
Research Article


Effects of Hyperspectral Data Transformations on Urban Interclass Separations using a Support Vector Machine 

M.A. Misman,
H.Z.M. Shafri
and
Raja M. Kamil Raja Ahmad



ABSTRACT

This study investigated the performance of different data types used in a hyperspectral data classification process. Data in the form of spectral reflectance, first derivative spectra and wavelet coefficients were used as inputs for the Support Vector Machine (SVM) algorithm used to classify five different classes. The first derivative spectra gave a lower classification accuracy (35.6%) than the spectral reflectance (82%) and the use of wavelet coefficients further improved the classification accuracy to 100%. Proper selection of the wavelet transformation method, the mother wavelet, the number of vanishing moments and the decomposition level could improve classification accuracy. In summary, wavelet coefficients could maximise discrimination capability as compared to the spectral reflectance and first derivative spectra.





Received: April 23, 2010;
Accepted: July 05, 2010;
Published: July 27, 2010


INTRODUCTION The availability of various remote sensing data provides an opportunity for users to fully utilise the data to achieve their goal with a maximum success rate. These opportunities arise because remote sensing technology has been used in many applications including mapping, the military, meteorology, agriculture and others. Different types of data have their own usefulness. Multispectral data with high spatial resolution is suitable for mapping and hyperspectral data is more suitable for subtle discrimination.
Hyperspectral data have been widely used because of their capability in discriminating
between subtle variations among similar features and they could improve the
user’s capability for gaining a greater understanding of various features.
Although, it has been extensively used in agricultural applications (Rao
et al., 2007; Gong et al., 2002;
Mutanga et al., 2004; Chappelle
et al., 1992; Hansen and Schjoerring, 2003),
hyperspectral data have also been extensively used for urban and suburban applications
(Heiden et al., 2007; Bassani
et al., 2007).
Several methods such as clustering (GomezChova et al.,
2009; Oldeland et al., 2010) or indices (Oldeland
et al., 2010; Huang et al., 2009;
Kuckenberg et al., 2008; Devadas
et al., 2008) are possible for discriminating between one feature
and others. Classification (Laverington, 2010; Wang
and Sousa, 2009; Lucas et al., 2008; Wilson
et al., 2004; CastroEsau et al., 2004)
is also one of the methods used to discriminate between features. There are
several factors that can contribute to the success of a classification process,
including the selection of data type for use as input for the classification
(Koger et al., 2003; Yang
et al., 2009; Ouma et al., 2008),
the selection of the optimal band (Phillips et al.,
2009; Keshava, 2004; Murakami,
2004; Serpico and Moser, 2007) and the use of classifiers
(Shafri and Ramle, 2009; Clark et
al., 2005; Yang et al., 2009; Erbek
et al., 2004; Ali et al., 2009; Nelson,
1981). These are the major factors that will impact the classification accuracy.
For hyperspectral data, the original data are normally in the form of the spectral
reflectance. Exploitations or transformations of the original data could be
performed to improve the classification results. Calculating the spectral derivative
is one of the transformations that can be applied to the reflectance data and
several studies have shown that derivative data could achieve better results
than the reflectance data (Tsai and Philpot, 1998; Han,
2002). Wavelet transformations can also be used as a data transformation
method and have been used for applications like feature extraction, image compression,
feature detection and others (Hsu, 2007; Koger
et al., 2003; Galford et al., 2008;
Bruce et al., 2002; Loum
et al.,2007; Ping et al., 2009; Raju
et al., 2008).
 Fig. 1:  The
study area and the AISA image 
However, the best data type has yet to be determined for use in classifying
different urbanarea features while using data from an Airborne Imaging Spectrometer
for Application (AISA). This is particularly important in the Malaysian context
because AISA is currently the only airborne hyperspectral sensor available through
a commercial data provider. Thus, this study focused on determining the effects
of data input selection on the classification of interclass features by using
AISA data.
MATERIALS AND METHODS The data used in this study were acquired in November 2009 over an area in Kuala Lumpur, Malaysia (Fig. 1a). With a pixel size of approximately 1x1 m, the AISA image covers a spectral region of 4001000 nm with a spectral resolution of approximately 5 nm for 128 bands. The AISA image used in this study is shown in Fig. 1b. The three major steps involved in this study were image processing, data transformation and classification. The general methodology of this study is shown in Fig. 2.
Image processing: The AISA image required processing before feature
extraction could be made. The image was converted to reflectance because it
was collected in radiance mode to minimise the effects of the atmosphere. The
conversion method used was a logresidual method in ENVI software. After the
conversion to reflectance was made, a Minimum Noise Fraction (MNF) transformation
was performed.
 Fig. 2: 
Methodology flowchart 
This step minimised the noise in the image. Twenty seven firstMNF bands were
used to obtain a denoised image that had less noise than the original image
obtained using the Inverse MNF method. Next, the Pixel Purity Index (PPI) was
performed to find the pixels of highest purity in the image that represent certain
features.
Feature extraction: The feature extraction was performed based on the PPI result. Five features were selected for further processing: vegetation, water, road, concrete and rooftop. One hundred pixels were extracted from the denoised image for each feature. Each pixel represented the signal for that particular feature. Those signals were used for further processing. Figure 3 shows the average spectral reflectance for the selected features. Data transformation: The methodology for data transformation is shown in Fig. 4. Three major steps were involved in this process and included spectral derivative conversion followed by wavelet conversion of the spectral reflectance and derivative data using a Continuous Wavelet Transform (CWT) and a Discrete Wavelet Transform (DWT). Selection of the mother wavelet was also performed along with selection of the number of vanishing moments and the level of decompositions.
The original spectral reflectance data can be transformed into other types
of dimensional data by applying mathematical operations. Using transformed data
may provide better information and understanding than using the original data.
For instance, the spectral derivative enhances the spectral differences in certain
parts of the spectrum, removes multiplicative factors and reduces the effect
of the soil background (Tsai and Philpot, 1998; Gong
et al., 2001).
 Fig. 3: 
Average spectral reflectance for vegetation, water, road,
concrete and rooftop features 
All of the spectral reflectance samples used in this study were transformed
into first derivative by using Eq. 1 (Dawson
and Curran, 1998).
where, FDR is the firstderivative reflectance at wavelength i, R_{(j)} is the reflectance at wavelength j, R_{ÆÆ(j+1)} is the reflectance at wavelength j+1 and Δ_{ë} is the difference in wavelength between j and j+1. Wavelet transformations were then applied to the spectral reflectance and the firstderivative data. A wavelet is a mathematical function used to divide a given function (or signal) into different scale components and a wavelet transform is the representation of a function by wavelets. Wavelets have advantages over traditional Fourier transform for representing functions that have discontinuities and sharp peaks. Wavelets also have advantages for deconstructing and reconstructing a signal. Discrete Wavelet Transform (DWT) and Continuous Wavelet Transform (CWT) are two types of wavelet transformations.
There are several types of mother wavelets and each mother wavelet has its
own characteristics. The mother wavelets are different in terms of their orthogonality,support,
regularity and symmetry.
 Fig. 4: 
Flowchart for the methodology of data transformation 
 Fig. 5: 
Wavelet families representing (a) Haar, (b) Daubechies, (c)
Symlet, (d) Coiflet, (e) Biorthogonal, (f) Reverse Biorthogonal and (g)
Discrete Meyer 
 Fig. 6: 
Example of a threelevel DWT decomposition 
As a result, wavelet coefficients vary according to the selection of mother
wavelet. This study uses only seven mother wavelets (Fig. 5ag):
Haar, Daubechies, Symlet, Coiflet, Biorthogonal, Reverse Biorthogonal and Discrete
Meyer.
Both DWT and CWT were used in this study. Each spectral reflectance and firstderivative spectrum was transformed or decomposed into wavelet coefficients with nine levels of decomposition. For DWT, each decomposition process produced two types of coefficients, the approximation coefficient (cA) and the detail coefficient (cD). Their values are illustrated in Fig. 6. Only the detail coefficient from each level was used as an input for classification. DWT and CWT were tested to determine the best method of wavelet transformation for use in distinguishing features. Both the spectral reflectance and the first derivative data were decomposed using the selected wavelet families to study the effects of mother wavelet selection on discrimination capability and to investigate which of the mother wavelets performed best. Haar, Daubechies (db1db20), Symlet (sym1sym20), Coiflet (coif1coif5), Biorthogonal (bior1.1bior6.8), Reverse Biorthogonal (rbio1.1rbio6.8) and Discrete Meyer (dmey) are the seven mother wavelets that were tested in this study. They were decomposed, levelbylevel, up to level nine. Since the transformation was applied to both the spectral reflectance and the firstderivative spectra, the analysis generally produced two groups of wavelet coefficient data. One group contained the wavelet coefficients derived from the spectral reflectance dataset and the other group was derived from the firstderivative dataset. The classification of datasets containing the spectral reflectance, the firstderivative and the wavelet coefficients were performed after the wavelet transformation process was completed.
Support vector machine classification: Classification was one of the
methods used for information extraction. Various supervised and unsupervised
classification algorithms may be used to assign data to one possible class.
The choice of classifier (i.e., decision rule) depends on the nature of input
data and the desired output (Jensen, 2005).
 Fig. 7:  Illustration
of the SVM process in twodimensional space. Blue dots represent data
from group 1 while red dots represent data from group 2 
The Support Vector Machine (SVM) method was selected because it is considered
the most suitable classifier for handling limited samples (Chi
et al., 2008). SVM is a supervised learning method used for classification
and regression. It constructs a separating hyperplane between two sets of data
in ndimensional space. The hyperplane will maximise the margin between the
two data sets. A good separation is achieved by the hyperplane that has the
largest distance separating the neighbouring data points of both classes. Larger
margins will lower the generalisation error of the classifiers.
The spectral reflectance, the first derivative of spectral reflectance, the wavelet coefficients derived from the spectral reflectance and the wavelet coefficients derived from the firstderivative spectra were used as inputs for SVM classification. Fifty samples from each group were used as a training dataset. To eliminate any potential bias, only the remaining separate samples were used as testing dataset. The performance of the spectral reflectance, firstderivative and allwavelet coefficients data were evaluated based on the classification accuracy (Fig. 7). RESULTS AND DISCUSSION
First derivative transformation: The values of the spectral reflectance
data depend on many factors including sun illumination. Use of the firstderivative
data has an advantage over use of the reflectance data because it minimises
the effects of sun illumination. Several studies have established the usefulness
of the firstderivative data for separating classes more effectively than the
reflectance data (Tsai and Philpot,1998;
Han, 2002).
 Fig. 8: 
Firstderivative spectra of the spectral reflectance of each
class 
Figure 8 shows the firstderivative spectra of the average
spectral reflectance for each class.
Wavelet coefficients: Different wavelet coefficient values were obtained
from the decomposition of different mother wavelets. There were also differences
in wavelet coefficient values obtained from the spectral reflectance and firstderivative
spectra. Although, the spectral reflectance and the first derivative spectra
were decomposed using the same mother wavelet, their wavelet coefficient values
resulting from decomposition were different. It has been suggested for some
time that the use of spectral derivatives can reduce the illumination and other
effects, thus, the use of spectral derivatives as input data may offer further
benefits when applying wavelet analysis (Blackburn and Ferwerda,
2008). In addition, the number of vanishing moments used for the decomposition
also affected the resulting wavelet coefficient values. Figure
9 shows an example of the wavelet coefficient values obtained from spectral
reflectance data and firstderivative spectra obtained from a vegetation sample.
Generally, the coefficient values of each mother wavelet varied because of different
characteristics the mother wavelets. The coefficient values derived from firstderivative
spectra were also smaller than the coefficient values derived from the spectral
reflectance. More visualization of wavelet transformation of spectral reflectance
and firstderivative using DWT and CWT methods by Blackburn
and Ferwada (2008).
Classification results: The performance of all data types that were tested in this study was assessed by their classification accuracies. Generally, higher classification accuracy indicates better feature discrimination.
The classification results of the spectral reflectance and the firstderivative
data are shown in Table 1. The classification accuracy for
the spectral reflectance was higher than the classification accuracy of the
first derivative.
 Fig. 9: 
An example of Haar, Daubechies and Symlet coefficient values
obtained from a spectral reflectance sample and its derivative spectrum
using the DWT method and a ninelevel decomposition. (a) Spectra reflectance,
(b) First derivative spectra, (c, d) Haar coeddicients, (e, f) Daubechies
coefficient and (g, h) Symlet coefficient 
Table 1: 
Classification accuracies of reflectance and firstderivative
spectra 

The classification accuracy achieved by using spectral reflectance was 82%
and the classification accuracy of first derivative was only 35.6%. The classification
accuracy drops significantly when using derivativetransformed data. Several
other studies also obtained similar results in which the classification accuracy
of derivative is lower than the classification accuracy of spectral reflectance
(Jones et al., 2010; Li
and He, 2008; Zhang et al., 2006). For this
study, the derivative spectra could not improve the discrimination capability
and had worse discrimination capability than the spectral reflectance. The low
classification accuracy achieved by using the derivative data may be caused
by noise in the spectra because the spectral derivative is very sensitive even
to small variations in the data. This factor was believed to have contributed
to the low overall classification accuracy achieved when using derivative spectra.
The classification accuracy for wavelet coefficients can vary with the selection
of transformation method, mother wavelet, number of vanishing moments and decomposition
level. Some of the mother wavelets gave higher classification accuracies than
the original dataset, while some gave lower classification accuracies. This
difference shows that not all mother wavelets were suitable for feature discrimination.
Koger et al. (2003), Zhang
et al. (2005) and Bruce et al. (2002)
got similar result pattern when classifying using different mother wavelet coefficients
data sets. Koger et al. (2003) studied on detecting
pitted morningglory in soybean using wavelet analysis. The classification accuracies
were varied for most of the mother wavelets. Only several mother wavelets gave
same accuracy. This indicated that classification accuracy is depending on the
selection of mother wavelet. Selecting the number of vanishing moments was also
as critical as the selection of the mother wavelet. This result is shown in
Fig. 10.
Figure 11 shows the maximum accuracy of every level while
using wavelet coefficients derived from reflectance and spectralderivative
data with DWT transformation. In general, the maximum accuracy increased with
the number of decompositions used. The accuracy for onelevel decomposition
was lower than the accuracy at higherlevel decompositions. This result may
be due to the noise occurring at every level. Lower decomposition levels may
contain more noise than the higher decomposition levels.
 Fig. 10: 
Classification accuracies achieved with the Daubechies wavelet
while using different numbers of vanishing moments 
 Fig. 11: 
Maximum classification accuracies achieved from each level
of decomposition while using the DWT wavelet coefficient. This data was
based on the maximum accuracy for each level, independent of the mother
wavelet, the number of vanishing moments and the level of decomposition 
The coefficient values at lower decomposition levels were small and smaller
coefficient values were often assumed to be noise in the data. Besides that,
the number of coefficients at a lower decomposition level was greater than the
number of coefficients at higher decomposition levels. Wavelet coefficients
at lower scale are sensitive to narrow or local spectral features because they
are derived from highpass filters, which is similar to the derivative spectra
(Zhang et al., 2006). Those factors could affect
classification accuracy with regard to the selection of decomposition level.
Figure 11 also suggests that the classification accuracy
while using wavelet coefficients derived from reflectance was higher than the
accuracy of wavelet coefficients derived from derivativespectra.
 Fig. 12: 
Classification accuracies when using the Daubechies wavelet
with different numbers of vanishing moments 
An exception was the threelevel decomposition. This result shows that wavelet
coefficients derived from spectral reflectance are more reliable for discriminating
between features than wavelet coefficients derived from derivative spectra.
Maximum classification accuracy could be achieved by using wavelet coefficients
derived from the reflectance when using a fivelevel decomposition. Wavelet
coefficients derived from derivative spectra only gives maximum accuracy when
using wavelet coefficients with a sevenlevel decomposition.
The same pattern of results was also achieved when using wavelet coefficients
that had been transformed using CWT transformation (Fig. 12).
Lower classification accuracy was achieved when using lower levels of decomposition
and wavelet coefficients derived from reflectance gave better classification
accuracies than wavelet coefficients derived from derivative spectra. The only
difference between DWT and CWT was that the CWT transformation could not produce
wavelet coefficients that provided a maximum classification accuracy of 100%
(as with DWT).
Table 2: 
Classification accuracies of the reflectance, the firstderivative
spectra, the best wavelet coefficients of reflectance and the derivative 

This result showed that the DWT transformation was much better than the CWT
transformation for discriminating features.
The classification accuracy for wavelet coefficients derived from firstderivative spectra did not produce better accuracies than wavelet coefficients derived from the spectral reflectance. This may have been due to noise contamination in the original data, thus resulting in poor classification results. However, there was an improvement in classification accuracy for wavelet coefficients of derivative spectra compared to the classification accuracy of the derivative spectra. This result proved that wavelet transformation could be used as one of the transformation method for improving the output result and achieving a maximum success rate. The overall accuracy of the data types analysed in this study is summarised in Table 2. The spectral reflectance gave better classification results than the derivative spectra. Furthermore, the wavelet coefficients of the spectral reflectance and the derivative spectra may maximise the classification accuracy. The results using Symlet 12 and reflectance data can be considered better as it requires only fivelevel decomposition compared to the use of Haar wavelet with derivative spectra that requires eightlevel decomposition. Decomposition to a higher level would demand more processing time. Thus, Symlet 12 with a fivelevel decomposition transformed from reflectance data by using the DWT transformation was the best wavelet coefficient for discriminating between features.
Although this study has proven that Symlet 12 is the best wavelet coefficients
in discriminating between features, this wavelet will not guarantee to give
the best results when applied to other data sets. Previous studies by Bruce
et al. (2002), Koger et al. (2003)
and Zhang et al. (2005) and have shown differences
in terms of the performance of the best mother wavelet achieved from their studies.
CONCLUSIONS In summary, the selection of data types was essential to achieve the desired output with minimum error. The selection of wavelet transformation method, mother wavelet, number of vanishing moments and level of decomposition also played an important role in achieving better results. This study showed that the firstderivative spectra do not necessarily provide better classification accuracies than the spectral reflectance data. Some of the mother wavelets derived from spectral reflectance or from derivative spectra resulted in lower classification accuracies than the original dataset and some of the mother wavelets produced improvements in the classification results. This study showed that the potential of using wavelettransformed data for discriminating features with a maximum success rate. Better ways of dealing with noise could be investigated to utilise the spectral derivative fully in future studies using this data. ACKNOWLEDGMENTS The authors would like to thank the Ministry of Higher Education Malaysia for the research grant and UPM for providing a graduate scholarship to aid in the completion of this research. Aeroscan Precision (M) Sdn. Bhd. is also acknowledged for providing the test data.

REFERENCES 
1: Ali, S.A., N. Sulaiman, A. Mustapha and N. Mustapha, 2009. Improving accuracy of intentionbased response classification using decision tree. Inform. Technol. J., 8: 923928. CrossRef  Direct Link 
2: Bassani, C., R.M. Cavalli, F. Cavalcante, V. Cuomo and A. Palombo et al., 2007. Deteriotion status of asbestoscement roofing sheets assessed by analyzing hyperspectral data. Remote Sensing Environ., 109: 361378. CrossRef 
3: Blackburn, G.A. and J.G. Ferwerda, 2008. Retrieval of chlorophyll concentration from leaf reflectance spectra using wavelet analysis. Remote Sensing Environ., 112: 16141632. CrossRef 
4: Bruce, L.M., C.H. Koger and J. Li, 2002. Dimensionality reduction of hyperspectral data using discrete wavelet transform feature extraction. IEEE Trans. Geosci. Remote Sensing, 40: 23312338. CrossRef 
5: CastroEsau, K.L., G.A. SanchezAzofeifa and T. Caelli, 2004. Discrimination of lianas and trees with leaflevel hyperspectral data. Remote Sensing Environ., 90: 353372. CrossRef 
6: Chappelle, E.W., M.S. Kim and J.E. McMurtrey, 1992. Ratio Analysis of Reflectance Spectra (RARS): An algorithm for the remote estimation of the concentrations of chlorophyll a, chlorophyll b and carotenoids in soybean leaves. Remote Sens. Environ., 39: 239247. CrossRef  Direct Link 
7: Chi, M., R. Feng and L. Bruzzone, 2008. Classification of hyperspectral remotesensing data with primal SVM for smallsized training dataset problem. Adv. Space Res., 41: 17931799. CrossRef 
8: Clark, M.L., D.A. Roberts and D.B. Clark, 2005. Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales. Remote Sensing Environ., 96: 375398. CrossRef 
9: Dawson, T.P. and P.J. Curran, 1998. A new technique for interpolating the reflectance red edge position. Int. J. Remote Sensing, 19: 21332139. CrossRef 
10: Devadas, R., D.W. Lamb, S. Simpfendorfer and D. Backhouse, 2008. Evaluation ten spectral vegetation indices for identifying rust infection in individual wheat leaves. Precision Agric., 10: 459470. CrossRef 
11: Erbek, S.F., C. Ozkan and M. Taberner, 2004. Comparison of maximum likelihood classification method with supervised artificial neural network algorithms for land use activities. Int. J. Remote Sensing, 25: 17331748. CrossRef 
12: Galford, G.L., J.F. Mustard, J. Melillo, A. Gendin, C.C. Cerri and C.E.P. Cerri, 2008. Wavelet analysis of MODIS time series to detect expansion and intensification of rowcrop agriculture in Brazil. Remote Sensing Environ., 112: 576587. CrossRef 
13: GomezChova, L., G. CampsValls, L. Bruzzone and J. CalpeMaravilla, 2009. Mean map kernel methods for semisupervised cloud classification. IEEE Trans. Geosci. Remote Sensing, 48: 207220. CrossRef 
14: Gong, P., R. Pu and R.C. Heald, 2002. Analysis of In situ hyperspectral data for nutrient estimation of giant sequoia. Int. J. Remote Sensing, 23: 18271850. CrossRef 
15: Gong, P., R. Pu and B. Yu, 2001. Conifer species recognition: Effects of data transformation. Int. J. Remote Sensing, 22: 34713481. CrossRef 
16: Han, L., 2002. Spectral reflectance of Thalassia testudinum with varying depths. Int. Geosci. Remote Sensing Symp., 4: 21232125. CrossRef 
17: Hansen, P.M. and J.K. Schjoerring, 2003. Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalize defferent vegetation indices and partial least squares regression. Remote Sensing Environ., 86: 542553. CrossRef 
18: Heiden. U., K. Segl, S. Roessner and H. Kaufmann, 2007. Determination of robust spectral features for identification of urban surface materials in hyperspectral remote sensing data. Remote Sensing Environ., 111: 537552. CrossRef 
19: Hsu, P.H., 2007. Feature extraction of hyperspectral image using wavelet and matching persuit. ISPRS J. Photogrammetry Remote Sensing, 62: 7892. CrossRef 
20: Huang, C., E.L. Geiger, W.J.D. van Leeuwen and S.E. Marsh, 2009. Discrimination of invaded and native species sites in a semidesert grassland using MODIS multitemporal data. Int. J. Remote Sensing, 30: 897917. CrossRef 
21: Jensen, J.R., 2005. Introductory Digital Image Processing: A Remote Sensing Perspective. 3rd Edn., Prentice Hall, Upper Saddle River, New Jersy, ISBN10: 0132058405
22: Jones, T.G., N.C. Coops and T. Sharma, 2010. Employing groundbased spectroscopy for treespecies differentiation in the Gulf Islands National Park reserve. Int. J. Remote Sensing, 31: 11211127. CrossRef 
23: Keshava, N., 2004. Distance metrics and band selection in hyperspectral processing with applications to material identification and spectral libraries. IEEE Trans. Geosci. Remote Sensing, 42: 15521565. CrossRef 
24: Koger, C.H., L.M. Bruce, D.R. Shaw and K.N. Reddy, 2003. Wavelet analysis of hyperspectral reflectance data for detecting pitted morning glory (Ipomoea lacunose) in soybean (Glycine max). Remote Sensing Environ., 86: 108119. CrossRef 
25: Kuckenberg, J., I. Tartachnyk and G. Noga, 2008. Temporal and spatial changes of chlorophyll fluorescence as a basis for early and precise detection of leaf rust and powdery mildew infections in wheat leaves. Precision Agric., 10: 3444. CrossRef 
26: Laverington, D.W., 2010. Discrimination of sedimentary lithologies using hyperion and landsat thematic mapper data: A case study at Melville Island, Canadian high Artic. Int. J. Remote Sensing, 31: 233260. CrossRef 
27: Li, X. and Y. He, 2008. Discriminating varieties of tea plant based on VisNIR spectral characteristics and using artificial neural networks. Biosyst. Eng., 99: 313321. CrossRef 
28: Loum, G., C.T. Haba, J. Lemoine and P. Provent, 2007. Texture characterisation and classification using full wavelet decomposition. J. Applied Sci., 7: 15661573. CrossRef  Direct Link 
29: Lucas, R., P. Bunting, M. Patersen and L. Chisholm, 2008. Classification of Australian forest communities using aerial photography, CASI and HyMap data. Remote Sensing Environ., 112: 20882103. CrossRef 
30: Murakami, T., 2004. Seasonal variation in classification accuracy of forestcover types by a single band or band combinations. J. For. Res., 9: 211215. CrossRef 
31: Nelson, R.F., 1981. A comparison of two methods for classifying forestland. Int. J. Remote Sensing, 2: 4960. CrossRef 
32: Mutanga, O., A.K. Skidmore and H.H.T. Prins, 2004. Predicting In situ pasture quality in the Kruger National Park, South Africa, using continuumremoved absorption features. Remote Sensing Environ., 89: 393408. CrossRef 
33: Oldeland, J., W. Dorigo, L. Lieckfeld, A. Lucieer and N. Jurgens, 2010. Combining vegetation indices, constrained ordination and fuzzy classification for mapping seminatural vegetation units from hyperspectral imagery. Remote Sensing Environ., 114: 11551166. CrossRef 
34: Ouma, Y.O., J. Tetuko and R. Tateishi, 2008. Analysis of cooccurrence and discrete wavelet transform textures for differentiation of forest and nonforest in veryhighresolution opticalsensor imagery. Int. J. Remote Sensing, 29: 34173456.
35: Phillips, R.D., L.T. Watson, R.H. Wynne and C.E. Blinn, 2009. Feature reduction using a singular value decomposition for the iterative guided spectral class rejection hybrid classifier. ISPRS J. Photogrammetry Remote Sensing, 64: 107116. CrossRef 
36: Ping, Z., T. LiZhen and X. DongFeng, 2009. Speech recognition algorithm of parallel subband HMM based on wavelet analysis and neural network. Inform. Technol. J., 8: 796800. CrossRef  Direct Link 
37: Raju, U.S.N., B.E. Reddy, V.V. Kumar and B. Sujatha, 2008. Texture classification based on extraction of skeleton primitives using wavelets. Inform. Technol. J., 7: 883889. CrossRef  Direct Link 
38: Rao, N.R., P.K. Garg and S.K. Ghosh, 2007. Development of the agricultural crops spectral library and classification of crops at cultivar level using hyperspectral data. Precision Agric., 8: 173185. CrossRef 
39: Serpico, S.B. and G. Moser, 2007. Extraction of spectral channels from hyperspectral images for classification purposes. IEEE Trans. Geosci. Remote Sensing, 45: 484495. CrossRef 
40: Shafri, H.Z.M. and F.S.H. Ramle, 2009. A comparison of support vector machine and decision tree classifications using satellite data of Langkawi Island. Inform. Technol. J., 8: 6470. CrossRef  Direct Link 
41: Tsai, F. and W. Philpot, 1998. Derivative analysis of hyperspectral data. Remote Sensing Environ., 66: 4151. CrossRef 
42: Wang, L. and W.P. Sausa, 2009. Distinguishing mangrove species with laboratory measurements of hyperspectral leaf reflectance. Int. J. Remote Sensing, 30: 12671281. CrossRef 
43: Wilson, M.D., S.L. Ustin and D.M. Rocke, 2004. Classification of contamination in salt marsh plants using hyperspectral reflectance. IEEE Trans. Geosci. Remote Sensing, 42: 10881095. CrossRef 
44: Yang, C., J.H. Everitt and H.B. Johnson, 2009. Applying image transformation and classification techniques to airborne hyperspectral imagery for mapping Ashe juniper infestations. Int. J. Remote Sensing, 30: 27412758. CrossRef 
45: Zhang, J., B. Rivard, A. SanchezAzofeifa and K. CastroEsau, 2006. Intra and interclass spectral variability of tropical tree species at La Selva, Costa Rica: Implications for species identification using HYDICE imagery. Remote Sensing Environ., 105: 129141. CrossRef 
46: Zhang, X., N.H. Younan and C.G. O`Hara, 2005. Wavelet domain statistical hyperspectral soil texture classification. IEEE Trans. Geosci. Remote Sensing, 43: 615618. CrossRef 



