Land-cover classification using SAR and MS image with ANN classifiers

Transcription

1 Land-cover classificaion using SAR and MS image wih ANN classifiers S N Omkar 1, Adii Kanjolia 2, Sandesh C 3 and Ashoka Vanjare 4 ABSTARCT: In his work, gray-level co-occurrence marices (GLCM) have been used o quaniaively evaluae saisical exural parameers for a SAR image and o generae a filered image o feed o ANNs for classificaion for land cover. Prior o performing he exural analysis, an adapive filer was applied o reduce he effec of radar-sysem-generaed coheren speckle o produces an image approximaing local one while preserving edge definiion. A feaure se was han chosen ha bes classifies he SAR image ino he aimed classes. The feaures are seleced based on heir discriminaion abiliy and classificaion accuracy. And a las, he hree ANNs used were compared using he image formed by he chosen feaures in combinaion. Index Terms: GLCM, SAR, exure, exure measures, speckle, adapive filer; classificaion, mulispecral, arificial neural neworks. Secion 1. Inroducion Land cover refers o he physical and biological cover a he surface of he earh, including waer, vegeaion, bare ground, man-made srucures, ec. Land cover informaion acs as imporan piece of informaion for assessmen of land uilizaion and resources. I can indeed play a significan role in planning for opimized usage of he resources in hand. Due o is echnological robusness and cos-effeciveness, remoe sensing has been increasingly used o derive land cover informaion hrough eiher manual inerpreaion or auomaed classificaion. The laer is more desirable bu can be less effecive for classifying heerogeneous landscapes. Arificial neural neworks are commonly conceived o have he capabiliy of improving auomaed classificaion accuracy due o heir disribued srucure and srong capabiliy of handling complex phenomena.many experimens have shown ha MLP neural neworks are more accurae for land cover classificaion han radiional saisical mehods (Bischof e al., 1992; Civco, 1993; Serpico e al., 1996; Chen e al., 1997) The characerisics of daa acquired by opical and synheic aperure radar (SAR) sensors grealy differ. Mulispecral saellies such as Landsa provide informaion on he energy scaered and radiaed by he Earh s surface in differen wavelenghs, from he visible o he hermal infrared, providing he abiliy o discriminae beween differen land cover classes such as vegeaed areas, waer surfaces, and urban ceners. SAR sensors such as he European Remoe Sensing saellies (ERS) ½ provide measuremens in ampliude and phase relaed o he ineracion of he Earh s surface wih microwaves. These acquisiions (C-band) are characerized by high reurns from buildings in urban areas and low and very low values from vegeaed areas and waer surfaces, respecively. Wihin residenial areas, furher discriminaion is achievable because he low-densiy areas are generally characerized by lower backscaering, given he wide srees and he presence of rees. This means ha SAR sensors provide informaion ha may no be obained from opical sensors, and herefore, daa fusion poenially provides improved resuls in he classificaion process compared o he convenional single-source classificaion resuls [27]. Consideraion of he visual inerpreaion of radar images reveals ha he inrinsic spaial variabiliy, or "exure," of he image, beyond ha caused by speckle, is a valuable feaure in discriminaing among differen land-use ypes. Texure may, in fac, be more useful han image one in inerpreing radar images. Many works says ha exure filering can improve he classificaion of SAR images. [1][2][3]. Alhough no precise definiion of exure exiss, cerain conceps of exure can be defined [4].Texure involves he spaial disribuion of gray levels in a local region. I conains imporan informaion abou he srucural arrangemen of surfaces and heir relaionship o heir neighboring surfaces. Some sudies [5][6][7][8][9] have indicaed ha classificaion based on exure migh be more robus han classificaion based on gray values alone. F. Pacifi, e al.,[28] inroduced a novel mehod o use a combinaion of Mulispecral and SAR daa o achieve improved resuls in he classificaion process compared o heconvenional single-source classificaion resuls, wih he usage of Arificial Neural Nework Classifier. This paper proposes a similar idea, o use a more robus informaion in he form of SAR 1 Chief Research Scienis, Indian Insiue of Science ; conac : omkar@aero.iisc.erne.in 2 Suden, IIT Indore ; conac : ee @iii.ac.in 3 Suden, IIT Kharagpur ; conac : sandeshc@cse.iikgp.erne.in 4 Research Assisan, Indian Insiue of Science ; conac : ashokavanjare@gmail.com

2 and MS daa o produce a classificaion ha is an improvemen over he convenional saisical mehods using ANN classifiers and compare he impac of each of hose classifiers on he classificaion problem. The paper is organized as follows: An idea of he sudy area under consideraion is presened in Secion 2, followed by a brief descripion abou saisical characerisics of a Radar Image in Secion 3. Secion 4 is used o describe he mehods used o process he SAR daa and he oucomes in he due process is also presened. A brief descripion of he hree Arificial Neural Nework Classifiers used, viz., he MLP(Muli-Layer Percepron), RBF- NN(Radial Basis Funcion - Neural Nework) and he recenly developed complex-valued neural nework echnique CC- ELM(Circular Complex valued Exreme Learning Machine), is presened in Secion 5. This secion is hen followed by Secion 6, where he preparaion of he Daa Se (Training and Tesing) for he applicaion of supervised mehods, is presened along wih he challenges faced in due course. Secion 7 presens he performance evaluaion of he various parameers of exure filering, in order o sele for he parameers which are bes for he improvemen of he final classificaion of he full-image. Secion 8 is dedicaed o he qualiaive analysis of he classified images generaed from he hree classifiers using he daa se prepared form he resuls obained in Secion 7.Finally, Secion 9 summarizes he main conclusions from his sudy. Secion 2.Sudy area and daa: The sudy area is urban region of Bangalore Ciy (12 58 N, E), locaed in he Souh-Eas region of Karnaaka Sae in Souhern India. The images used in his paper were aken in he year 2014 and are given below. Figure A. Figure B. A.) Radar daa: Our daa se is a C-band ( cm wavelengh, 5.35 GHz) single-polarized RISAT-1 SAR daa. The polarizaion is HH (horizonal ransmi, verical received). The resoluion of his image is approximaely 18 X 18 m. Area = square kms. B.) Mulispecral daa: A 30 m resoluion, specral daa has been used for he purpose of classificaion of he same region. Secion 3. Saisical Characerisics of a Radar Image:

3 The saisical variaions observed in a radar image of a disribued arge are aribued o wo sources: signal fading and inrinsic scene variabiliy. The coheren naure of SAR gives rise o pixel-o-pixel variaions in image inensiy (or a relaed quaniy) ha accouns o coheren fading (or speckles). The spaial variabiliy in he scaering properies of he ground cells ha consiue a arge gives rise o an inrinsic scene exure. Speckle in SAR images is a scaering phenomenon. Speckle appearing in synheic aperure radar (SAR) images is due o he coheren inerference of waves refleced from many elemenary scaerers. This effec causes a pixel-o-pixel variaion in inensiies, and he variaion manifess iself as a granular noise paern in SAR images [10]. Speckle can be reaed as a random process governed by signal fading and was considered o be saisically independen of he exural variaions associaed wih he spaial variaions of he scaering properies of various visually "uniform" disribued arges in he SAR image. Considering his he whole experimenaion was done. Secion 4.SAR Daa processing: Speckle Filering: Firs, speckles were removed using adapive filers. One such approach was earlier used by some auhors. [11][12].Speckle reducion filers for he images ha conain high frequency informaion such as edges and exure, should be adapive o he local exure informaion. These adapive filers can smooh speckle in homogeneous areas while preserving exure and high frequency informaion in heerogeneous areas. Numerous adapive filers have been proposed in he pas and he few famous are he Lee [14], he Kuan [15], he Fros [16], and he modified Lee and modified Fros filers proposed by Lopes e al. [17]. For ha, we choose modified fros filer. Modified adapive filers are effecive in preserving edges and exure in he images, A Comparison of Digial Speckle Filers by Zhenghao Shi and KO B. Fung, proven his resul [18]. Boh modified lee and fros filers are equivalen for small window sizes. Modified Fros filer was chosen for his work[17]. Parameers chosen: The size of he window should be relaed o he size of he objecs, relaive o he spaial resoluion of he scene. A 5 *5 seemed a good choice for he daa. The filer was applied via Envi 4.8 sofware package, which gives us his filer as an inbuil-feaure. The damping facor was aken o be 0.1 and coefficien of variaion for homogeneous areas (Cu) and for heerogeneous areas (C max), was aken as ha for a single look image as and respecively, following he deducions of[17].the speckle-filered image is shown below.

4 Texure Filering: Nex, he speckle-filered image was exure filered, based on GLCM exure model [19]. GLCM exure model: In saisical exure analysis, exure feaures are compued from he saisical disribuion of observed combinaions of inensiies a specified posiions relaive o each oher in he image. According o he number of inensiy poins (pixels) in each combinaion, saisics are classified ino firs-order, second-order and higher-order saisics. The Gray Level Co-occurrence Marix (GLCM) mehod is a way of exracing second order saisical exure feaures. A GLCM is a marix where he number of rows and columns is equal o he number of gray levels, in he image. Suppose an image o be analyzed is recangular and has N x resoluion cells in he horizonal direcion and N y resoluion cells in he verical direcion. Le he gray one appearing in each resoluion cells be quanized o N g levels. Le L x ={1,2,.,N x} be he horizonal spaial domain, L y ={1,2,..., Ny} be he verical spaial domain and G ={1,2...,Ng} be he se of N g quanized gray ones. The se L y x L x is he se of resoluion cells of he image ordered by heir row- column designaions. The image I can be represened as a funcion which assigns some gray one in G o each resoluion cell or pair of coordinaes in L y x L x, such ha I: L y x L x --> G GLCM is a marix of relaive frequencies P ij wih which wo neighbouring resoluion cells separaed by disance d occur on he image, one wih gray one i and he oher wih gray one j. This marix is a funcion of he angular relaionship beween he neighbouring resoluion cells as well as a funcion of he disance beween hem. Formally, for angles quanized o 45 o inerval he unnormalied frequencies are defined as follow: Le I (k,l) = i, I (m,n) = j where : # denoes he number of elemens in he se. Noe ha hese marices are symmeric. Texural feaures : They are compued from a co-occurrence marix of relaive frequencies of gray levels a neighbouring pixels. Many saisics may be derived from he GLCM. 8 of hem which are commonly used have been applied in his paper. These feaures are saed in he able according o heir groups. Group 1 measures are for smoohness ; group 2 for homogeneiy and group 3 for general sasics.

5 Noaions : The exure measures have been defined below : Texure measuremen requires he choice of window size, quanizaion levels, displacemen value, and orienaion facors for each exure measure. In his paper, each exure measure/feaure have been calculaed in all 4 orienaion angles (0, 45, 90, 135 ) and feaure measures of he marices of all he four orienaions arehen averaged, o ge a direcionally invarian marix. Displacemen value d was chosen as 1.Window size was again 5*5 and quanizaion level as 64, which was sufficien and efficien as esablished by [20], who said ha quanizaion level above 24 performs well. PARAM IMAGE 5

6

7 Secion 5. Arificial Neural Nework (ANN) Classifiers In his secion, he hree ANNs used, viz., he MLP(Muli-Layer Percepron), RBF-NN (Radial Basis Funcion - Neural Nework) and CC-ELM (Circular Complex valued Exreme Learning Machine) are discussed and finally a nework model is proposed for he purpose of classificaion using he Mulispecral and SAR (Synheic Aperure Radar) daa. Secion 5.1 The Muli-Layer Percepron Neural Nework Classifier (MLP) Mulilayer Percepron (MLP) model is a feed forward nework. The MLP model chosen for he daa se is a hree layered nework as shown in figure 1.

8 Figure 1. The inpu layer has he number of nodes equal o he dimension of he inpu vecor of each sample in he daa se, viz., 8 (7 Mulispecral + 1 Radar bands) and 10 (7 Mulispecral + 3 Radar bands) resp. for each of he phase of running he classifiers on individual exural parameer filered SAR images and of ha when is ran on 3 exural parameers ogeher.the hidden layer neurons possess a non-linear acivaion funcion in he form of a sigmoid funcion in his case. The oupu of sigmoid funcion of he h h hidden neuron is given by, y h = 1 / (1+exp(-a h), where, (1) a h = (x i*w hi) + b h (2) Here, x i is he inpu from he i h node, w hi is he weigh connecing h h hidden neuron o he i h inpu node and b h is he bias conneced o hh hidden neuron. An MLP wih a single hidden-layer is used as shown in figure 1 is used in his problem. The increase in he hidden neurons resuls in he increase in nework classificaion because, more number of neurons resuls in fiing beer boundaries beween he classes in daa disribuion, as each neuron has a capabiliy o draw an imaginary hyper-plane in he sample space and hus creaing a pariion. There exiss a hreshold value afer which he increase in he number of hidden neurons amouns o decrease in he classificaion accuracy as i sars o over-fi and becomes less generalized. Hence, he number of hidden neurons is opimally se depending on he overall performance of he nework. The oupu layer neurons have a purely-linear acivaion funcion, i.e., he acivaion of an oupu neuron is i s oupu. For oupu vecor is an m-dimensional vecor, where, m is he number of classes presen in he final classified image. The oupu vecor is considered o be of he form, Y = [ Y 1 Y k Y m ] T (3) Y k = 1 if C =k, else 0 (4) Where, c is he class o which he h sample/observaion belongs o. The error is obained by calculaing he sum of squared difference beween he Oupu and Targe value for each sample. Thus he error is given by, E = i=1 No (Y i - T i) 2 (5) Where N o is he number of oupu nodes ; Y i is he oupu of he i h node and T i is he arge value. The raining on he neural nework is carried ou by presening raining se o he nework and he weighs are adjused hrough ieraions. The weighs are iniialized randomly. In every ieraion, he nework oupu is calculaed and he error is deermined. This error is used o adjus he weighs of he nework by using Scaled Conjugae Gradien mehod. The raining is coninued ill error is minimized o a desired value. Training aims a minimizing he oupu error and opimizing he nework weighs. Afer raining, he nework is checked for he measure of generalizaion and performance using he esing se. Secion 5.2 Radial Basis Funcion Neural Nework Classifier (RBF-NN) A radial basis neural nework (RBF-NN) is a commonly used alernaive o ML-NN. RBF-NN is also a feed forward nework. I is srucurally similar o MLP-NN as shown in figure 2, bu he acivaion funcion in he hidden layer nodes is called radial basis acivaion funcion. The oupu of he acivaion funcion depends on he locaion of he cenre of he funcion and he spread of he funcion. The oupu of a radial basis funcion can be defined as, ɸ(x) = exp ( - x c 2 / 2*σ 2 ) (6) where, c is he cenre of he RBF uni, x is he inpu and σ is he spread of he RBF uni.

9 Figure 2. Figure 3. The inpus are firs normalized using suiable normalizaion. This acivaion funcion in he hidden layer produces a nonzero response when he inpu falls wihin kernel funcion. Each hidden uni has is own recepive fields in inpu space. The weighs connecing he inpus o he hidden layer decide he spread of he acivaion funcion and he weighs connecing he hidden layer o he oupus is used as a scalar muliplier o he hidden layer oupus. The nework oupu is he sum of weighed hidden layer oupus. The raining of he nework involves he adjusmen of he wo ses of weighs and updaing of ceners of he hidden nodes. The mean squared error is deermined beween he nework oupu and he arge oupu values. The ceners and weighs are iniialized randomly. The weighs and ceners are opimized using Gradien Descen Back Propagaion algorihm, by minimizing he insananeous mean squared error. The error cos funcion is defined as shown in equaion (5).The raining algorihm aims a minimizing he error and he opimizaion of he weighs and he cener locaion. The raining is carried ou ill he arge performance is reached. The nework is hen esed for performance and generalizaion using he esing daase. Secion 5.3 Circular Complex Exreme Learning Machine (CC-ELM) R. Saviha [23] e al., presened a fas learning fully complex-valued exreme learning machine classifier, referred o as Circular Complex-valued Exreme Learning Machine (CC-ELM) for handling real-valued classificaion problems. CC-ELM is a single hidden layer neworkwih non-linear inpu and hidden layers and a linear oupu layer. A circular ransformaionwih a ranslaional/roaional bias erm ha performs a one-o-one ransformaionof real-valued feaures o he complex plane is used as an acivaion funcion for he inpuneurons. The neurons in he hidden layer employ a fully complex-valued Gaussian-like( sech ) acivaion funcion. The inpu parameers of CC-ELM are chosen randomly andhe oupu weighs are compued analyically. The inpu and oupu configuraions of CC-ELM are similar o ha of MLP and RBF-NN. In CC-ELM he inpu layer acivaion funcion is Circular Transformaion, which maps he real valued daa o he complex domain, as in figure 4. The ransformaion is given by, z = sin ( ax + ibx + α ) (7) where, x is he inpu vecor for he h sample/observaion,such ha each of x i is normalized in [0,1]; 0 < a, b 1, and 0 <α< 2π.

10 Figure 4. The bias erm ensures randomness in he disribuion of he feaure vecor and prevens overlapping in he complex plane. This also resuls in he efficien use of orhogonal decision boundaries. The above funcion is analyic and bounded in [0,1] and he inpu is mapped is mapped o [0,1] hence, as menioned in [24] i is a valid acivaion funcion. The weighed response of he inpu layer is sen o he hidden layer. The hidden layer employs a Radial Basis acivaion funcion as in he RBF neworks bu in he complex domain. The hidden layer response of j h hidden neuron is, z h = sech ( σ j T ( z - c j ) ) (8) Where, σ j ϵ C m is a scaling facor of he acivaion funcion, c j ϵ C m is cener of he acivaion funcion. As menioned in he previous secion, each neuron has 2 orhogonal surfaces as decision boundaries. This divides he feaure space ino 4 secions insead of 2, as in he case of real valued neurons. This conribues o higher classificaion efficiency as shown by Nia [25]. The plo of magniude and phase of he acivaion funcion is shown in figure 5 and figure 6, respecively. Each of hese hidden neurons is conneced o he neurons in he oupu layer. The oupu neurons calculae he weighed sum of he response of hidden layer neurons, which is effecively he oupu of he nework. The arges are class coded as, Y = [ Y 1 Y k Y m ] T (9) Y k = 1+i if C =k, else -1-i (10) Figure 5. Figure 6. Given he marix of oupu weighs as V oxh and he response of hidden neurons as H, and he oupu of he nework Y, he raining procedure reduces o finding he leas squares soluion V oxh of he sysem: Y = VH (11) In he case where number of hidden neurons is equal o he number of raining samples, H is inverible. Bu in a pracical scenario, number of hidden neurons is much less han he number of raining samples; hence H may no be inverible.

11 Hence here may no exis a H -1 and V ha saisfies he above equaion. Therefore he soluion wih he minimum norm of leas squares is, Ѵ = YH ǂ (12) Where, H ǂ is he Moore-Penrose Pseudo inverse of he hidden layer oupu marix [26] and Y is he complex-valued coded class label. Secion 5.4 Proposed Model for he Nework Archiecure Secion 6. Preparaion of he Daa Se Secion 6.1 Down sampling he High Resoluion SAR daa Figure 7. In order o use boh he Mulispecral and SAR daa in parallel o generae he daa se, i is imporan ha boh he daa can be linked in a way o benefi from he muual informaion ha hey conain. The simples way is o down sample he Higher resoluion SAR image (18m) o he resoluion of he Mulispecral image (30m) and hen co-regiser boh he images. The mehod employed o down-sample he SAR image was based a simple weighed average. Secion 6.2 Co-regisraion of MS and SAR image Figure 8 The co-regisraion of he wo daa in he form of MS and SAR images is a quinessenial componen o proceed furher wih he classificaion. The co-regisraion was done manually by examinaion using he ERDAS IMAGINE 9 sofware and hen corresponding laeral shifs were applied on he images. Figure 9 : Before Co-regisraion Figure 10 : Afer Co-regisraion Secion 6.3 Exracing he Training and Tesing daa se from he MS and SAR images

12 The Training and Tesing daa se were prepared by exracing pixels from he co-regisered image using pre-consruced AOIs (Area of Inres), carefully designed for each of he 4 classed (LAND,URBAN,VEG,WATER). The design process of he AOIs were consanly verified wih Google Earh daa. The properies of he daa se prepared is as follows, Figure 11 : AOI designed o exrac WATER pixels Secion 7. Performance evaluaion of he various parameers of exure filering Clausi,e al. [22], menioned ha hree exure measures ha provide differen informaion aken in combinaion can give beer classificaion han individual exure feaures being used. The goal of his secion was o find he parameers of exure filering for he SAR image, which sui bes for improving he classificaion accuracies. The 8 parameers chosen were ASM, CON, COR, DIS, ENT, HOM, MEAN, VAR. The classifiers were run on he raining and esing daa se for hese 8 parameers individually, i.e., 8 inpus ( 7 MS bands + 1 Texure Filered

13 SAR image corresp. o a single parameer ) were fed o he classifiers. The resuls are presened in Table 1. And Table 2. corresp. o resuls obained from R-RBF and CC-ELM respecively. Legend ASM 6 - CON 7 - COR 8 - DIS 9 - ENT 10 - HOM 11 - MEAN 12 - VAR Table 1. Table 2. I was noed from he above resuls ha ENT and COR perform he bes, and he hird parameer o be chosen is sill in doubs. Hence four combinaions of 3 parameers aken ogeher were formed, so ha he classifiers could jusify as o which one gives he bes resuls. The four combinaions chosen were, 13 - ENT,COR,MEAN 14 - ENT,COR,CON 15 - ENT,COR,DIS 16 - ENT,COR,HOM The classifiers were ran on he above four combinaions of daa, i.e, 10 inpus ( 7 MS + 3 SAR exure filered, where he 3 SAR inpus corresp. o one parameer of exure filering for each of he combinaions) were fed o he ANNs o find which one gives he bes resuls of hem all. Table 3. And Table 4. provide he resuls obained from R-RBF and CC-ELM resp. on he four combinaions. Table 3.

14 Table 4. I is observed from he above resuls ha he combinaion 16 provides he bes resuls in boh he cases. Hence he combinaion of ENT, COR and HOM is chosen o prepare he daa se for he cause of classifying he enire daa. Secion 8. Qualiaive Analysis of he Classified Images obained from he ANN classifiers The following resuls were observed on he raining and he esing daa ses. Secion 8.1 Running he classifiers on Paches of he enire daa The classifiers were run on paches of images from he bigger daa and he resuls are presened below and are followed by a qualiaive analysis of he resuls obained. Original MS MLP R-RBF CC-ELM

15 For he above resuls i can be observed ha he MLP resuls are saisfacory o a fair exen. The CC-ELM resuls on paches 4 and 5 are no convincing as i misclassifies many VEG pixels as LAND pixels and deecs a fewer URBAN pixels han wha i had o. R-RBF on pach 4 produces similar resuls o wha CC-ELM produced, viz., misclassified VEG pixels as LAND pixels. Secion 8.2 Running he classifiers on he Full daa Below are he resuls obained afer classificaion of he enire image using he daa se prepared from he resuls of Secion 7, using he ANN classifiers. MLP R-RBF CC-ELM

16 The R-RBF classified image shows ha here has been a lo of misclassified pixels, mainly oher caegories (especially URBAN) misclassified as WATER pixels. The CC-ELM resuls show almos no valid classificaion on he bigger image, hough i showed accepable resuls on he smaller paches. Secion 9. Conclusion I is difficul o define he mos suiable exure measure o incorporae ino he per-pixel classificaion of SAR image, bu he findings presened in his paper are imporan for considering which exure measures can be used for accurae classificaion. Firs i was inferred ha exure measures ha are bes for land-use/cover classificaion in SAR land images are Inverse Difference Momen, Enropy and Correlaion. The resuls obained via MLP classifier are saisfacory, while here were a lo of shor comings in he resuls obained from R-RBF and CC-ELM. The possible improvemens o re-work on can be: Auomaed co-regisraion of MS and SAR daa, insead of he manual co-regisraion as done in he above work. Generae a beer raining and esing daa se, if he already generaed daa se isn sufficien. Invesigae as o why he resuls obained were didn include much variaion even wih so many combinaions of daa. Trying o achieve more generalized performance for he ANN classifiers insead of over-fiing he Training daa se. Invesigae as o why R-RBF and CC-ELM didn provide beer resuls han MLP, on he same daa se, as is proven heoreically. Invesigae why CC-ELM isn providing good resuls on he enire image. Acknowledgemens: We are graeful for he conribuion by Vivekananda Shankaracharya, S N Omkar, J. Senhilnah for heir immaculae work on crop classificaion using supervised learning echniques and heir guidance regarding he proceedings of he projec. References: [1] Ulaby, F.T.; Kouyae, F.; Brisco, B.; Williams, T.H.L., "Texural Infornaion in SAR Images," Geoscience and Remoe Sensing, IEEE Transacions on, vol.ge-24, no.2, pp.235,245, March 1986 doi: /TGRS [2] Barber, David G., e al. "A comparison of second-order classifiers for SAR sea ice discriminaion." Phoogrammeric engineering and remoe sensing 59.9 (1993): [3] Berberoglu, Suha, e al. "Texure classificaion of Medierranean land cover." Inernaional Journal of Applied Earh Observaion and Geoinformaion 9.3 (2007): [4] M. Tuceryan and A. K. Jain. Texure analysis. In C. H. Chen, L. F. Pau, and P. S. P. Wang, ediors, Handbook of Paern Recogniion and CompuerVision, pages World Scienific Publishing Company, 1993 [5] D. G. Barber, M. E. Shokr, R. A. Fernandes, E. D. Soulis, D. G. Fle, and E. F. LeDrew. A comparison of second-order classifiers for SAR sea ice discriminaion. Phoogrammeric Engineering &Remoe Sensing, 59: , [6] R. T. Franko and R. Chellappa. Lognormal randomfield models and heir applicaions o radar image synhesis. IEEE Trans. Geosc. Remoe Sensing, 25: , [7] Murni, N. Darwis, M. Masur, and D. Hardiano. A exure classificaion experimen for SAR radar images. In Proceedings of Paern Recogniion in Pracice IV, pages , Vlieland, Neherlands, June [8] A. H. Schisad and A. K. Jain. Texure analysis in he presence of speckle noise. In Proceedings of heinernaional Geoscience and Remoe Sensing Symposium (IGARSS), pages , Houson, Texas, May 1992.

17 [9] M. E. Shokr. Texure measures for sea-ice classificaion from radar images. In Proceedings of he InernaionalGeoscience and Remoe Sensing Symposium(IGARSS), pages , Vancouver, Canada, [10] Lee, J.S., Poier, E., Polarimeric Radar Imaging - From basics o applicaions, CRC Press (2009)) [11] Holmes, Quenin A; Nuesch, Daniel R.; Shuchman, R.A, "Texural Analysis And Real-Time Classificaion of Sea-Ice Types Using Digial SAR Daa," Geoscience and Remoe Sensing, IEEE Transacions on, vol.ge-22, no.2, pp.113,120, March 1984 doi: /TGRS \ [12] Solberg, Anne H. Schisad, and Anil K. Jain. "Texure fusion and feaure selecion applied o SAR imagery." IEEE Transacions on Geoscience and Remoe Sensing 35.2 (1997): [13] Schisad, AH.; Jain, AK., "Texure Analysis in he Presence of Speckle Noise," Geoscience and Remoe Sensing Symposium, IGARSS '92. Inernaional, vol.2, no., pp.884,886, May 1992 doi: /IGARSS [14] Jong-Sen Lee, "Digial Image Enhancemen and Noise Filering by Use of Local Saisics," Paern Analysis and Machine Inelligence, IEEE Transacions on, vol.pami-2, no.2, pp.165,168, March 1980 doi: /TPAMI [15] Kuan, Darwin T.; Sawchuk, AA; Srand, Timohy C.; Chavel, P., "Adapive Noise Smoohing Filer for Images wih Signal- Dependen Noise," Paern Analysis and Machine Inelligence, IEEE Transacions on, vol.pami-7, no.2, pp.165,177, March 1985 doi: /TPAMI [16] Fros, Vicor S.; Siles, Josephine Abbo; Shanmugan, K.S.; Holzman, J., "A Model for Radar Images and Is Applicaion o Adapive Digial Filering of Muliplicaive Noise," Paern Analysis and Machine Inelligence, IEEE Transacions on, vol.pami-4, no.2, pp.157,166, March 1982 doi: /TPAMI [17] Lopes, A; Touzi, R.; Nezry, E., "Adapive speckle filers and scene heerogeneiy," Geoscience and Remoe Sensing, IEEE Transacions on, vol.28, no.6, pp.992,1000, Nov 1990 doi: / [18] Zhenghao Shi; Fung, K. B., "A comparison of digial speckle filers," Geoscience and Remoe Sensing Symposium, IGARSS '94. Surface and Amospheric Remoe Sensing: Technologies, Daa Analysis and Inerpreaion., Inernaional, vol.4, no., pp.2129,2133 vol.4, 8-12 Aug 1994 doi: /IGARSS [19]Haralick, R.M.; Shanmugam, K.; Dinsein, Is'Hak, "Texural Feaures for Image Classificaion," Sysems, Man and Cyberneics, IEEE Transacions on, vol.smc-3, no.6, pp.610,621, Nov doi: /TSMC [23]. Fas learning Circular Complex-valued Exreme Learning Machine (CC-ELM) for real-valued classificaion problems R. Saviha, S. Suresh,, N. Sundararajan [24]. Kim, T., Adali, T.: Fully complex muli-layer percepron nework for non-linear signal processing. Journal of VLSI signal processing 32(1/2), (2002) [25]. Tohru Nia: Orhogonaliy of decision boundaries in complex valued neural neworks. Neural Compuing Jan; 16 (1): [26]. Orega, J.M.: Marix Theory. Pienum Press, New York (1986) [27]. A. H. Schisad Solberg, A. K. Jain, and T. Tax, Mulisource classificaion of remoely sensed daa: Fusion of Landsa TM and SAR images, IEEETrans. Geosci. Remoe Sens., vol. 32, no. 4, pp , Jul [28]. Pacifici, Fabio, e al. "Urban mapping using coarse SAR and opical daa: Oucome of he 2007 GRSS daa fusion cones." Geoscience and Remoe Sensing Leers, IEEE 5.3 (2008):