Agricultural Information Research 19(3), 2010. 71–78
Available online at www.jstage.jst.go.jp/
Original Paper
High Spatial Resolution Hyperspectral Mapping for Forest Ecosystem at Tree Species Level Gang Shen, Kenshi Sakai* and Yoshinobu Hoshino Faculty of Agriculture, Tokyo University of Agriculture and Technology, 183-8509, Japan
Abstract Integrated management of forest ecosystems requires an accurate and all-sided mastery of the forest information, of which forest ecosystem cover especially at tree species level is the most basic and important component. The study investigated and demonstrated the mapping potential of the forest ecosystem at tree species level from high spatial resolution hyperspectral images. The mapping performances of eight conventional classification methods including Maximum Likelihood (ML), Mahalanobis Distance (MaD), Minimum Distance (MD), Parallelepiped (P), Binary Encoding (BE), Spectral Angle Mapper (SAM), Spectral Information Divergence (SID) and Support Vector Machine (SVM), were verified based on two noise treatments (noise fraction and noise removal) and three leaf growth periods (tender leaf period, young leaf period and adult leaf period). It could be confirmed that noise removal obviously contributed to improving the classification agreement and young leaf period was most suitable for mapping forest ecosystem at tree species level from high spatial resolution hyperspectral images. ML, P, BE and SID were not considered appropriate according to good results with overall accuracy and kappa coefficient exceeding 85% and 0.80 respectively. Though MD also produced a very high classification agreement, it could not cover up its poor potential to identify tree species by spectral features. Even if SVML, SVMP, SVMR and SVMS performed the stablest and could generate good results across three periods, the best result was obtained by SAM. Except that the difference was significant between MD and SVMS at the 5% significance level in tender leaf period, the comparative tests did not provide more proof to show the significant difference between the methods considered appropriate.
Keywords high spatial resolution hyperspectral image, mapping, forest ecosystem
Introduction
ing and digital image analysis techniques, especially high spatial resolution hyperspectral images which have 48 or more bands
Forest ecosystems, which are defined as terrestrial ecosystems
with a spectral resolution of 20 nm or smaller and a spatial reso-
with the tree canopy covers more than 10% of the ground area
lution of 5 m or less (Aspinall et al. 2002a), facilitate the acquisi-
(Matthews et al. 2000), supply a wide range of commodities
tion and extraction of forest ecosystem information.
sought by an expanding human population, such as structural
From high spatial resolution hyperspectral images, Cho et al.
materials, fuels and medicines, along with a wide range of critical
(2009) mapped three forest structural parameters including mean
ecosystem services including nutrient cycling, climate regulation,
diameter-at-breast height (DBH), mean tree height and tree den-
maintaining water balance and carbon sequestration (Klenner et
sity of a closed canopy beech forest with accuracy 72.4%, 67.4%
al. 2009). Managing forests simultaneously for wood, biodiver-
and 53.6%, respectively. Dehaan et al. (2007) quantified the dis-
sity, carbon sequestration, energy, water quality, flood control,
tribution of blackberry in open canopies with an overall accuracy
habitat and recreation is the 21st century challenge (Burger
92%. Aspinall (2002b) got an excellent agreement of mapping
2009). Integrated management requires an accurate and all-sided
depicting large woody debris and three Populus spp. within a
mastery of the forest information, of which forest ecosystem
single genus in a riparian area in Yellowstone National Park,
cover is the most basic and important component. Remote sens-
Wyoming, USA. However, there have been little studies on mapping the forest ecosystems at tree species level from high
* Corresponding Author E-mail:
[email protected]
spatial resolution hyperspectral images. Furthermore, tree species composition is a primary attribute of forest ecosystems (Barbier 71
High Spatial Resolution Hyperspectral Mapping
et al. 2009), and single species or species groups of forest ground
annual mean temperature is 15°C, and monthly mean temperature
vegetation can be used as indicators for site conditions (Khanina
is highest in July and lowest in February, 33.1°C and 3.1°C
et al. 2007).
respectively. The area receives an annual mean precipitation of
This study was designed to: (i) investigate the mapping poten-
1,750 mm and an annual mean snowfall of less than 10 cm.
tial of the forest ecosystems at tree species level from high spatial
The landscape is partially deforested and subsequent regenera-
resolution hyperspectral images; (ii) verify the mapping per-
tion has created complex mosaic of different vegetation commu-
formance of eight conventional classification methods including
nities. Quercus (Quercus serrata, Quercus acutissima, Quercus
Maximum Likelihood (ML), Mahalanobis Distance (MaD),
glauca) is dominant, with Pinus (Pinus densiflora and Pinus
Minimum Distance (MD), Parallelepiped (P), Binary Encoding
thunbergii), Cinnamomum camphora, Cryptomeria japonica,
(BE), Spectral Angle Mapper (SAM), Spectral Information
Phyllostachys bambusoides, Castanea mollissima, Prunus spp.,
Divergence (SID) and Support Vector Machine (SVM); (iii)
Diospyros kaki, Prunus mume and grassland also occurring in
understand the effects of noise treatments and leaf growth periods
various mixtures. Additionally, the other types of landuse include
on the classification performance.
orchard, road, water surface and buildings.
Data and Methodology
Image data acquisition and pre-processing
Site descriptions
for three different periods, which represented tender leaf period,
High spatial resolution hyperspectral images of the study area The focus of this study is an incity forest park located within
young leaf period and adult leaf period respectively, and coinci-
the east of Hachioji city in Japan (Fig. 1), a field science educa-
dent field data area were used in the study. The images were
tion research center affiliated with Tokyo University of Agricul-
acquired on 10 April, 22 May, 27 June in 2003 by an Airborne
ture and Technology. The landscape covers approximate 65 ha
Imaging Spectrometer for Applications (AISA) Eagle system
with an elevation range of 140–180 m above sea level. The
(Pasco Co. Ltd., Tokyo, Japan). AISA Eagle sensor was mounted
Fig. 1
72
Colorful image of the study area by bands 40 (660.85 nm), 25 (561.38 nm), and 8 (451.87 nm) as R, G, B channels.
Agricultural Information Research 19(3),2010
on an aircraft flying along SW-NE path at an altitude of 1,198 m
and remove noise. MNF transforms is usually adopted to deter-
during cloud-free periods in the daytime. The images on 10 April
mine the inherent dimensionality of image data, to segregate
and 27 June were collected with 1.5×1.5 m spatial resolution but
noise in the data, and to reduce the computational requirements
for on 22 May with 2.0×2.0 m spatial resolution. All images
for subsequent processing (Boardman and Kruse 1994). It is a
included 72 bands located in the 406.92–898.55 nm spectral
linear transformation that consists of the following two separate
range with 6.3 nm nominal spectral resolution.
principal component analysis rotations: the first rotation decorre-
Images were georectified and atmospherically calibrated using
lates and rescales the noise in the data by the noise covariance
the modified flat field method. Spectrum of a sample field were
matrix; the second rotation implements a standard principal com-
obtained using a spectrophotometer during the flight of AISA
ponent analysis based on previous result (Green et al. 1988). The
Eagle on that day, which were used for atmospheric calibration.
effects of noise treatments were investigated by three treatments
These images were then processed to surface reflectance and
such as no treatment, noise fraction and noise removal.
imported into the Environment for Visualizing Images (ENVI
In this study, eight well-developed supervised classification
4.5, ITT Visual Information Solutions) software for the sub-
methods, including Maximum Likelihood (ML), Mahalanobis
sequent analysis.
Distance (MaD), Minimum Distance (MD), Parallelepiped (P), Binary Encoding (BE), Spectral Angle Mapper (SAM), Spectral
Ground data collection
Information Divergence (SID) and Support Vector Machine
Based on the thematic vegetation map from the field investiga-
(SVM), were used to classify the forest vegetation at tree species
tion conducted in 2003, the study determined twelve ground
level from high spatial resolution hyperspectral images. ML
cover classes including Cinnamomum camphora, Pinus,
assumes that the statistics for each class in each band are normally
Cryptomeria japonica, Shadow & gap, Water surface, Grassland,
distributed and calculates the probability that a given pixel
Castanea mollissima, Bareland, Quercus, Diospyros kaki,
belongs to a specific class (Evans 1998). In Mahalanobis
Phyllostachys bambusoides, Prunus mume. A common rectangu-
Distance method, unknown pixels are assigned on basis of the
lar area with 468×400 pixels for three different periods was
distance referred to the mean vectors and covariance matrix
selected for training and validation. In order to obtain pure distri-
(Maesschalck et al. 2000). MD calculates the Euclidean distance
bution areas of mono tree species, supplementary surveys were
from each unknown pixel to the mean vectors for each class
implemented in 2009 referring to the minimum noise fraction
(Wacker and Landgrebe 1972). Parallelepiped classifies images
(MNF) transform images. The colorfully composed image of the
using an n-dimensional parallelepiped with boundaries defined
top three principal component bands resulted from MNF trans-
based upon the standard deviation from the mean of each selected
form accumulatively contributed 46.25% information toward
class (Sivakumar et al. 2004). BE classifies all unknown pixels
the image, which clearly indicated pure distribution areas of tree
into the classes with the greatest number of match bands by
species by different colors. Cluster sampling for training was
encoding the data and reference spectrum into zero and one based
adopted to contain all spectral features of one ground cover as
on whether a band falls below or above the spectral mean (Mazer
possible, along with the validation area determination. The
et al. 1988). SAM determines the spectral similarity between two
ground data was also imported into the ENVI software for deter-
spectra by calculating the angle between the two spectra, treating
mining image-derived class endmembers and validation pixels.
them as vectors in a space with dimensionality equal to the num-
Considering adequate and relative coincident amounts of valida-
ber of bands (Kruse et al. 1993). SID compares the similarity
tion pixels for each class, nine classes including Cinnamomum
between two pixels by measuring the probabilistic discrepancy
camphora, Pinus, Cryptomeria japonica, Shadow & gap, Water
between two corresponding spectral signatures (Chang 2000).
surface, Grassland, Castanea mollissima, Bareland and Quercus
The basic idea of SVM is to find a hyperplane which separates the
were used as validation analysis.
n-dimensional data perfectly (Vapnik 1995) and four popular kernel functions (Linear, Polynominal, Radial Basis Function and
Image processing
Sigmoid) were used here, which were abbreviated to SVML,
Image of 22 May was resampled to 1.5×1.5 m spatial resolu-
SVMP, SVMR and SVMS respectively. However, ML and MaD
tion coincident with the other two periods. In order to depress the
were only applied to noise removal images due to the limitation of
negative impacts of the spectral feature average on the identifica-
training pixel numbers.With noise treatments and classification
tion potential of ground cover, training pixels of each class were
methods mentioned previously, images of three continuous
further subdivided into subclasses by the natural cluster of spec-
periods were analyzed to understand the impacts of leaf growth
tral features. All subclasses were used as endmember classes for
periods on the classification performance.
classification and then merged before validation analysis. The most commonly used MNF transform was employed to segregate 73
High Spatial Resolution Hyperspectral Mapping
Assessment and comparison for classification performance As a simple cross-tabulation of the mapped class label against
k1 – k2 z = --------------------------2 2 σ k1 + σ k2
that observed in the ground or reference data for a sample of cases
where k1 and k2 represent respectively the kappa coefficients for
at specified locations, the confusion matrix is at the core of accu-
two classifications and σ2k1 and σ2k2 their associated variance
racy assessment (Canters 1997, Foody 2002). Based on the con-
(Congalton and Green 1998).
fusion matrix, overall accuracy (OA) and kappa coefficient (KC) were adopted to weigh the classification performance. Overall
Results and Discussion
accuracy is the percentage of cases correctly allocated, which exceeding 85% is considered good (Thomlinson et al. 1999).
The effects of noise treatments on classification agreement
Kappa coefficient is a statistical measure of classification agree-
With the generality of noise in high spatial resolution hyper-
ment for qualitative (categorical) items, which exceeding 0.80
spectral images, it is very necessary to make certain how noise
represents almost perfect agreement of the mapped data and
treatments affect the classification agreement. For that reason,
validation data (Landis and Koch 1977). Furthermore, one of the
different noise-treated images including no treatment images,
most widely promoted means for classification accuracy compar-
noise fraction images and noise removal images were classified
ison is through the comparison of kappa coefficients (Foody
with eight methods for every period. All classification agreement
2009). With this approach, the statistical significance of the
(kappa coefficient) was shown in Fig. 2.
difference between two independent kappa coefficients is calculated by a z-value with
From Fig. 2, under the circumstances of no noise treatment, only SAM and SID produced results with kappa coefficient exceeding 0.70 in the young leaf period. On the whole, there were
Fig. 2
74
The effects of noise treatments on classification agreement
Agricultural Information Research 19(3),2010
not satisfactory results with kappa coefficient exceeding 0.80 at
type, the reflectance spectrum also depends on other factors such
all for eight methods across three leaf growth periods from the
as the leaf growth periods. Thus, an essential component of
original high spatial resolution hyperspectral images without
mapping forest ecosystem is the identification of key leaf growth
noise treatment.
periods to determine when one vegetation species can best dif-
In the tender leaf period, noise fraction and especially noise
ferentiate from others. In the study, the images of the study area
removal remarkably improved the classification agreement for
for three continuous periods , which represented tender leaf
MD, BE, SAM, SID, SVML, SVMP, SVMR and SVMS, and
period, young leaf period and adult leaf period respectively, were
conversely noise fraction cut down it for Parallelepiped. In the
analyzed to understand the effects of leaf growth periods on the
young leaf period, noise removal greatly promoted the classifica-
classification overall accuracy for every method. Fig. 3 gave the
tion agreement for MD, BE, SAM, SVML, SVMP, SVMR and
comparison of overall accuracy in three periods for eight methods
SVMS, but lowered it for P and SID. At the same time period,
and accuracy value was the best result of that period for each
noise fraction improved the classification agreement slightly for
method.
SVML, SVMP, SVMR and considerably for MD and BE, but
For MD, BE, SVMP and SVML, overall accuracy declined
conversely lowered it for P, SAM and SVMS. Similar to the ten-
gradually from tender leaf period, young leaf period to adult leaf
der leaf period, noise removal was consistent with improving the
period, but conversely for SVMS. Relative to tender leaf period,
classification agreement for MD, P, BE, SAM, SID, SVML,
it seemed that images of adult leaf period and especially young
SVMP, SVMR and SVMS in the adult leaf period. Noise fraction
leaf period were more suitable to identify the forest vegetation for
obviously improved the classification agreement just for MD,
P, MaD and SAM. From the young leaf period, SID and SVMR
BE, SAM and SID, whereas lowered it for P, SVML, SVMP,
received higher overall accuracy relative to tender leaf period and
SVMR and SVMS. From Fig. 2, it could not be confirmed
especially adult leaf period, which was just opposite to ML.
whether noise fraction was able to improve the classification
Fig. 4 displayed the comparison of performance stability for clas-
agreement. However, noise removal was consistent with improv-
sification methods.
ing the classification agreement cross different methods and leaf growth periods , except for P and SID in the young leaf period.
ML, MaD and P manifested the worst stability with standard deviation great than 5 cross three periods. The stability for MD,
Therefore, it could be confirmed that noise removal is a very
BE, SAM and SID was moderate and SVML, SVMP, SVMR and
important process to obtain the high quality mapping from high
SVMS performed stablest. In general, SVML, SVMP, SVMR
spatial resolution hyperspectral images. Meanwhile, it also dem-
and SVMS were tolerant toward subtle variation in spectral fea-
onstrated the excellent potential of MNF transform to separate
tures induced by leaf growth periods. Although the performance
and remove the noise of high spatial resolution hyperspectral
of MD and SAM fluctuated obviously with the different leaf
images.
growth periods, they submitted the most satisfactory results with overall accuracy greater than 90% in tender leaf period and young
The effects of leaf growth periods on the classification overall
leaf period, respectively.
accuracy Vegetation has its own unique spectral signature, which can be used for identification of vegetation type. For the same vegetation
Fig. 3
The comparative tests of classification difference Classification accuracy statements are often compared in
The effects of leaf growth periods on the classification overall accuracy
75
High Spatial Resolution Hyperspectral Mapping
Fig. 4
The stability of classification methods
Table 1 The z-values between excellent classification methods MD Tender leaf period
SAM
MD
SVML 1.5875
SVML
SVMP
SVMR
1.3007
1.4938
2.6761
–0.2416
–0.0530
1.1956
SVMP
0.1846
SVMR Young leaf period
MaD MD
1.3996 1.2243
–0.3895
–1.4038
–0.3890
–0.3397
–0.4423
–0.3532
0.0408
0.0043
–0.0769
0.3582
0.4479
0.3132
0.2140
1.3540
–0.0326
–0.1181
0.3441
SAM SVML SVMP
–0.0754
SVMR Adult leaf period
SVMS
SAM SVML SVMP
–0.0895
0.3043 0.4073
–1.1288
–0.7730
–0.8856
0.3528
0.2277
0.0512
–0.1209
–0.3176
SVMR
–1.1286
–0.1872
remote sensing research and the central focus of such compara-
level. The difference was significant between SVMP and SVMS
tive analyses has been the magnitude of the difference between
at the 20% level. In the young leaf period, SAM was more out-
the accuracy values (Foody 2009). The methods with excellent
standing than MaD and SVMS only at the 20% level. In the adult
performance were filtrated by a combination of overall accuracy
leaf period, there was no significance between methods beyond
and kappa coefficient exceeding 85% and 0.80 respectively. As a
the 20% level, which proved that the forest in the adult leaf period
result, MD, SVML, SVMP, SVMR and SVMS can be used to
promoted the mixture of spectra and complicated the classifica-
map the forest ecosystem of tender leaf period; MaD, MD, SAM,
tion so that it was difficult to show apparent difference between
SVML, SVMP, SVMR and SVMS can be used to map the forest
classification methods. Meanwhile, the fact that more methods
ecosystem of young leaf period; SAM, SVML, SVMP, SVMR
received excellent results in young leaf period proved it was more
and SVMS can be used to map the forest ecosystem of adult leaf
suitable for forest ecosystem mapping at tree species level from
period. Moreover, the statistical significance of the difference
high spatial resolution hyperspectral images.
between two excellent classification was calculated by a widely promoted z-value (Table 1).
Forest ecosystem mapping of the study area at tree species level
In the tender leaf period, MD was obviously more excellent
The most excellent method in every period was used to map the
than the others. The difference was significant between MD and
forest ecosystem of the study area in that period at tree species
SVMS at the 5% significance level. The difference was sig-
level and the results were presented in Fig. 5. Compared with
nificant between MD and SVML, SVMR at the 15% level. The
ground data, (b) reflected the classes and their distributions of
difference was significant between MD and SVMP at the 20%
forest ecosystem more factual than (a) and (c). (a) misclassified
76
Agricultural Information Research 19(3),2010
Fig. 5
The forest ecosystem mapping of the study area for three periods at tree species level. (a) was mapped with MD method for tender leaf period; (b) was mapped with SAM method for young leaf period; (c) was mapped with SVML method for adult leaf period.
Cryptomeria japonica into Grassland on the same positions of A
species on account of sparse canopy in the tender leaf period, and
and B which were labelled in (b) of Fig. 5. Meanwhile, there was
overmuch overlap of leaves between different tree species
an obvious confusion between Pinus and Cryptomeria japonica
complicated the distinction of them in the adult leaf period. The
in (a) on the same position of C. (c) misclassified Pinus into
mapping performance of eight conventional classification meth-
Prumnus mume on the same position of C and misclassified Pinus
ods including Maximum Likelihood (ML), Mahalanobis Distance
into Cinnamomum camphora on the same position of D. (c) also
(MaD), Minimum Distance (MD), Parallelepiped (P), Binary
did not identify Pinus on the same position of E. In the tender leaf
Encoding (BE), Spectral Angle Mapper (SAM), Spectral Infor-
period, the spectrum from stem and background were the most
mation Divergence (SID) and Support Vector Machine (SVM)
important interferences on account of sparse canopy. However,
were also verified in the study. ML, P, BE and SID were not
overmuch overlap of leaves from different tree species compli-
considered appropriate to identify and map the forest ecosystem
cated the distinction of different tree species in the adult leaf
at tree species level from high spatial resolution hyperspectral
period. Thus, it could be made certain that young leaf period was
images according to criteria with overall accuracy and kappa
more suitable for mapping the forest ecosystem at tree species
coefficient exceeding 85% and 0.80 respectively. MaD per-
level, which resulted in previous results. Simultaneously, the
formed well just occasionally. Though MD also produced very
results also proved the outstanding potential of SAM for mapping
high classification agreement, Fig. 2 and Fig. 4 indicated an obvi-
the forest ecosystem at tree species level.
ous drop-off of mapping potential in the rear two periods, which devulged its poor potential to identify tree species by spectral
Conclusions
features. Visible misclassification in Fig. 5 (a) conflicted with its high agreement which might be aroused by randomicity of artifi-
The study has demonstrated the excellent mapping potential of
cial sampling in validation areas. Even if SVML, SVMP, SVMR
the forest ecosystem at tree species level from high spatial resolu-
and SVMS performed stablest and could generate good results
tion hyperspectral images. As the most commonly used process-
across all three periods, the best result was obtained by SAM.
ing technique, MNF transform could effectively separate and
Except that the difference was significant between MD and
remove the noise of high spatial resolution hyperspectral images.
SVMS at 5% significance level in tender leaf period, the compar-
Although whether noise fraction was able to improve the classifi-
ative tests did not provide more proof to show the significant dif-
cation agreement was unregular, it could be confirmed that noise
ference between the excellent methods considered appropriate.
removal obviously contributed to improving the classification
This study provided forest managers with scientific basis of map-
agreement for absolute majority methods cross different leaf
ping selection for forest ecosystem from high spatial resolution
growth periods. There were always some methods to produce
hyperspectral images by the view of noise treatments and classi-
good results for three periods, whereas majority excellent results
fication periods.
were generated in young leaf period. It might be interpreted that stem and background interfered greatly with distinction of tree 77
High Spatial Resolution Hyperspectral Mapping
Acknowledgement This study was funded by the Japan Society for the Promotion of Science (JSPS) grant-in-aid for scientific research project (No. 14360148).
References Aspinall, R. J., W. A. Marcus and J. W. Boardman (2002a) Considerations in Collecting, Processing, and Analysing High Spatial Resolution, Hyperspectral Data for Environmental Investigations, Journal of Geographical Systems, 4: 15–29. Aspinall, R. J. (2002b) Use of logistic regression for validation of maps of the spatial distribution of vegetation species derived from high spatial resolution hyperspectral remotely sensed data, Ecological Modelling, 157: 301–312. Barbier, S., F. Gosselin and P. Balandier (2009) Influence of tree species on understory vegetation diversity and mechanisms involved—A critical review for temperate and boreal forests, Forest Ecology and Management, 254: 1–15. Boardman, J. W and F. A. Kruse (1994) Automated spectral analysis: a geological example using AVIRIS data, North Grapevine Mountains, Nevada, “Proceedings of the Tenth Thematic Conference on Geological Remote Sensing”, Environmental Research Institute of Michigan, San Antonio, TX, 407–418. Burger, J. A. (2009) Management effects on growth, production and sustainability of managed forest ecosystems: Past trends and future directions, Forest Ecology and Management, 258: 2335–2346. Canters, F. (1997) Evaluating the uncertainty of area estimates derived from fuzzy land-cover classification, Photogrammetric Engineering and Remote Sensing, 63: 403–414. Chang, C. I. (2000) An information theoretic-based approach to spectral variability, similarity and discriminability for hyperspectral image analysis, IEEE Transactions on information theory, 46: 1927–1932. Cho, M. A., A. K. Skidmore and I. Sobhan (2009) Mapping beech (Fagus sylvatica L.) forest structure with airborne hyperspectral imagery, International Journal of Applied Earth Observation and Geoinformation, 11: 201–211. Congalton, R. G and K. Green (1998) “Assessing the accuracy of remotely sensed data: principles and practices”, Lewis Publishers, New York. Dehaan, R., J. Louis, A. Wilson, et al. (2007) Discrimination of blackberry (Rubus fruticosus sp. agg.) using hyperspectral imagery in Kosciuszko National Park, NSW, Australia, ISPRS Journal of Photogrammetry & Remote Sensing of Environment, 62: 13–24. Evans, F. (1998) “An investigation into the use of maximum likelihood classifiers, decision trees, neural networks and conditional probabilistic networks for mapping and predicting salinity”, MSc thesis, University of Western Australia.
78
Foody, G. M. (2002) Status of land cover classification accuracy assessment, Remote Sensing of Environment, 80: 185–201. Foody, G. M. (2009) Classification accuracy comparison: hypothesis tests and the use of confidence intervals in evaluations of difference, equivalence and non-inferiority, Remote Sensing of Environment, 113: 1658–1663. Green, A. A., M. Berman, P. Switzer, et al. (1988) A transformation for ordering multispectral data in terms of images quality with implications for noise removal, IEEE Transactions on Geoscience and Remote Sensing, 26: 65–74. Khanina, L., M. Bobrovsky, A. Komarov, et al. (2007) Modeling dynamics of forest ground vegetation diversity under different forest management regimes, Forest Ecology and Management, 248: 80– 94. Klenner, W., A. Arsenault, E. G. Brockerhoff, et al. (2009) Biodiversity in forest ecosystems and landscapes: A conference to discuss future directions in biodiversity management for sustainable forestry, Forest Ecology and Management, doi:10.1016/j.foreco.2009.10.037. Kruse, F. A., A. B. Lefkoff, J. W. Boardman, et al. (1993) The spectral image processing system (SIPS)-interactive visualization and analysis of imaging spectrometer data, Remote Sensing of Environment, 44: 145–163. Landis, J. R and G. Koch (1977) The measurement of observer agreement for categorical data, Biometrics, 33: 159–174. Maesschalck, R. D., D. Jouan-Rimbaud and D. L. Massart (2000) The mahalanobis distance, Chemometrics and Intelligent Laboratory Systems, 50: 1–18. Matthews, E., R. Payne, M. Rohweder, et al. (2000) “Pilot analysis of global ecosystems: forest ecosystems”, World Resources Institute, Washington, D.C. Mazer, A. S., M. Martin, M. Lee, et al. (1988) Image processing software for imaging spectrometry data analysis, Remote Sensing of Environment, 24: 201–210. Sivakumar, M. V. K., P. S. Roy, K. Harmsen, et al. (2004) “Satellite remote sensing and GIS applications in agricultural meteorology”, World Meteorological Organisation, Switzerland. Thomlinson, J. R., P. V. Bolstad and W. B. Cohen (1999) Coordinating methodologies for scaling landcover classifications from site-specific to global: steps toward validating global map products, Remote Sensing of Environment, 70: 16–28. Vapnik, V. N. (1995) “The nature of statistical learning theory”, SpringerVerlag, New York. Wacker, A. G and D. A. Landgrebe (1972) Minimum distance classification in remote sensing, Laboratory for Applications of Remote Sensing, Purdue University. Received January 7, 2010 Accepted May 24, 2010 Agro-informatics & Technology
