Content uploaded by Dar A. Roberts
Author content
All content in this area was uploaded by Dar A. Roberts on May 17, 2016
Content may be subject to copyright.
This article was downloaded by: [University of California Santa Barbara]
On: 30 January 2013, At: 12:39
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
Remote Sensing Letters
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/trsl20
Identifying Santa Barbara's urban tree
species from AVIRIS imagery using
canonical discriminant analysis
Mike Alonzo a , Keely Roth a & Dar Roberts a
a Department of Geography, University of California, Santa,
Barbara, CA, USA
Version of record first published: 30 Jan 2013.
To cite this article: Mike Alonzo , Keely Roth & Dar Roberts (2013): Identifying Santa Barbara's
urban tree species from AVIRIS imagery using canonical discriminant analysis, Remote Sensing
Letters, 4:5, 513-521
To link to this article: http://dx.doi.org/10.1080/2150704X.2013.764027
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-
conditions
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation
that the contents will be complete or accurate or up to date. The accuracy of any
instructions, formulae, and drug doses should be independently verified with primary
sources. The publisher shall not be liable for any loss, actions, claims, proceedings,
demand, or costs or damages whatsoever or howsoever caused arising directly or
indirectly in connection with or arising out of the use of this material.
Remote Sensing Letters
Vol. 4, No. 5, May 2013, 513–521
Identifying Santa Barbara’s urban tree species from AVIRIS imagery
using canonical discriminant analysis
MIKE ALONZO∗, KEELY ROTH and DAR ROBERTS
Department of Geography, University of California, Santa, Barbara, CA, USA
(Received 29 October 2012; in final form 2 January 2013)
In this research, we classify 15 common urban trees in downtown Santa Barbara,
California, using crown-level canonical discriminant analysis (CDA) on airborne
visible/infrared imaging spectrometer (AVIRIS) imagery. We compare the CDA
classification accuracy against results obtained from stepwise discriminant analysis.
We also examine the impact of various crown-level aggregation techniques and
training sample size on classification results. An overall classification accuracy
of 86% was achieved using CDA. Species-specific results were highest for dense
crowns with high normalized difference vegetation index values. Bands chosen
using forward feature selection spanned AVIRIS full spectral range illustrating a
need for retaining a full complement of spectral information. Nevertheless, there is
some indication that bands along the green edge, green peak and yellow edge are
particularly valuable for discriminating structurally similar urban trees.
1. Introduction
Urban forest management and parameterization of ecosystem models are depen-
dent on identification of urban trees to the species level (Santamour 1990, La´
can
and McBride 2008, Nowak et al. 2008, White and Zipperer 2010). Presently, most
tree identification is done on the ground and is conducted only for either a sample
or a jurisdictional subset of a city’s trees. Plot-based sample inventories require a
significant investment in trained labour and, while useful, do not produce spatially
explicit results. Remote sensing platforms can provide wall-to-wall maps charac-
terizing vegetation throughout a city. However, the identification of tree species in
biodiverse environments remains a challenge due to the high degree of spectral
variability within a given species or even a single crown.
Hyperspectral imagery has improved our ability to resolve subtle spectral differ-
ences among species. Xiao et al. (2004) used spectral mixture analysis applied to
airborne visible/infrared imaging spectrometer (AVIRIS; Green et al. 1998) data to
map 22 urban tree species in Modesto, CA, with 70% accuracy. Zhang and Qiu (2012)
identified 40 species in Dallas, Texas, with 69% accuracy using a neural network
approach. Crown-level, as opposed to pixel-level, classification may further enhance
our ability to make useful maps, with less ‘salt and pepper’ noise (Benz et al. 2004,
Myint et al. 2011). For instance, Clark et al. (2005) reported 85% pixel-level accuracy
in discriminating seven tropical rainforest tree species and 90% accuracy using mean
*Corresponding author. Email: mike.alonzo@geog.ucsb.edu
Remote Sensing Letters
ISSN 2150-704X print/ISSN 2150-7058 online © 2013 Taylor & Francis
http://www.tandfonline.com
http://dx.doi.org/10.1080/2150704X.2013.764027
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
514 M. Alonzo et al.
spectra from manually-delineated crowns. Automatic crown delineation in heteroge-
neous forests remains a challenge due to variable crown structure and inconsistent
relationships between tree height and crown width (Jing et al. 2012). That said, Zhang
and Qiu (2012) reached 94% crown detectionaccuracy using ‘tree climbing’ and ‘donut
expanding’ methods to segment a lidar point cloud.
Canonical discriminant analysis (CDA; Klecka 1980, Zhao and Maclean 2000) is a
data reduction technique well suited for separating potentially overlapping classes in
high dimensional feature space. It is similar to principal components analysis (PCA)
and minimum noise fraction (MNF), but rather than summarizing total image band
variation, CDA summarizes between class variance among ggroups or classes (Klecka
1980, Zhao and Maclean 2000). As such, CDA requires aprioriknowledge of the
training data set’s class variable (in this case, tree species). The derived canonical
discriminant functions are linear combinations of the original variables where the
coefficients maximize the between-group separation. Conceptually, CDA offers several
advantages for classification of hyperspectral imagery. First, reducing the data set
to g-1 canonical variables instead of poriginal variables avoids the ill-posed prob-
lem where pis greater than the number of observations, n. Second, in selecting only
those canonical variables with significant discriminating power, one effectively par-
titions target information from unexplained noise (Zhao and Maclean 2000). Third,
CDA has its basis in common multivariate statistical methods. Therefore, unlike pop-
ular classifiers such as support vector machines, it offers methodological transparency
and interpretable by-products (e.g. standardized coefficients) that allow one to further
characterize the separability of their classes. Pu and Liu (2011) used in-situ, leaf-level
spectra from 13 urban species in Tampa, Florida, to reach 90% classification accuracy
using segmented CDA and 89% accuracy using stepwise discriminant analysis (SDA)
versus 80% accuracy for PCA with a linear discriminant analysis (LDA) classifier.
In this research, we take a step towards spatially explicit urban forest inventory by
improving our ability to discriminate among 15 urban tree species using AVIRIS data
acquired over Santa Barbara, California. Specifically, we used CDA to classify multi-
ple pixels in each tree crown then labelled the crown with the mode pixel classification
result (i.e. ‘winner-take-all’). This method was compared to classification using full
spectrum LDA and SDA on a subset of bands to better understand the impact of
dimension reduction on crown-level classification accuracies. Training set sizes were
also varied to examine the relative degradations in accuracy by method. Finally, we
hypothesized that bands representing the full spectral range, from the visible through
the shortwave infrared, would be necessary for accurate urban forest species identifi-
cation. To test this hypothesis, we analysed the distribution of forward feature selected
bands contributing most to reducing the overall misclassification rate (MCR).
2. Materials and methods
2.1 Study area and sampling
This study was conducted in downtown Santa Barbara, California (34.42◦N,
119.69◦W) (figure 1). Santa Barbara is situated on a coastal plain between the Pacific
ocean to the south and the Santa Ynez range to the north. It benefits from year-round
mild temperatures associated with its Mediterranean climate. Given the amenable
climate, the urban forest supports a wide variety of native, introduced, and invasive
species. The City of Santa Barbara (2010) includes over 450 species in their database of
publically-managed trees. In this study, we chose to focus on 15 of the most common
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
Urban forest species classification from AVIRIS imagery 515
Figure 1. Study area: downtown Santa Barbara.
species in this database (table 1). Palm species were excluded because their average
crown size was smaller than one image pixel (3.7 m). It is our intent that these species
will be classified in a follow-up study incorporating lidar data, which may increase our
ability to map at a fine scale.
The sample included 722 crowns with each species represented by at least 41 crowns.
The crowns were delineated in ArcGIS (ArcGIS 9.3, Esri, Redlands, CA, USA) using
an August 2010 lidar canopy height model at 1 m grid size and a geospatial database
Table 1. Producer’s and user’s accuracies by species.
Species name Tree type Producer’s accuracy (%) User’s accuracy (%)
Eucalyptus ficifolia BP 94 73
Eucalyptus globulus BP 94 97
Ficus microcarpa BP 97 87
Jacaranda mimosifolia BP 97 96
Liquidambar styraciflua BD 73 80
Lophostemon confertus BP 80 46
Magnolia grandiflora BP 95 93
Olea europaea BP 92 99
Pittosporum undulatum BP 89 99
Platanus racemosa BD 82 76
Podocarpus gracilior BP 90 87
Pyrus kawakamii BP/D71 71
Quercus agrifolia BP 79 70
Schinus terebinthifolius BP 82 94
Ulmus parvifolia BD 63 77
Notes: BP, broadleaf persistent; BD, broadleaf deciduous.
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
516 M. Alonzo et al.
(a)(b)
N75 m
Figure 2. (a) Examples of crown objects overlaid on a 1 m lidar canopy height model. (b)The
same crown objects overlaid on a 3.7 m colour infrared (R, 724 nm; G, 648 nm; B, 550 nm)
display of AVIRIS imagery.
from the City of Santa Barbara with species information (figure 2). Species were visu-
ally confirmed using Google Street View imagery from 2011 under leaf-on conditions.
Individual tree crowns were selected only if the length of the minor axis was greater
than 7 m, meaning each crown could contain at least four AVIRIS pixels. The majority
of crowns are street trees, but some are on school or park property. The sample was
randomly split into training and validation subsets. The size of the training subset was
varied to quantify the effect of training sample size on classification results.
2.2 Image processing and spectra extraction
AVIRIS is a 224 channel system that measures radiance from 365 to 2500 nm with
a34
◦field of view and an instantaneous field of view of 1 mrad (Green et al. 1998).
Two AVIRIS scenes over the study area were acquired at approximately 1150 and
1420 Pacific Standard Time in November 2010, with solar zenith angles of 50.5◦and
54.1◦, respectively. The flight-lines were flown at average altitudes of 4421 and 3963 m
with a Twin Otter aircraft resulting in pixel sizes of 3.7 and 3.4 m, respectively. Surface
reflectance was retrieved on each flight-line using ATCOR-4 (ReSe Applications, Wil,
Switzerland) (Richter and Schläpfer 2002) and 178 bands were retained for analysis.
The reflectance image was registered to the lidar data set using Delaunay triangulation.
To ensure all relevant pixels were extracted for each crown object, we resampled the
AVIRIS data to 1 m with nearest neighbour resampling to match the resolution of
the delineation. Redundant spectra created in this resampling were discarded prior to
analysis.
Spectra were extracted from each full crown and sunlit portion based on a vari-
able normalized difference vegetation index (NDVI; Rouse et al. 1973) threshold.
An NDVI threshold was used to limit the pixels included in the analysis to only the
purest canopy spectra while retaining multiple pixels per crown whenever possible.
The average NDVI for all crowns was 0.61. Therefore, the initial extraction threshold
was set to 0.6. If no pixels in a given crown met that criterion, all pixels above 0.5 were
extracted. If no pixels met this second threshold, the single pixel with the maximum
NDVI value was selected. On average, eight unique pixels were extracted from each
full crown for a total of 5755 spectra. Sixty-two crowns (9%) were represented by only
one spectrum.
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
Urban forest species classification from AVIRIS imagery 517
2.3 Data reduction and classification
In the present study, we highlight the capacity of CDA to classify crowns using a
reduced data set. We also offer a comparison to results obtained using stepwise band
selection as well as to results from LDA on all 178 bands. For CDA, 14 canonical
variates were generated from canonical coefficients derived from training spectra. This
is the maximum number of canonical variates allowed given 15 groups; each exhibited
statistically significant discriminating power (α=0.01). For SDA, we ran forward
feature selection on randomly selected training crowns 100 times to choose a subset of
bands based on an MCR minimization criterion.
LDA classification was conducted at the pixel level on the CDA variates, the full
SDA subsets, the first 14 bands of each SDA subset (for direct comparison to the
CDA variates) and all 178 bands. LDA is also based on the creation of linear com-
binations of the discriminating variables that maximize the between-group differences
(Fisher 1936, Yu et al. 1999). With LDA, an observation is assigned to the group with
the highest classification function score. Full crowns and sunlit crowns received their
classification based on the mode pixel-classification result (‘winner-take-all’). We com-
pared the winner-take-all approach with classification of mean crown spectra, another
crown-level classification technique. One mean spectrum was created for each crown
by averaging each pixel that met the NDVI threshold. To test the effects of training
sample size at the crown level, the classification was run with randomly selected train-
ing sets of 40, 30, 20, 15, 10 and 5 crowns per species. Accuracies were assessed in each
case using the set of crowns (and constituent pixels) held out from algorithm training.
3. Results
3.1 Classification accuracies
Over 100 sampling iterations, an average crown-level classification accuracy of 86%
(kappa value =0.85) was achieved using CDA for both full crowns and sunlit por-
tions only. LDA on all bands produced the same accuracy. SDA selected, on average,
45 bands (SDA-45) and yielded an average accuracy of 84%. The mean accuracies for
runs using the largest training sample size (n=40) were tested for significant difference
using two sample t-tests. The SDA result was significantly lower than the LDA or CDA
results (α=0.01). Using only the first 14 forward selected bands from each SDA sub-
set reduced the average accuracy to 43%. Classification accuracies decreased at similar
rates with reduced training set size for CDA, SDA-45 and LDA (figure 3). Accuracies
at the pixel level were generally about 10% lower than crown-level accuracies. This
relationship remained stable at all sample sizes and across all techniques. Classification
accuracy of crown mean spectra was much more sensitive to reductions in the size of
the training data set. While a training set of 40 crowns produced an accuracy of 83%,
a training set of 10 resulted in 24% correct classification. This sensitivity to small sam-
ple sizes is likely due to the increased influence of outlier spectra included within a
crown due to registration error or legitimate morphological characteristics (e.g. sparse
crown, pruning). The winner-take-all approach may be less sensitive to outlier spectra
because outliers will rarely represent the majority class within a crown.
Individual species classification accuracies were assessed (table 1). The highest clas-
sification accuracies were for species having large and densely foliated crowns. This
was especially true of Ficus microcarpa,Eucalyptus globulus and Magnolia grandiflora,
whose average crown areas in this study were 235, 197 and 161 m2, respectively. Two
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
518 M. Alonzo et al.
Figure 3. Decline in average overall classification accuracy with decreasing number of crowns
in training set. Each data point represents the average of 100 sample draws. The number and
wavelengths of the SDA bands selected depended on the training set. An average of 45 bands
were selected with a minimum of 27 and a maximum of 70.
of the three species with the lowest accuracies also had the smallest crown areas:
Lophostemon confertus (82 m2)andPyrus kawakamii (86 m2). Two notable excep-
tions to this trend were Ulmus parvifolia, with the third largest average crown area
but the second lowest classification accuracy and Pittosporum undulatum, with the
third smallest crown area and 94% accuracy. Coregistration error is a likely cause of
problems for crowns made up of few pixels. This error may be pronounced in cases
of tall trees and off-nadir viewing angles leading to significant spatial displacement of
crowns between the lidar and hyperspectral data sets. For two large-crowned species,
Ulmus parvifolia and Platanus racemosa, low accuracies may be due to sparsely foliated
crowns. This result supports previous research showing that leaf level spectral features
are most pronounced in areas with high leaf area index (Asner and Martin 2008).
3.2 Band selection
We hypothesized that accurate classification of urban tree species would require spec-
tral information from the full complement of AVIRIS bands. In this research, the
visible range (VIS) includes bands from 394 to 734 nm, the near infrared (NIR) spans
744–1313 nm, shortwave infrared 1 (SWIR1) is from 1443 to 1802 nm and shortwave
infrared 2 (SWIR2) is from 2018 to 2425 nm. Previous work has demonstrated the
importance of the NIR (Cochrane 2000), the NIR and SWIR1 (Gong et al. 1997,
Clark et al. 2005) and the VIS (van Aardt and Wynne 2001) to tree species classifica-
tion. In this research employing forward feature selection, bands were indeed chosen
in each region of the spectrum (figure 4(a)). Nevertheless, there was a notable emphasis
on bands in the VIS range especially when only considering the first 14 selected bands
(figure 4(b)). This grouping corresponds to relatively high coefficients of variation
(CV) in this region for the entire data set.
Examining the frequency of band selection and the corresponding CV values, it is
clear that the green edge, green peak and yellow edge exhibited high degrees of varia-
tion through the data set and were also important determinants of species separability.
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
Urban forest species classification from AVIRIS imagery 519
(a)
(b)
100
90
80
70
60
50
40
30
20
10
0
No. of times band selected
550 nm
550 nm
675 nm
675 nm
1207 nm
1722 nm
2038 nm 2427 nm
2038 nm
2287 nm
100
90
80
70
60
50
40
30
20
10
0
400 600 800 1000 1200
Wavelen
g
th (nm)
1400 1600 1800 2000 2200 2400
No. of times band selected
Figure 4. Frequency of stepwise selected bands is aligned with the overall data set’s coefficient
of variation (dashed line) and its normalized grand mean spectrum (solid line). (a) Frequency
with which each band was chosen using an MCR minimization criterion. On average, 45 bands
were selected per run. (b) Frequency with which the first 14 bands were selected using SDA with
an MCR minimization criterion.
In some cases, high CV values did not align with frequently selected bands. It is sur-
mised that in these cases, the high CV is an artefact of low signal-to-noise ratios
(e.g. 414 nm, 2437 nm) or an artefact from the reflectance retrieval, where a minor-
mismatch between AVIRIS-measured radiance and ATCOR-4 modelled radiance can
create pronounced spikes or troughs in reflectance. Further investigation is needed
to better understand why the NIR region was the least frequently selected. This may
be attributable to high degrees of within-species NIR variability causing irretrievable
between-species spectral overlap. It also may be the case that selective planting and
pruning in urban areas has limited canopy structural variation and thus dampened
the between-species NIR signal.
4. Conclusions
This research implemented canonical, stepwise and linear discriminant analysis for the
purpose of classifying urban tree species at the crown level. An average overall accu-
racy of 86% was attained using 14 variables from data rotated to maximize separability
using CDA. This 14-variable accuracy is the same as that which was attained using all
178 bands in an LDA classifier. Fourteen bands selected using SDA yielded an aver-
age accuracy of 43% indicating that more spectral information must be retained to
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
520 M. Alonzo et al.
adequately characterize the data set’s variability. Still, VIS bands along the green edge,
green peak and yellow edge were most frequently selected by SDA for their capacity to
reduce the overall MCR. This may indicate the relative importance of subtle chemical
differences among species that, due to their preferential selection for urban planting,
demonstrate less interspecific structural variability. Research is in progress to further
determine the utility of structural information for urban tree species identification.
Early results indicate that the fusion of lidar and hyperspectral data holds promise for
increasing classification accuracy especially of smaller trees.
Acknowledgements
The authors thank the Naval Postgraduate School (Award # N00244-11-1-0028) for
funding this research and Seth Peterson for general guidance. AVIRIS radiance
imagery was supplied by the Jet Propulsion Laboratory.
References
ASNER,G.andMARTIN, R., 2008, Spectral and chemical analysis of tropical forests: scaling
from leaf to canopy levels. Remote Sensing of Environment,112, pp. 3958–3970.
BENZ, U.C., HOFMANN, P., WILLHAUCK, G., LINGENFELDER,I.andHEYNEN, M., 2004,
Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready
information. ISPRS Journal of Photogrammetry and Remote Sensing,58, pp. 239–258.
CITY OF SANTA BARBARA, 2010, Publically managed trees database. Available online at: http://
www.santabarbaraca.gov/government/maps/ (accessed October 2010).
CLARK, M., ROBERTS,D.andCLARK, D., 2005, Hyperspectral discrimination of tropical
rain forest tree species at leaf to crown scales. Remote Sensing of Environment,96,
pp. 375–398.
COCHRANE, M., 2000, Using vegetation reflectance variability for species level classification of
hyperspectral data. International Journal of Remote Sensing,21, pp. 2075–2087.
FISHER, R., 1936, The use of multiple measures in taxonomic problems. Annuals of Eugenics,7,
pp. 179–188.
GONG, P., PU,R.andYU, B., 1997, Conifer species recognition: an exploratory analysis of in
situ hyperspectral data. Remote Sensing of Environment,62, pp. 189–200.
GREEN, R., EASTWOOD, M., SARTURE, C., CHRIEN, T., ARONSSON, M., CHIPPENDALE,B.,
FAUST, J., PAV R I, B., CHOVIT, C., SOLIS,M.andOLAH, M., 1998, Imaging spectroscopy
and the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS). Remote Sensing of
Environment,65, pp. 227–248.
JING, L., HU, B., LI,J.andNOLAND, T., 2012, Automated delineation of individual tree crowns
from lidar data by multi-scale analysis and segmentation. Photogrammetric Engineering
and Remote Sensing,78, pp. 1275–1284.
KLECKA, W., 1980, Discriminant Analysis (Beverly Hills, CA: Sage Publications).
LA´
CAN,I.andMCBRIDE, J.R., 2008, Pest Vulnerability Matrix (PVM): a graphic model for
assessing the interaction between tree species diversity and urban forest susceptibility to
insects and diseases. Urban Forestry & Urban Greening,7, pp. 291–300.
MYINT, S.W., GOBER, P., BRAZEL, A., GROSSMAN-CLARKE,S.andWENG, Q., 2011, Per-
pixel vs. object-based classification of urban land cover extraction using high spatial
resolution imagery. Remote Sensing of Environment,115, pp. 1145–1161.
NOWAK, D., CRANE, D., STEVENS, J., HOEHN, R., WALTON,J.andBOND, J., 2008, A ground-
based method of assessing urban forest structure and ecosystem services. Arboriculture
and Urban Forestry,34, pp. 347–358.
PU,R.andLIU, D., 2011, Segmented canonical discriminant analysis of in situ hyperspectral
data for identifying 13 urban tree species. International Journal of Remote Sensing,32,
pp. 2207–2226.
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
Urban forest species classification from AVIRIS imagery 521
RICHTER,R.andSCHLÄPFER, D., 2002, Geo-atmospheric processing of airborne imaging spec-
trometry data. Part 2: atmospheric/topographic correction. International Journal of
Remote Sensing,23, pp. 37–41.
ROUSE, J.W., HAAS, R.S., SCHELL, J.A. and DEERING, D.W., 1973, Monitoring vegetation
systems in the Great Plains with ERTS. In Proceedings,3rd ERTS Symposium,NASA
SP-351, vol. 1, pp. 48–62.
SANTAMOUR JR, F.S., 1990, Trees for urban planting: diversity, uniformity, and common sense.
In Proceedings of the Seventh Conference of the Metropolitan Tree Improvement Alliance
(METRIA), vol. 7, pp. 57–65.
VAN AARDT, J.A. and WYNNE, R.H., 2001, Spectral separability among six southern tree
Species. Photogrammetric Engineering & Remote Sensing,67, pp. 1367–1375.
WHITE, R.H. and ZIPPERER, W.C., 2010, Testing and classification of individual plants for
fire behaviour: plant selection for the wildland–urban interface. International Journal
of Wildland Fire,19, pp. 213–227.
XIAO, Q., USTIN,S.L.andMCPHERSON, E.G., 2004, Using AVIRIS data and multiple-masking
techniques to map urban forest tree species. International Journal of Remote Sensing,
25, pp. 5637–5654.
YU, B., OSTLAND, M., GONG,P.andPU, R., 1999, Penalized discriminant analysis of in situ
hyperspectral data for conifer species recognition. IEEE Transactions on Geoscience and
Remote Sensing,37, pp. 2569–2577.
ZHANG,C.andQIU, F., 2012, Mapping individual tree species in an urban forest using airborne
lidar data and hyperspectral imagery. Photogrammetric Engineering & Remote Sensing,
78, pp. 1079–1087.
ZHAO,G.andMACLEAN, A., 2000, A comparison of canonical discriminant analysis and prin-
cipal component analysis for spectral transformation. Photogrammetric Engineering &
Remote Sensing,7, pp. 841–847.
Downloaded by [University of California Santa Barbara] at 12:39 30 January 2013
