literature review

H s

A a

b

a

A R A

K H S M M

1

s w fi l i g i m s s i o m

0 d

International Journal of Applied Earth Observation and Geoinformation 12 (2010) 261–269

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and Geoinformation

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

yperspectral image classification by a variable interval spectral average and pectral curve matching combined algorithm

. Senthil Kumara,∗, V. Keerthia, A.S. Manjunatha, Harald van der Werffb,1, Freek van der Meerb,1

National Remote Sensing Centre, Data Processing Area, Balanagar, Hyderabad 500 625, India International Institute for Geo-information Science and Earth Observation (ITC), Hengelosestraat 99, 7500AA Enschede, The Netherlands

r t i c l e i n f o

rticle history: eceived 24 June 2009 ccepted 19 March 2010

eywords: yperspectral image pectral curve matching ultiresolution analysis ixed pixel classification

a b s t r a c t

Classification of hyperspectral images has been receiving considerable attention with many new appli- cations reported from commercial and military sectors. Hyperspectral images are composed of a large number of spectral channels, and have the potential to deliver a great deal of information about a remotely sensed scene. However, in addition to high dimensionality, hyperspectral image classification is com- pounded with a coarse ground pixel size of the sensor for want of adequate sensor signal to noise ratio within a fine spectral passband. This makes multiple ground features jointly occupying a single pixel. Spectral mixture analysis typically begins with pixel classification with spectral matching techniques, followed by the use of spectral unmixing algorithms for estimating endmembers abundance values in the pixel. The spectral matching techniques are analogous to supervised pattern recognition approaches, and try to estimate some similarity between spectral signatures of the pixel and reference target. In this paper, we propose a spectral matching approach by combining two schemes—variable interval spectral average (VISA) method and spectral curve matching (SCM) method. The VISA method helps to detect

transient spectral features at different scales of spectral windows, while the SCM method finds a match between these features of the pixel and one of library spectra by least square fitting. Here we also compare the performance of the combined algorithm with other spectral matching techniques using a simulated and the AVIRIS hyperspectral data sets. Our results indicate that the proposed combination technique exhibits a stronger performance over the other methods in the classification of both the pure and mixed

ly.

class pixels simultaneous

. Introduction

Thanks to many operational airborne and spaceborne hyper- pectral instruments onboard, tremendous interest has been itnessed in the utilization of hyperspectral imaging for many elds of research, including agriculture, geology, military surveil-

ance, etc. Unlike conventional panchromatic or even multispectral maging systems that gather broadband reflectance from the round, HSI employs an imaging spectrometer to detect and extract nformation in numerous spectral bands. A great deal of infor-

ation captured from hyperspectral instruments allows in-depth pectral analysis of a target not possible with conventional multi-

pectral images. However, classification of a hyperspectral image s quite a challenging task due to a very high dimensionality in rder to account these bands data. In addition, almost all pixels are ixed class type due to the fact that ground pixel size of the sen-

∗ Corresponding author. Tel.: +91 40 2388 4447; fax: +91 40 2387 9307. E-mail address: senthilkumar [email protected] (A. Senthil Kumar).

1 Tel.: +31 53 4874228; fax: +31 53 4874335.

303-2434/$ – see front matter © 2010 Elsevier B.V. All rights reserved. oi:10.1016/j.jag.2010.03.004

© 2010 Elsevier B.V. All rights reserved.

sor is coarser than that of panchromatic or multispectral systems in order to meet adequate sensor signal to noise ratio within a fine spectral passband. This invariably makes multiple ground features jointly occupying a single pixel (Plaza et al., 2005). The HSI classi- fication is best viewed from a high dimensional signal processing paradigm due to the ‘curse of dimensionality” (Landgrebe, 2002; Shaw and Manolakis, 2002). Many new schemes have been pro- posed in recent years to classify the hyperspectral images (Guo et al., 2008; Wang et al., 2007; Jia and Richards, 1999; Morgan et al., 2002).

Spectral mixture analysis is an analytically intensive exercise, and is commonly achieved with two steps. In the first step, the hyperspectral classification is attempted with the application of spectral matching techniques on HSI image. In the second step, pixel purity analysis is carried with spectral unmixing algorithms to estimate endmembers abundance values of each pixel (Chang,

2003). In this paper, we are concerned only with the spectral matching techniques. These techniques belong to supervised pat- tern recognition approaches, and aim to estimate some similarity between spectral signatures of the pixel and reference target. The similarity measures can be broadly put under two categories:

2 d Earth

s e p t a b d 2 n ( a 2

t t p o ( ( h e a s e e a m

b i t t a a r 4

F (

62 A. Senthil Kumar et al. / International Journal of Applie

tochastic measures and deterministic measures. In the first cat- gory, one defines some statistical parameters such as divergence, robability, entropy, etc. as spectral information criteria, while in he deterministic category, deterministic criteria such as spectral ngle, the distance, correlation, etc. are defined to find a match etween an unknown pixel and a reference spectrum from a stan- ard spectral library (Liu and Chang, 2006; Van der Meer et al., 003). A variety of new classification approaches based on neural etworks, support vector machines have been recently attempted Subramanian et al., 1997; Valls et al., 2006; Watanachaturaporn nd Arora, 2004; Homayouni and Roux, 2004; Van der Werff et al., 007).

In this paper, we propose a spectral matching technique which akes the advantages of both multiresolution analysis and spec- ral correlation fitting in a least square sense to achieve both the ure and mixed pixels classification. This is done with combination f two schemes—variable interval spectral average (VISA) method Kumar et al., 2006) and spectral curve matching (SCM) method Van der Meer, 2000). More specifically, the VISA method would elp to detect ‘transient’ signal variations of spectral data at differ- nt scales of windows, while the SCM method would help to find match between these transient features of the pixel and one of

pectra from standard spectral library by least square fitting. Our xperiment with both simulated and AVIRIS hyperspectral data xhibits a strong performance when the transient VISA features re combined with the SCM method when compared to individual ethods when applied independently. The rest of the paper is organized as follows: Section 2 describes

riefly the methodology of the proposed hybrid method with a brief ntroduction to its constituent methods. Section 3 discusses about he application of these methods and general spectral matching echniques for classifying both a simulated hyperspectral image

nd AVIRIS hyperspectral dataset collected over the predominantly gricultural area of the Indiana State, USA. Section 3 includes esults and discussion and our conclusions are given in Section .

ig. 1. (a) A sample spectral profile of the Kaolinite mineral. The normalized VISA outpu =0.15 �m), respectively.

Observation and Geoinformation 12 (2010) 261–269

2. Methodology

The basic elements of the VISA and SCM are briefly described here. The input to the proposed approach is spectral reflectance of the HSI pixel and reference targets. Necessary radiometric pro- cedures must be applied to estimate from count-to-radiance and radiance to spectral reflectance. These are generally provided by the data distribution agency.

2.1. The VISA algorithm

The VISA method is a variant of time average method proposed for locating coherent packets in a turbid liquid flow (Hudgins and Kaspersen, 1999; Kumar et al., 2006). In the VISA method, we basi- cally compute significant variations, if any, present in the signal by estimating ‘local’ or ‘short-interval’ variances. The short-interval variance is defined by

f̃ (�) = f̂ 2(�) − [f̂ (�)]2 (1) Here f̂ (�) is the average energy of the spectrum within the interval �, estimated from integral

f̂ (�) = 1 �

∫ �+�/2 �−�/2

f (x)dx window : � > 0 (2)

The application of (2) results in a number of local maxima, say, M. The significant maxima are treated as “signatures” by setting a threshold � to suppress unwanted background from entering into the analysis. Mathematically, the VISA output can be written as

Dvisa(�) = {� ∈ M : f̂ (�) ≥ �} (3)

By changing the spectral interval size �, it is possible to detect

the transient features of the signal at different scales of resolu- tion, much similar to multiresolution analysis (MRA) by wavelets of Mallat (1989). However, there is no sub-sampling of the signal at different scales involved in the VISA, thus making interpretation at

t plots for spectral window sizes of (b) 5 (=0.05 �m), (c) 9 (=0.09 �m) and (d) 15

d Earth Observation and Geoinformation 12 (2010) 261–269 263

d t r c

V

s a a l t F 1 a a p e t

e f u p l m w ‘ t

2

d a s i s o

b

w S f

F

a d

F

Fig. 3. (a) Spectral curve plots of A, K and their mixtures. (b) SCM fit function plots for A, K and A + K spectra for the VISA output of a mixed pixel case.

Table 1 Simulated dataset: individual class accuracies from three methods.

Class name Classification accuracy (%)

VISA SCM Hybrid

Alunite (A) 88.39 94.03 97.21 Dolomite (D) 97.93 98.84 98.07 Illite (I) 98.67 97.36 99.84 Kaolinite (K) 91.49 0.00 98.52 Mixed D and I 26.74 72.00 85.25 Mixed A and I 53.66 61.09 85.73 Mixed A and K 83.29 41.59 96.62

A. Senthil Kumar et al. / International Journal of Applie

ifferent scales easier than the MRA. For spectral analysis, we need o compare the image spectra at a pixel with that of targeted mate- ials in the standard spectral library (SSL). This can be achieved by omputing co-variance between these two spectra:

OL(�) = Ô(�)L(�) − Ô(�)L̂(�) (4) The variables O(�) and L(�) represent image pixel and SSL data

pectra, respectively. The peak detection in this case is from a bivari- te set by a simple extension of the univariate in (3). The VISA was pplied on the Kaolinite mineral spectrum taken from JPL spectral ibrary as shown in Fig. 1. This mineral has a unique broad absorp- ion band centered around 2.2 �m. Results of the VISA are shown in ig. 1(b–d) for the selected spectral wavelength windows of 5, 9 and 5. While the two spikes are distinctly centered around 2.14 �m nd 2.22 �m at window size 5, they appear to merge and appear s a broad absorption band at spectral window of 0.15 �m. Appro- riate threshold (�) is to be estimated; it was found sufficient to stimate the background signal outside the absorption band for he hyperspectral data (shown as dashed line in Fig. 1(b–d)).

For the classification purpose, firstly, both sum-square differ- nce of peak wavelengths (SSDP) and sum-square difference of ull width half-maximum (SSDF) between the reference pixel val- es and the HSI pixel are computed. Secondly, treating these two arameters as independent measures, the pixel is assigned to a class

abel independently for which each of these values were at its mini- um. Finally, the pixel would then be assigned to that class label for hich both the class labels are same; or else, the pixel is declared as

unclassified’, thus giving both these measures equal bias in judging he final decision (Kumar et al., 2006).

.2. The SCM algorithm

The spectral curve-matching algorithm basically estimates the egree of similarity between the two spectral data by means of nalytic matching the complete shape of image spectra to library pectra within a certain wavelength of range. A least square fit (LSF) s calculated between the image and each of reference members’ pectra separately. The RMSE of this fit yields on overall goodness f fit measure, after computing correlation coefficients as

= ∑

OcLc − ( ∑

Oc ∑

Lc)/n∑ L2c − (

∑ Lc)

2 /n

= ∑

OcLc − ( ∑

Oc ∑

Lc)/n∑ O

2 c − (

∑ Oc)

2 /n

(5)

here Lc and Oc represent continuum removed spectra from the SL and the image pixel, respectively. Finally, fit value is computed rom (5) as

= (bb′)1/2 (6) The reference spectrum which has the highest fit value is

ssigned unbiased to the image pixel independent of its absorption epth (Van der Meer, 2000).

ig. 2. Simulated image and selected region of interests for classification analysis.

Avg. %acc. 77.17 66.42 94.46 Overall acc. (%) 98.21 88.57 99.71 Kappa coefficient 0.9694 0.8082 0.9951

Table 2 Comparative classification results of conventional methods.

Class name Classification accuracy (%)

SAM SDS SCS SSV

Alunite (A) 88.15 87.86 82.70 97.21 Dolomite (D) 95.81 87.18 94.24 99.53 Illite (I) 87.42 94.25 78.11 98.14 Kaolinite (K) 90.28 90.28 99.41 95.87 Mixed D and I 58.63 28.00 17.83 78.74 Mixed A and I 61.18 61.90 52.42 98.50 Mixed A and K 61.85 61.35 69.33 60.17

Avg. %acc. 77.62 72.97 70.58 89.74 Overall acc. (%) 65.06 60.42 63.26 62.24 Kappa coefficient 0.5697 0.5101 0.5446 0.5360

264 A. Senthil Kumar et al. / International Journal of Applied Earth Observation and Geoinformation 12 (2010) 261–269

Fig. 4. AVIRIS image data and their ground truth map with class labels.

Table 3 Class confusion matrix for the VISA classification scheme.

Class A B GP GM GT C CN CM SC SN SM O HW SS WT WD UN Total

A 17 0 0 0 0 0 1 1 0 0 0 0 3 0 0 0 3 25 B 5 168 10 1 2 15 2 16 10 3 6 4 1 0 0 20 23 286 GP 0 0 62 0 0 0 1 2 0 0 0 0 0 0 0 0 23 88 GM 0 0 0 17 0 0 0 0 0 0 0 0 0 0 0 0 3 20 GT 0 3 0 0 180 0 0 0 0 0 0 0 0 0 0 0 12 195 C 0 0 0 0 0 81 3 0 0 1 4 0 0 0 0 0 21 110 CN 0 0 0 0 0 0 130 3 0 0 0 0 0 0 0 0 27 160 CM 0 0 0 0 0 0 0 128 0 0 0 0 0 0 0 0 70 198 SC 0 0 0 0 0 0 0 0 73 0 7 0 0 0 0 0 16 96 SN 0 0 0 0 0 2 4 0 11 235 9 0 0 0 0 0 39 300 SM 0 0 0 0 0 0 47 0 28 108 318 0 0 0 0 0 59 560 O 0 0 0 0 0 0 0 0 0 0 0 12 0 0 0 0 10 22 HW 0 0 0 0 0 0 0 0 0 0 0 0 268 0 0 0 55 323 SS 0 0 0 0 0 0 0 0 0 0 1 0 0 56 0 0 6 63 WT 0 1 0 0 0 0 1 0 0 7 4 0 0 0 129 0 26 168 WD 0 1 4 0 6 1 0 0 0 0 0 0 0 0 0 427 21 460 Average accuracy = 75.02% Overall accuracy = 74.85% Kappa coefficient = 0.75

Table 4 Class confusion matrix for the SCM classification scheme.

Class A B GP GM GT C CN CM SC SN SM O HW SS WT WD Total

A 24 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 25 B 4 73 10 0 38 8 0 1 0 0 0 5 0 16 6 125 286 GP 43 0 17 6 0 2 3 1 1 7 0 0 0 5 3 0 88 GM 0 0 0 14 0 0 6 0 0 0 0 0 0 0 0 0 20 GT 0 0 3 0 189 0 0 0 0 0 0 1 0 0 0 2 195 C 0 0 0 0 0 44 50 1 0 4 4 7 0 0 0 0 110 CN 0 0 0 0 0 0 146 1 13 0 0 0 0 0 0 0 160 CM 0 0 0 0 0 8 6 108 53 17 6 0 0 0 0 0 198 SC 0 0 0 0 0 2 0 6 71 17 0 0 0 0 0 0 96 SN 0 0 0 0 0 13 1 1 6 261 18 0 0 0 0 0 300 SM 0 1 0 0 0 13 20 14 90 228 185 0 0 0 1 8 560 O 0 3 1 0 5 0 0 0 0 0 0 12 1 0 0 0 22 HW 142 0 0 125 0 0 0 0 0 0 0 0 56 0 0 0 323 SS 0 0 0 0 0 2 2 6 8 1 0 0 0 44 0 0 63

WT 0 0 0 0 0 0 0 0 WD 0 0 0 0 0 0 0 0 Average accuracy = 64.26% Overall accuracy = 6

0 1 0 0 0 1 166 0 168 0 0 0 0 0 0 0 460 460 0.83% Kappa coefficient = 0.663

A. Senthil Kumar et al. / International Journal of Applied Earth Observation and Geoinformation 12 (2010) 261–269 265

Table 5 Confusion matrix for proposed classification scheme.

Class A B GP GM GT C CN CM SC SN SM O HW SS WT WD Total

A 18 0 7 0 0 0 0 0 0 0 0 0 0 0 0 0 25 B 0 189 1 0 19 0 0 0 0 0 0 3 0 0 0 74 286 GP 0 0 77 1 4 0 3 1 0 2 0 0 0 0 0 0 88 GM 0 0 0 20 0 0 0 0 0 0 0 0 0 0 0 0 20 GT 0 0 0 0 194 0 0 0 0 0 0 0 0 0 0 1 195 C 0 0 0 0 0 85 0 3 3 7 12 0 0 0 0 0 110 CN 0 0 0 0 0 0 141 0 17 0 2 0 0 0 0 0 160 CM 0 0 0 0 0 0 4 126 50 4 14 0 0 0 0 0 198 SC 0 0 0 0 0 0 0 3 53 9 31 0 0 0 0 0 96 SN 0 0 0 0 0 0 5 30 1 236 28 0 0 0 0 0 300 SM 0 0 0 2 7 0 21 0 33 47 450 0 0 0 0 0 560 O 0 0 0 0 5 0 0 0 0 0 0 17 0 0 0 0 22 HW 4 0 0 28 0 0 0 0 0 0 0 0 291 0 0 0 323

cy = 83

2

f e a r t

t a r c s V v h L i

3

s i o T

3

t a t t t c s v p o

i m a v u

mixed pixels were to be considered as separate classes. Evidently, some of the mixed pixels were grouped to either as unclassified or wrongly with other classes. In contrast, the hybrid method has provided the best classification results.

Table 6 AVIRIS 92AV3C dataset: individual class accuracies for the proposed (hybrid) scheme as against its constituents methods.

Class name Accuracy (%)

VISA SCM Hybrid

Alfalfa 68.00 96.00 72.00 Corn-notill 81.25 91.25 88.13 Corn-min 64.65 54.55 63.64 Corn 73.64 40.00 77.27 Grass/pasture 70.45 19.32 87.5 Grass/trees 92.31 96.92 99.49 Grass/pasture-mowed 85.00 70.00 100 Hay-windrowed 82.97 17.34 90.09 Oats 54.55 54.55 77.27 Soybeans-notil 78.33 87.00 78.67 Soybeans-min 56.79 33.04 80.36 Soybeans-clean 76.04 73.96 55.21 Wheat 76.79 98.81 85.12 Woods 92.83 100 100

SS 0 0 0 0 0 0 0 0 WT 0 0 0 0 0 0 0 0 WD 0 0 0 0 0 0 0 0 Average accuracy = 82.25% Overall accura

.3. The hybrid (VISA + SCM) scheme

In the hybrid algorithm, we first extract the multi-scale salient eatures of HSI pixel resulting from the VISA scheme. Second, we mploy shape functions to fit with these features at each level in least mean square error sense with the spectral features for the

eference materials. Finally, we select that target material which he fit value has the highest value.

To carry out the above steps, a subset of probable materials spec- ra of relevance is selected from the SSL. Let these be denoted as Li, nd short interval self-variances (VLL) of each Li are computed by eplacing O with Li in Eq. (4). For each HSI pixel O(i, j), short interval ross-variances (VOL) between O with Li are computed. In second tep, correlation coefficients are computed from Eq. (5) by setting LL and VOL equal to the variables Oc and Lc, respectively. The fit alue F is then computed for each Li with input, and one with the ighest fit value is assigned to a class label corresponding to that i spectrum. These two steps are repeated for all pixels in the HSI mage to complete the classification.

. Results and discussions

To test the performance of the proposed method, two hyper- pectral datasets were considered. One was a laboratory simulated mage and the other was AVIRIS sensor 92AC3C data cube collected ver a test site called Indian Pine in north-western Indiana, USA. his dataset is downloadable from the webpage (AVIRIS, 1992).

.1. Results with simulated data

The simulated image was prepared with four mineral spectra aken from spectral library: dolomite (hereafter “D”), illite (here- fter “I”), alunite (hereafter “A”), and kaolinite (hereafter “K”). The otal size of the image is 200 samples × 200 scan lines with 19 spec- ral channels covering wavelength 2.009–2.325 �m. Fig. 2 shows he simulated image and regions of interest (ROIs) selected for lassification performance evaluation. The simulated image was ynthesized to facilitate the classification performance analysis of arious schemes both for unmixed and mixed minerals as inde- endent class types fairly very accurately without any hindrance f sensor and/or atmospheric related issues.

The result of a single image pixel of mixed class type is graph-

cally illustrated in Fig. 3 to show the capability of the proposed

ethod. The spectral curves of A, K from the spectral library data nd their combine (50% A + 50% K) obtained by averaging of two alues at each wavelength are shown in Fig. 3(a). The SSDF val- es obtained after applying the VISA method were for 0.00023,

0 2 1 0 0 60 0 0 63 0 0 0 0 0 3 143 22 168 0 0 0 0 0 0 0 460 460 .28% Kappa coefficient = 0.814

0.0029 and 0.067, respectively for the A, K and A + K, suggesting that the mixed pixel would be wrongly classified as the alunite by the VISA scheme when applied independently. The VISA applica- tion on the A + K spectrum for window size of 7 is shown in Fig. 3(b) and denoted here as ‘signal’. Superposed on this curve are the least square fitted line curves corresponding to VISA outputs of the A and K minerals’ spectra. Large deviations of these curves from the actual the signal curve are quite obvious (respective F values were found to be 0.92 and 0.97). The combination of the VISA and SCM can be seen to be a good representation of the output with the fit value F is as high as 0.996.

Table 1 summarizes the results obtained from the three methods when applied on the simulated image data. The regions of interest (ROI) of both pure and mixed minerals pixels were subjected to these classification schemes. It was evident that both the VISA and SCM methods performed equally well for identifying the pure pix- els. The exception was the SCM failing to recognize the K samples in the presence of its mixtures and other minerals chosen. On the other hand, their ability to discriminate the mixed minerals pix- els was only moderate for many classes. As mentioned earlier, the

Bldg–grass–tree–driv 58.74 25.52 66.08 Stone–steel towers 88.89 69.84 95.24

Avg. %acc. 75.08 64.26 82.25 Overall acc. 74.85 60.83 83.28 Kappa coefficient 0.75 0.663 0.814

266 A. Senthil Kumar et al. / International Journal of Applied Earth Observation and Geoinformation 12 (2010) 261–269

s m t ( R t a I e

F f

Fig. 5. Spectral signatures for corn, soybeans and their sub-classes.

To compare the classification performances of the proposed cheme with existing approaches, we have employed spectral angle apper (SAM), spectral distance similarity (SDS), spectral correla-

ion similarity (SCS) and spectral similarity value (SSV) techniques Homayouni and Roux, 2004), and applied over the same datasets.

esults from these methods are summarized in Table 2. Comparing he average percentage accuracies, the VISA and the SAM provide lmost similar results. However, as far as the mixed pixel (D and ) class, barring the SCS, the VISA method has yielded the poor- st classification accuracy for a mixed class (D and I) case. Also, it

ig. 7. Illustration of the proposed algorithm on soya-min pixel. The solid curves show the rom the SCM. F is the fit value obtained at each resolution.

Fig. 6. Correlation values of land cover classes in AVIRIS data with respect to soya- min (SM) class.

should be noted that for both the mixed and unmixed class pixels, the average percentage accuracy with the proposed method was the highest value. This has in turn resulted in high overall accu- racy and kappa coefficient values for the proposed method when compared to those values from the other methods.

3.2. Results with AVIRIS image data

The AVIRIS data set, 92AC3C, is a useful dataset extensively explored for testing many classification techniques applicable to

output from the VISA for different window sizes and dashed curves the LSF function

A. Senthil Kumar et al. / International Journal of Applied Earth Observation and Geoinformation 12 (2010) 261–269 267

el un

p o 2 a f t m

Fig. 8. The SCM classification scheme for sample pix

redominantly agricultural area covering many crops and a variety f their subclasses. The data was collected in June 1992 and with

20 spectral channels covering a spectral range between 0.4 �m nd 2.5 �m. Details on these spectral channels can be obtained rom website (AVIRIS, 1992). Since the present aim of the study is o judge relative classification performance analysis of the hybrid

ethod and its constituents, some from 220 channels have been

Fig. 9. The VISA profiles for SM, C

der study. Here LSF is done at actual spectral curve.

removed for brevity. The water absorption channels (numbered 104–108, 150–163 and 220) which have least signals and meant

for atmospheric correction, and those channels (1–3, 103, 109–112, 148–149, 164–165 and 217–219) which are plagued with noise were excluded for spectral analysis here. Thus, the final set con- tained a total of 185 remaining spectral channel data. The image and its wall-to-wall ground truth map are shown in Fig. 4.

, CM and SN classes spectra.

2 d Earth

r W 1 t s f a z T T e d t a w o c d

a w v c b

F C

68 A. Senthil Kumar et al. / International Journal of Applie

The scene digital numbers (DNs) were first converted to adiance (L) by using the expression L = (DN − 1000)/500 (in

cm−2 nm−1 sr−1) as per information found in website (AVIRIS, 992). With no other atmospheric characteristics available at the ime of acquisition, apparent spectral reflectance model was con- idered appropriate. The apparent reflectance (�) can be derived rom the equation � = (L�)/(E0(�) cos �), where E0 is the exo- tmospheric solar irradiance at band center �c and �, the solar enith angle, following the method proposed by Chavez (1996). here are about 16 classes identified and available with the dataset. he reference spectra were obtained from each of these classes by xtracting visually identified uniform areas (typically, 2 × 2 win- ow size). The regions of interest (ROIs) are selected for each class o study the performance of the proposed method. Window sizes nd number of peaks for the VISA scheme are class-specific and ere selected accordingly and the SCM fits were applied to each

f these before taking average as the final parameter for class allo- ation. Tables 3–5 give classification performances of the schemes iscussed in Section 2.

From Tables 3–5, considerable increase in both overall and aver- ge classification accuracies is noted with the proposed method

hen compared to the component methods when applied indi-

idually. The proposed method has also resulted in a high kappa oefficient by about 18.5% when compared to the SCM. It should e noted that the VISA classification scheme has resulted in about

ig. 10. Illustration of the proposed algorithm on SM representative pixel. The solid curv , CM and SN.

Observation and Geoinformation 12 (2010) 261–269

13.4% pixels as unclassified due to the classification logic described in Section 2.1. Table 6 provides the class-wise performance results for all the classes. The crops corn and soybeans and their subclasses have a strong overlapping of their spectral signatures, as shown in Fig. 5. The digital correlation of land cover classes, for example, with soybeans-min (SM) class, taken here as reference, is shown in Fig. 6. As can be noted from Fig. 6, this class not only has high correlation with its subclasses of soybeans (SN, SC), but also with corn and its subclasses as well. The classification of these classes by only spectral signatures is a tedious task. Similarly, grass com- bined with pastures, pastures-mowed and trees can be shown to be another set of the spectrally confused classes.

Fig. 7 illustrates graphically the performance of the proposed method at multiple resolutions. The VISA and corresponding LSF profiles were shown for different window sizes. Fit value of each resolution was computed and the average of these was considered for final classification. The average of fit values was equal to 0.9565. This result can be compared with the SCM approach whose output is shown in Fig. 8. The mismatch between the reference and soya- min pixels spectra was obvious and led to a poor fit value of 0.7909, and was wrongly classified as the soya-clean class for which the fit

value was about 0.9322.

Figs. 9 and 10 illustrate graphically how the proposed method would perform classification on the spectrally confused classes. Again, the SM pixel was considered for illustration. Fig. 9 gives the

es show the VISA profile of SM and dashed curves are LSF function from the SCM of

d Earth

V n a t A w t h p f f

4

i e a b s m l p o m

A

J t L e f w

R

A

C

Wang, L., Jia, X., Zhang, Y., 2007. A novel geometry based feature selection technique

A. Senthil Kumar et al. / International Journal of Applie

ISA outputs (at window size of 5) of SM, C, CM and SN. It can be oted that though peaks appear almost at similar locations, there re minute differences in terms of spikes magnitudes. Fig. 10 shows he results of combining the VISA outputs with the LSF function. s can be seen, the auto-correlation has the highest value for SM, hile the LSF curves show mismatching at many wavelength loca-

ions between the SM and other classes. For e.g., the CM and SM ave different matching at 2.0 �m. Thus, the proposed method is roven superior to either the VISA or the SCM, especially for con- used and mixed pixels case when they are applied independently or hyperspectral image classification.

. Conclusions

The proposed hybrid scheme demonstrates the performance of ts spectral matching capabilities in comparable or in most cases xceeding the performances of its existing schemes. This can be ttributed to the fact that the hybrid scheme takes advantages of oth of its constituents in that it facilities multiresolution analy- is of the spectral signatures through the variable interval average ethod and facilitates digital correlation of these signatures in a

east square fitting model. Experimental results indicate that the roposed combination technique exhibits a stronger performance ver the other methods in the classification of both the pure and ixed class pixels simultaneously.

cknowledgements

The authors would like to acknowledge with thanks to Dr. V. ayaraman, Director, NRSC for his special interest and support to his research. Financial support provided by the Physical Research aboratory under the PLANEX project is also sincerely acknowl- dged with thanks. They also wish to thank anonymous reviewers or their valuable and constructive comments and suggestions, hich enhanced this paper greatly.

eferences

VIRIS, N.W. Indiana’s Indian Pines, 1992. 92AV3C hyperspectral dataset

downloadable from: ftp://ftp.ecn.purde.edu/biehl/MultiSpec/92AV3C (orig- inal image) and ftp://ftp.ecn.purde.edu/biehl/PC MultiSpec/ThyFiles.zip (ground truth). The spectral channel characteristics can be found in: http://cobweb.ecn.purdue.edu/∼biehl/MultiSpec/aviris documentation.html.

hang, C.I., 2003. Hyperspectral Imaging: Techniques for Spectral Detection and Classification. Kluwer Academic Publishers.

Observation and Geoinformation 12 (2010) 261–269 269

Chavez, P.S., 1996. Image based atmospheric corrections-revisited and improved. Photogrammet. Eng. Remote Sens. 62, 1025–1036.

Guo, B., Gun, S.R., Damper, R.I., Nelson, J.D.B., 2008. Customizing kernel function for SVM-based hyperspectral image classification. IEEE Trans. Image Process. 17, 622–629.

Homayouni, S., Roux, M., 2004. Material Mapping from Hyperspectral Images using Spectral Matching in Urban Area, Paper downloadable from webpage: http://www.isprs.org/new/congresses/istanbul2004/comm7/papers/10.pdf.

Hudgins, L., Kaspersen, J.H., 1999. Wavelets and detection of coherent structures in fluid turbulence. In: Van der Berg, J.C. (Ed.), Wavelets in Physics. Cambridge University Press, Cambridge.

Jia, X., Richards, J.A., 1999. Segmented principal components transformation for effi- cient hyperspectral remote sensing image display and classification. IEEE Trans. Geosci. Remote Sens. 37, 538–542.

Kumar, A.S., Jayabharathi, S., Manjunath, A.S., Rao, K.M.M., 2006. Fast searching of spectral library database using variable interval spectral average method. Proc. SPIE 6405, 64051. In: Multispectral, Hyperspectral and Ultraspectral Remote Sensing Technology and Applications.

Landgrebe, D., 2002. Hyperspectral image data analysis. IEEE Signal Process. Mag. 19, 17–28.

Liu, W., Chang, C.I., 2006. Sample spectral correlation based measures for sub- pixels and mixed pixels in hyperspectral imagery. Proc. SPIE 6233, 62330. In: Algorithms and Techniques for Multispectral, Hyperspectral and Ultraspectral Imagery.

Mallat, S., 1989. A theory of multiresolution signal decomposition: the wavelet rep- resentation. IEEE Trans. Pattern Anal. Mach. Intell. 2, 674–693.

Morgan, M.M., Henneguelle, A., Crawford, M.M., Ghosh, J., Neuenschwander, A.L., 2002. Best bases Bayesian hierarchical classifier for hyperspectral data analy- sis, Proc. International Geoscience and Remote Sensing Symposium, Toronto, Canada, June 24–28, 1434–1437.

Plaza, J., Chang, C.I., Plaza, A., Perez, R., Matinez, P., 2005. On the generation of training samples for neural network based mixed pixel classification. Proc. SPIE 5806, 139. In: Algorithms and Techniques for Multispectral, Hyperspectral and Ultraspectral Imagery.

Shaw, G., Manolakis, D., 2002. Signal processing for hyperspectral image exploita- tion. IEEE Signal Process. Mag. 19, 12–16.

Subramanian, S., Gat, N., Sheffield, M., Barhen, J., Toomariarr, N., 1997. Methodology for hyperspectral image classification using novel neural network. Proc. SPIE, 3071. In: Algorithms for Multispectral and Hyperspectral Imagery III.

Valls, G.C., Chova, L.G., Mari, J.M., Frances, J.V., Maravilla, J.C., 2006. Composite kernels for hyperspectral image classification. IEEE Geosci. Remote Sens. Lett. 3, 93–97.

Van der Meer, F., De Jong, S.M., Bakker, W., 2003. Imaging spectrometry: basic ana- lytical techniques. In: Imaging Spectrometry: Basic Principles and Prospective Applications. Kluwer Academic Publishers, The Netherlands.

Van der Meer, F., 2000. Spectral curve shape matching with continuum removed CCSM algorithm. Int. J. Remote Sens. 21 (16), 3179–3185.

Van der Werff, H., Van Ruitenbeek, F., Van der Meijde, M., Van der Meer, F., De Jong, S., Kalubandara, S., 2007. Rotation variant template matching for supervised hyperspectral boundary detection. IEEE Geosci. Remote Sens. Lett. 4 (1), 70–74.

for hyperspectral imagery. IEEE Geosci. Remote Sens. Lett. 4, 171–175. Watanachaturaporn, P., Arora, M.K., 2004. Support vector machines for classifica-

tion of multi- and hyperspectral data. In: Varshney, P.K., Arora, M.K. (Eds.), Advanced Image Processing Techniques for Remotely Sensed Hyperspectral Data. Springer-Verlag.