohman, & L. D. Bryant (Eds.), Existing vegetation classification and

mapping technical guide. Washington’ USDA Forest Service.

Tripod Data Systems, Inc. (2004). SOLO field. Corvallis’ Tripod Data

Systems, Inc.

Welch, R., & Ahlers, W. (1987). Merging multiresolution SPOT HRV and

landsat TM data. Photogrammetric Engineering and Remote Sensing,

53, 301–303.

Winthers, E., Fallon, D., Haglund, J., Demeo, T., Tart, D., Ferwerda, M.,

et al. (2003). Terrestrial ecological unit inventory technical guide.

USDA Forest Service, Washington Office-Ecosystem Management

Coordination Staff, 125 pp.