MULTISPECTRAL REMOTE SENSING IMAGE CLASSIFICATION WITH MULTIPLE FEATURES

Similar documents
FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION

Lecture 13: High-dimensional Images

Feature Reduction and Selection

Applying EM Algorithm for Segmentation of Textured Images

Using Fuzzy Logic to Enhance the Large Size Remote Sensing Images

Fuzzy Logic Based RS Image Classification Using Maximum Likelihood and Mahalanobis Distance Classifiers

An Image Fusion Approach Based on Segmentation Region

Detection of an Object by using Principal Component Analysis

A Fast Visual Tracking Algorithm Based on Circle Pixels Matching

A Probabilistic Approach to Detect Urban Regions from Remotely Sensed Images Based on Combination of Local Features

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision

Cluster Analysis of Electrical Behavior

Determining the Optimal Bandwidth Based on Multi-criterion Fusion

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data

An Improved Image Segmentation Algorithm Based on the Otsu Method

Classifier Selection Based on Data Complexity Measures *

CS 534: Computer Vision Model Fitting

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points;

Parallelism for Nested Loops with Non-uniform and Flow Dependences

Modular PCA Face Recognition Based on Weighted Average

Title: A Novel Protocol for Accuracy Assessment in Classification of Very High Resolution Images

IMAGE FUSION TECHNIQUES

Feature Selection for Target Detection in SAR Images

A Binarization Algorithm specialized on Document Images and Photos

Object-Based Techniques for Image Retrieval

IMAGE FUSION BASED ON EXTENSIONS OF INDEPENDENT COMPONENT ANALYSIS

A Shadow Detection Method for Remote Sensing Images Using Affinity Propagation Algorithm

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration

A PATTERN RECOGNITION APPROACH TO IMAGE SEGMENTATION

International Conference on Applied Science and Engineering Innovation (ASEI 2015)

BioTechnology. An Indian Journal FULL PAPER. Trade Science Inc.

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning

A fast algorithm for color image segmentation

Local Quaternary Patterns and Feature Local Quaternary Patterns

The Study of Remote Sensing Image Classification Based on Support Vector Machine

Maximum Variance Combined with Adaptive Genetic Algorithm for Infrared Image Segmentation

SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE

ALEXNET FEATURE EXTRACTION AND MULTI-KERNEL LEARNING FOR OBJECT- ORIENTED CLASSIFICATION

A ROBUST CHANGE DETECTION METHODOLOGY FOR TOPOGRAPHICAL APPLICATIONS. Booth Str. Ottawa, Ontario K1A 0E9 Canada

Support Vector Machine for Remote Sensing image classification

A New Feature of Uniformity of Image Texture Directions Coinciding with the Human Eyes Perception 1

BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION

Image Representation & Visualization Basic Imaging Algorithms Shape Representation and Analysis. outline

S1 Note. Basis functions.

B.N.Jagadesh* et al. /International Journal of Pharmacy & Technology

Unsupervised Learning and Clustering

Unsupervised Learning

Hybridization of Expectation-Maximization and K-Means Algorithms for Better Clustering Performance

Face Recognition University at Buffalo CSE666 Lecture Slides Resources:

Combination of Color and Local Patterns as a Feature Vector for CBIR

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1

The Research of Support Vector Machine in Agricultural Data Classification

Hybrid Non-Blind Color Image Watermarking

Scale Selective Extended Local Binary Pattern For Texture Classification

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization

Fuzzy C-Means Initialized by Fixed Threshold Clustering for Improving Image Retrieval

Human Face Recognition Using Generalized. Kernel Fisher Discriminant

Edge Detection in Noisy Images Using the Support Vector Machines

Keywords - Wep page classification; bag of words model; topic model; hierarchical classification; Support Vector Machines

Lobachevsky State University of Nizhni Novgorod. Polyhedron. Quick Start Guide

Type-2 Fuzzy Non-uniform Rational B-spline Model with Type-2 Fuzzy Data

Recognizing Faces. Outline

PERFORMANCE EVALUATION FOR SCENE MATCHING ALGORITHMS BY SVM

Research and Application of Fingerprint Recognition Based on MATLAB

Steps for Computing the Dissimilarity, Entropy, Herfindahl-Hirschman and. Accessibility (Gravity with Competition) Indices

Learning the Kernel Parameters in Kernel Minimum Distance Classifier

Course Introduction. Algorithm 8/31/2017. COSC 320 Advanced Data Structures and Algorithms. COSC 320 Advanced Data Structures and Algorithms

Fusion Performance Model for Distributed Tracking and Classification

EYE CENTER LOCALIZATION ON A FACIAL IMAGE BASED ON MULTI-BLOCK LOCAL BINARY PATTERNS

Text Similarity Computing Based on LDA Topic Model and Word Co-occurrence

A Background Subtraction for a Vision-based User Interface *

Positive Semi-definite Programming Localization in Wireless Sensor Networks

Solving two-person zero-sum game by Matlab

A New Approach For the Ranking of Fuzzy Sets With Different Heights

A Novel Term_Class Relevance Measure for Text Categorization

Multiclass Object Recognition based on Texture Linear Genetic Programming

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance

Comparison Study of Textural Descriptors for Training Neural Network Classifiers

Brushlet Features for Texture Image Retrieval

A Clustering Algorithm for Key Frame Extraction Based on Density Peak

The Improved K-nearest Neighbor Solder Joints Defect Detection Meiju Liu1, a, Lingyan Li1, b *and Wenbo Guo1, c

A Robust Method for Estimating the Fundamental Matrix

An Entropy-Based Approach to Integrated Information Needs Assessment

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task

Lecture 5: Multilayer Perceptrons

SRBIR: Semantic Region Based Image Retrieval by Extracting the Dominant Region and Semantic Learning

Novel Pattern-based Fingerprint Recognition Technique Using 2D Wavelet Decomposition

Contourlet-Based Image Fusion using Information Measures

Related-Mode Attacks on CTR Encryption Mode

Classification of Face Images Based on Gender using Dimensionality Reduction Techniques and SVM

Hyperspectral Image Classification Based on Local Binary Patterns and PCANet

Appearance-based Statistical Methods for Face Recognition

Combination of Local Multiple Patterns and Exponential Discriminant Analysis for Facial Recognition

TN348: Openlab Module - Colocalization

PRÉSENTATIONS DE PROJETS

Support Vector Machines

An Evaluation of Divide-and-Combine Strategies for Image Categorization by Multi-Class Support Vector Machines

An Anti-Noise Text Categorization Method based on Support Vector Machines *

Transcription:

MULISPECRAL REMOE SESIG IMAGE CLASSIFICAIO WIH MULIPLE FEAURES QIA YI, PIG GUO, Image Processng and Pattern Recognton Laboratory, Bejng ormal Unversty, Bejng 00875, Chna School of Computer Scence and echnology, Bejng Insttute of echnology, Bejng 0008, Chna E-MAIL: ynqan@bnu.edu.cn; pguo@eee.org Abstract: In ths paper, we propose to combne the spectral and texture features to compose the mult-feature vectors for the classfcaton of multspectral remote sensng mage. It usually s dffcult to obtan the hgher classfcaton accuracy f only consders one nd feature, especally for the case of dfferent geographcal objects have the same spectrum or texture specalty for a multspectral remote sensng mage. he spectral feature and the texture feature are composed together to form a new feature vector, whch can represent the most effectve features of the gven remote sensng mage. In ths way we can overcome shortcomngs of only usng the sngle feature and rase the classfcaton accuracy. he system classfcaton performance wth composed feature vector s nvestgated by expermentatons. By analyss of results we can learn how to combne the mult-feature vector can obtan a hgher classfcaton rate, and experments proved that the proposed method s feasble and useful n multspectral remote sensng mage classfcaton study. Keywords: Spectrum feature; exture feature; Multspectral remote sensng mage; Feature combnaton; Classfcaton.. Introducton In the last decades, remote sensng magery utlty has proved a powerful technology for montorng the earth s surface and atmosphere at a global, regonal, and even local scale. he volume of remote sensng mages contnues to grow at an enormous rate due to advances n sensor technology for both hgh spatal and temporal resoluton systems. As a consequence, an ncreasng quantty of multspectral mage acqured n many geographcal areas s avalable. here are many applcatons n analyss and classfcaton of remote sensng mage, such as n geology remote sensng, water area remote sensng, vegetaton remote sensng, sol remote sensng, mult-spectrum remote sensng and so on. In all these applcatons, to recognze the nterestng regons from the multspectral remote sensng mage s the ey processng step. In order to effectve rase the classfcaton accuracy, t s mportant to extract the most effectve features to represent the mage under study. Wth extracted features, a classfer s bult to recognze the nterested objects n remote sensng magery. here are two nds of classfcaton: supervsed and unsupervsed. In general, when we have lttle nowledge about gven mage, we have to adopt unsupervsed classfcaton technques. Among the unsupervsed methods, the fnte mxture model analyss has some advantages and t attracts many researchers nterest n mage segmentaton as well as other applcatons []. When buldng a classfer, we assumed that the data n the feature space as a mxture of Gaussan probablty densty dstrbuton, and the fnte mxture model s used to cluster the extracted features [4]. Spectral feature s regarded as one of the most mportant peces of nformaton for remote sensng mage nterpretaton. hs nd feature can be used to characterze most mportant contents for varous types of the remote sensng mages. It s beleved that the gray value plays an mportant role n the vsual systems for recognton and nterpretaton of the multspectral remote sensng mage data. On the other hand, the texture descrbes the attrbuton between a pxel and the other pxels around t [3]. exture features represent the spatal nformaton of an mage, can be regarded as an mportant vsual prmtve to search vsually smlar patterns n the mage. However, no unversally accepted mathematcal defnton of texture exsts, and texture analyss s even more dffcult n remote sensng mages [0]. Reed and Buf present a detaled survey of varous texture methods for mage analyss []. However, n classfcaton, t has the shortcomngs f only adoptng the texture analyss method, such as the edge between dfferent classes may be ncorrectly classfed, because texture feature extracton must be consdered based on a small regon, not a sngle pxel. Spectral feature such as gray value can be extracted based on a sngle pxel, but t -444-0973-X/07/$5.00 007 IEEE 360

has lmtaton as the representatve nformaton of an mage. owadays, most exstng classfcaton studes for remote sensng mage adopt only smple spectral or texture feature, or nvestgate wth ndependent manner [5][4]. to effcently extract useful features from remote sensng mage and to buldng a classfcaton system wth hgher accuracy become a challenge. In ths paper, we propose to combne the spectral and texture features to compose the mult-feature vectors for the classfcaton of multspectral remote sensng mage. he spectral features and the texture features, whch are extracted n the same mage, are composed to construct a new feature vector n mult-feature space for classfcaton.. Methodology Bascally, for multspectral remote sensng magery, ts data amount s large and maes t become dffcult to extract the man features f only consderng one spectral characterstc; especally for the dfferent geographcal objects have the same spectrum. When ts resoluton s hgh, t has abundant texture nformaton. hs becomes dffcult to dstngush some geographcal objects when nterested area has complcated texture nformaton. Also, the spectral characterstc s uncertanty f the geographcal objects have complcated surroundngs. Sometmes dfferent objects have the same spectrum, or same geographcal objects have dfferent spectrum. Whle texture refers to a pattern, t has propertes of homogenety that does not depend on the presence of only a sngle color or ntensty. Prevous attempts at modelng texture nclude the followng approaches: Marov random feld, co-occurrence matrces and Gabor flters and so on []. he approaches of extractng the features should be guded by the followng concerns: he features should carry enough nformaton about the mage and should nclude doman-specfc nowledge for ther extracton. hey should be easy to compute n order for the approach to be feasble for a large mage collecton and rapd clusterng. hey should relate well wth the remote sensng mage characterstcs snce users wll fnally determne the sutablty of the classfed mages... Spectral Features Spectrum s an mportant characterstc for the analyss of varous types of the remote sensng mages. In the prevous wor, fve dmenson reducton methods, such as the Eucld dstance measurement method (EDM), the dscrete measurement crtera functon method (DMCF), the mnmum dfferentated entropy method (MDE), the probablty dstance crteron method (PDC), and the prncple component analyss method (PCA) are adopted to extract the most avalable features form multspectral mages [9]. he purpose for dmenson reducton s that extracted spectral feature may not be redundancy and the parameters can be estmated well n lower feature space n whch the multspectral remote sensng mage s relatve easy nvestgated. We can suppose the orgnal remote sensng mage has D bands and the processed data has d dmensons after we decrease the bands. And we can defne the orgnal data vector as y = [ y, y,, yd ], the processed data vector as [ ] x = x, x,, xd. And the transformaton formula s: x= W y () A number of bands are reduced nto a sngle one n ths paper, so a multspectral mage can be dsplayed n the gray scale form. he detal descrpton of each dmenson reducton method can be found n reference [9]... exture Features We wll adopt the followng approaches for texture feature modelng problem: the gray level co-occurrence matrx (GLCM) [3], the hstogram measures (HM) [], the texture spectrum (S) [3], and the gray dfference statstcal quantty (GDSQ) [3].... Gray Level Co-occurrence Matrx For the GLCM texture feature, we use the angular second moment (ASM), also called the energy, the dssmltude (DIS), also called nerta, and the nverse dfference moment, and also called homogenety (HOM). In order to ncrease computaton effcent, we ntroduce an mproved method to avod calculatng the zero values n the gray level co-occurrence matrx. he mproved GLCM s constructed by p (, j), and the correspondng measures are: (, ) p j = graynum(, j) ( d, θ ) ( max mn ), j = mn,,max = j= (, ), () ASM = P j, (3) 36

= j= (, ) ( ) HOM = P j + j, (4) (, ). (5) DIS = j P j = j=... Hstogram Measures For the HM feature we adopt the average value (AVE), the standard covarance (SDCOV) and the entropy (E) measures. Suppose the number of pxels n an mage s P, for each gray value, the pxel number s p, the frequency hstogram H, can be expressed as p H ( ) =, = 0,,,55. (6) P It s used to descrbe the probablty densty curve of the gven mage. And those measures are expressed as 55 AVE = H ( ), (7) 56 = SDCOV H AVE H AVE 56.3. exture Spectrum 0 = ( )( ) 55 = 0 ( ), (8) E = H log H. (9) he S feature we adopted are angular second moment (ASM), the average value (AVE) and the standard covarance (SDCOV). he defntons of these measures are as follow [4]. 8 S () (0) ASM = = 8 8 = () AVE = S () 8 8 SDCOV = S S = = () 8 () () where S() s the th value of texture spectrum..4. Gray Dfference Statstcal Quantty For the GDSQ feature, we adopt the followng measures: the contrast, the angular second moment (ASM), the entropy (E) and the average value (AVE). he defntons for these measures are []: where p () ASM p =, (3) E = p p, (4) () lg () AVE = p (), (5) m s the gray dfference statstcs. 3. Fnte Mxture Model Analyss he unsupervsed classfcaton method s adopted because we can get better results n the case where there s a lac of pror nowledge about remote sensng mages. From the detaled descrpton of each dmenson reducton method [9], we now that except the PCA method, all the other methods need to label each pxel n the orgnal multspectral remote sensng mage. But we have lttle nowledge about whch pxel should belong to whch class. In order to resolve ths problem, n ths wor we adopt the random sample method. hat means, we frst assgn a class label for each pxel randomly. he fnte mxture model s adopted to analyze and the Expectaton- Maxmzaton (EM) algorthm [4] s used to estmate the model parameters. Wth the teratve EM algorthm, the mxture parameters can be estmated untl the lelhood L( Θ) functon reaches a local mnmum value. n n ( ) ( ) ( j) L( Θ ) = p x = p x j p n= n= j= α = ). (7) j. (6) he jont probablty dstrbuton can be expressed as ( ) α jg( x m ) p x, Θ =,,, ( α 0, Where (,, j ) G x m = j j= j= ( π ) d j exp ( x mj) j ( x mj) (8) s a general expresson of multvarate Gaussan dstrbuton. x denotes a random vector, d s the dmenson of x, and the parameter { α j, m j, j} j = Θ= s a set of fnte mxture model parameters vectors. Here α j s the mxng 36

weghts, m j s the mean vector, and Σ j s the covarance matrx of the jth component of the mxture model. Actually, these parameters are unnown, and usng how many Gaussan component denstes can best descrbe the jont probablty densty s also unnown. Redner [6] had proved that the lelhood functon was convergent and assured t could be close to a local mnmum value. Under the pre-assgned regon number, the posteror probablty can be descrbed as: P( j = x), P( j = x),, P( j = x ). We use Bayes decson to classfy x nto cluster j*. hs procedure s called Bayesan probablstc classfcaton. 4. Experments and Result Analyss EDM MDE PCA DMCF PDCM Fgure. he sngle band spectral mages obtaned wth dfferent feature extracton methods One multspectral remote sensng mage s adopted n ths paper. he platform for the remote sensng mage s Indan Remote Sensng (IRS) Satellte, whch was launched on n 989 by Inda. he remote sensor s Lnear Imagng Self Scanner (LISS), and the resoluton s 5.8 meters. he mage ncludes the Raas and Manasarovar rvers of Inda n July 000. 4.. Experments In Fgure, the spectrum mages obtaned from one multspectral remote sensng mage wth dfferent feature extracton methods are dsplayed, these mages are used n the experments. For these spectral remote sensng mage, the extracted texture measures are shown n able wth mage format. By analyzng the expermental results, we can now that though ASM measure can better descrbe ths mage s GLCM texture features, SDCOV measure can better descrbe ths mage s hstogram texture features, compared wth AVE measure for GDSQ texture features, these two measures do not good n representng ths mage s texture feature. herefore, n constrcutng the mult-feature vector whch can contan more effectve texture nformaton, we fnally adopt the AVE measure n GDSQ texture features as one component. In prncple, we can construct a hgh dmensonal vector usng as many spectral and textural features as possble. However, t wll become very dffcult to analyze the propertes of combned vector for classfcaton. In ths wor, we only consder a smple case, that s, to construct the mult-feature vector wth one spectral component and one textural component. In the experments, the combnaton of varous spectral features wth the dfferent texture features together s nvestgated ntensvely. he classfcaton results by usng texture feature and dfferent spectral nformaton combnaton are shown n table n mage form representaton. able 3 shows the classfcaton accurate rates for orgnal multspectral remote sensng mage. 4.. Results By analyzng the expermental results n table 3, we can fnd from the results shown n the row, when the same spectral feature extracton method s adopted, for dfferent texture feature, the hstogram measures method can better descrbe the texture characterstcs, texture spectrum feature taes second place n accuracy. From the results shown n the column, we can now when the same texture feature s adopted, for dfferent spectral feature extracton methods, the DMCF method can better descrbe the spectrum characterstc, the EDM method tae the second place n classfcaton accuracy. If the gray dfference statstcal quantty texture feature s adopted, wth the MED, PCA and PDCM spectral feature extracton methods, t cannot get the better classfcaton results. hese measures are not sutable to descrbe the features of the multspectral remote sensng mage. From the expermental results, we also now that 363

for dfferent geographcal objects, the edge of the objects can be better separated by the hstogram measures method. hen we can conclude that wth HM and DMCF combned feature vector, the hgher classfcaton accuracy can be reached. able. the texture features mages wth dfferent measures GLC M HM S GFSQ ASM DIS HOM AVE SDCOV E ASM AVE SDVOV ASM E AVE able. he classfcaton results wth dfferent feature combnaton exture / Spectrum DMCF EDM MDE PCA PDCM 5. Conclusons GLCM HM S GFSQ In ths paper, we propose to construct mult-feature vector to rase the multspectral remote sensng mage classfcaton accuracy. Fve methods are adopted to extract spectral feature, and four methods are used to extract texture features of the mage. Varous combnatons of spectral and textural features are expermental nvestgated n order to fnd the most effectve feature vector to represent the mage. From the results we can conclude that applyng proposed methodology can get the hgher accuracy n classfyng multspectral remote sensng mage. In the future wor, we ntend to combne more features to form a mult-feature vector, we also ntend to apply genetc algorthm to fnd the optmal combnaton of spectral and textural features n order to mprove the performance of mult-source data classfer n hgh dmenson space. able 3. the classfcaton accuracy for combnng dfferent spectral and texture features exture / Spectrum GLCM HM S GDSQ DMCF 94.86% 98.66% 95.86% 95.58% EDM 94.40% 97.60% 95.66% 95.60% MDE 9.84% 95.59% 93.0% 83.% PCA 9.9% 95.9% 9.63% 8.6% PDCM 93.98% 93.90% 95.5% 88.06% Acnowledgements he research wor descrbed n ths paper was fully supported by a grant from the atonal atural Scence Foundaton of Chna (Project o. 606750). he authors would le to than Ms. Yanqn an for her help n part of experment wor. References [] S. Sanjay-Gopal,. J. Hebert, Bayesan pxel classfcaton usng spatally varant fnte mxtures and the generalzed EM algorthm, IEEE rans. Image Processng, Vol. 7, o. 7, pp. 04-08, 998. [] Andrew R. Webb, Statstcal Pattern Recognton, Oxford Unversty Press, London, 999. [3] A. Barald, F. Parmnggan, An Investgaton on the exture Characterstcs Assocated wth Gray Level Co-occurrence Matrx Statstcal Parameters, IEEE ransacton on Geoscences and Remote Sensng, Vol.3, o., pp.93~303, 995. [4] Png Guo, Hanqng Lu, A Study on Bayesan Probablstc Image Automatc Segmentaton, Acta Optca Snca, Vol., o., pp.479~483, 00. (n Chnese) [5] Jang L, and R. M. arayanan, Integrated Spectral and Spatal Informaton Mnng n Remote Sensng 364

Imagery, IEEE rans. on Geoscences and Remote Sensng, Vol. 4, o. 3, pp.673~684, 004 [6] R. A. Redner, and H. F. Waler, Mxture denstes, maxmum lelhood and the EM algorthm, SIAM Revew, Vol. 6, o., pp.95~39, 984. [7] Brandt C. K. so, and Paul M. Mather, Classfcaton of Multsource Remote Sensng Imagery Usng a Genetc Algorthm and Marov Random Felds, IEEE rans. on Geoscences and Remote Sensng, Vol. 37, o.3, pp.55-60, 999. [8] Gunnar Jaob Brem, Jon Atl Benedtsson, and Johannes R. Svensson, Multple Classfers Appled to Multsource Remote Sensng Data, IEEE rans. on Geoscences and Remote Sensng, Vol.40, o.0, pp.9-99, 00. [9] YanQn an, Png Guo, M.R.Lyu. Comparatve Studes on Feature Extracton Methods for Multspectral Remote Sensng Image Classfcaton, Proceedngs of IEEE Internatonal Conference on Systems, Man and Cybernetcs, pp. 75-79, 005. [0] S. Shehar, P. R. Schrater, R. R. Vatsava,W. Wu, and S. Chawla, Spatal contextual classfcaton and predcton models for mnng geospatal data, IEEE ran. Multmeda, vol. 4, pp. 74 88, June 00. []. R. Reed and J. M. H. Buf, A revew of recent texture segmentaton and feature extracton technques, Comput. Vs., Image Process. Graph., vol. 57, no. 3, pp. 359 37, 993. [] Qulng Ruan, Dgtal Image Processng, Electroncs Industry Press, Bejng, pp.49~44, 00. (In Chnese) [3] We Xue, Wedong Gao, and Jhong Lu, Image Cuttng Method n Recognton of Chec exture, Cotton extle echnology, Vol.6, o.6, pp. 33~335, 003. (In Chnese) [4] Ketut Wanta, Whartn, Ryutaro atesh, and Agung Bud Harto, Spectral and extural Informaton of Multsensor Data for Land Use Classfcaton n Metropoltan Area, Proceedngs of IEEE 000 Internatonal Geoscence and Remote Sensng Symposum, pp.843~845, 000. 365