Classification of Satellite Images by Texture-Based Models Modulation Using MLP, SVM Neural Networks and Nero Fuzzy

Transcription

1 Internatonal Journal of Electroncs and Electrcal Engneerng Vol. 1, No. 4, December, 2013 Classfcaton of Satellte Images by Texture-Based Models Modulaton Usng MLP, SVM Neural Networks and Nero Fuzzy Gholam Reza Shahrar, Abbas Gharb, and Azm Rezae Motlagh Iranan Ol Termnal Company Emal: and Abstract The purpose of ths paper s the automatc dentfcaton of varous dstrcts n satellte mages usng the textural feature, whle comparng them by two methods of GLCM and Fourer Spectrum. The modulaton of dscrete volet and GLCM yelded a new method for the dentfcaton of the urban areas that s used as a crteron for measurng the development rate n the urban areas usng satellte mages. Through the modulaton of GLCM and spatal features, MLP Neural Network and salency measurements made t possble to determne the most mportant textural features for sub-metrc spatal resoluton magery of urban scenes. That s used as a crteron for measurng the development rate n the urban areas usng satellte mages. The results of smulaton usng MATLAB/IMAGE PROCESSING software on IKONOS database, from whch the mages have been collected, verfy the accuracy of the performance of ths system. network can be consdered as a specal category of nformaton that has to be taken nto account for the analyss. Ths expertse can be used to estmate the desred new poston and answer the queston of "what f". Other advantages nclude the followng Adaptve learnng: The ablty to learn how to do tasks based on tranng experence for ntroductory nformaton [1]. II. PROBLEM STATEMENT The geographsts have been workng on several urban areas on the map for a long tme. It seems more essental than ever to have an ntellgent and powerful system updated wth urban development and mprovements. The technques that process a wde area wth a hgh speed and accuracy are hghly needed, so the researchers are workng on these technques. Durng the last decade, dgtal magng tools have been progressed renderng t possble to vew the surface of an area wth ts spatal and spectral detals. Advanced aeral photography spectrometers are beng ncreasngly used n varous applcatons [1], [2]. However, the cost of the data receved from these mult-spectral sensors s really hgh. The researchers wth lmted fnancal resources would not afford to observe the requred data from hgh-frequences. The next problem for mult-spectral aeral data s ther hgh resoluton whch makes the feature dentfcaton, data processng, data extracton and obect classfcaton n the mages qute dffcult [2], [3]. Ths chapter deals wth the analyss of the method suggested n ths artcle and ts advantages and dsadvantages wll be observed on varous mages. The nput to ths system s the satellte mages collected from the source mentoned n whch s wdely used as a source n most works on urban zones classfcaton. These reference mages are classfed under the followng categores: Green space, street, hghway, resdental houses. Hence, we have four classes whch should be called nput mages and the areas where these classes are located have to be dentfed. The avalable mages are used to tran the neural system; Therefore, these samples were selected so that they would be as dfferent as possble. Through a consderable level of overlook, the neural networks can be called the electronc models of the neural system of Index Terms textural features, GLCM matrx, MLP neural network, feature vector, satellte mages. I. INTRODUCTION Satellte mages are beng wdely used n several felds of technology. Due to the abundance of the avalable data n ths feld, we need advanced automatc methods for extractng and provdng the user wth the requred data out of these mages. The Techncal Commttee of data extracton DFT s a subgroup of Land Remote Sensors team. Insttute of Electrcal & Electronc Engneers (IEEE) has been workng on ths feld for many years and t has publshed several works ncludng the most recent achevements of researchers n ths feld. The problem nvestgated n ths paper s the automatc dentfcaton of varous areas n a satellte mage. These areas are selected based on the applcaton of ths artcle. GLCM matrx s one of the maor elements n ths study. GLCM matrx s a 2D matrx and the (,) element s the number of co-occurrences of and. Neural networks, wth a remarkable ablty to derve meanng from complcated or ambguous data, for extract patterns and dentfy the methods and technques that knowledge of those wll be complex and dffcult for computer and humans, Wll be used. A traned neural Manuscrpt receved Aprl 25, 2013; revsed November 28, 2013 do: /eee

2 Internatonal Journal of Electroncs and Electrcal Engneerng Vol. 1, No. 4, December, 2013 human bran. The learnng and tranng mechansms of the bran s bascally rooted n experence. An artfcal neural network s an dea for data processng whch has been nspred by the bologcal neural system and deals wth data processng ust lke bran. These mages are used n ths way: a feature vector s extracted from the mages of each class and ths feature vector s used for tranng the neural network. In other words, the neural network s traned to show the hghway class at the output f the feature vector of ths class was ntroduced as the nput. A feature vector s obtaned for each mage and t corresponds to the same mage. The program reads an mage n every attempt and processes t to extract ts feature vector. Another pont consdered n the collecton of the neural network tutoral mages database s the hgh number of ts mages, as the hgher the number of tutoral samples the more the generalzaton possblty for the neural network. The other pont s dfferent szes of mages. The system s desgned so that t would be able to work on all mages wth any szes. When collectng the classes, the attempt was made to choose the mages that are totally dfferent n sze. The desgned system enables us to easly ncrease the number of classes. Here, we have consdered 4 classes; however, ths system can be used for any number of classes [3]. III. PROPOSED MLP NEURAL NETWORK FOR CLASSIFICATION OF SATELLITE IMAGES Ths chapter deals wth the analyss of the method suggested n ths artcle and ts advantages and dsadvantages wll be observed on varous mages. The nput to ths system s the satellte mages collected from the source mentoned n whch s wdely used as a source n most works on urban zones classfcaton. These reference mages are classfed under the followng categores: green space, street, hghway, resdental houses. Hence, we have four classes whch should be called nput mages and the areas where these classes are located have to be dentfed. The avalable mages are used to tran the neural system; therefore, these samples were selected so that they would be as dfferent as possble. Through a consderable level of overlook, the neural networks can be called the electronc models of the neural system of human bran. The learnng and tranng mechansms of the bran s bascally rooted n experence. An artfcal neural network s an dea for data processng whch has been nspred by the bologcal neural system and deals wth data processng ust lke bran. These mages are used n ths way: a feature vector s extracted from the mages of each class and ths feature vector s used for tranng the neural network. In other words, the neural network s traned to show the hghway class at the output f the feature vector of ths class was ntroduced as the nput. A feature vector s obtaned for each mage and t corresponds to the same mage. The program reads an mage n every attempt and processes t to extract ts feature vector. Another pont consdered n the collecton of the neural network tutoral mages database s the hgh number of ts mages, as the hgher the number of tutoral samples the more the generalzaton possblty for the neural network. The other pont s dfferent szes of mages. The system s desgned so that t would be able to work on all mages wth any szes. When collectng the classes, the attempt was made to choose the mages that are totally dfferent n sze [4]. The desgned system enables us to easly ncrease the number of classes. Here, we have consdered 4 classes; however, ths system can be used for any number of classes. Most conventonal methods of mage automatc classfcaton based on pxel color and tone and mportant characterstcs such as texture, shape, content wll be gnored completely. In general, segmentaton methods can be classfed nto the followng types [4], [5] based on pxels (cluster threshold Investment and down Based on Edge (flterng and mprovng) Based on the (growth area, connectng the area, segregated area Other methods nclude fuzzy technques, neural network methods (mlp neural network, neural network SOM), physcally based method (bcolor reflecton model and the model reflect the approxmate color) are texture-based methods. So many ways to classfy these two classes of SOM neural networks and fuzzy methods have been presented n several szes [5]. Although varous technques have been proposed and demonstrated sub-pxel classfcaton s to classfy the two classes (road and houses) has been successfully appled, however, dentfy some of of them are down. One of the methods that have been used so much, lnear spectral mxture analyss (LSMA) for extracton levels are, however, lmtatons LSMA n [6] s nvestgated. Because most statstcal methods for dgtal mage analyss based on the hypothetcal the data does have lmtatons. Furthermore, LSMA assumes that the spectral reflectvty of a pxel lnear combnaton of the the spectral reflect of ts land that s covered pxels. However, there are many examples of non-lnear combnaton. For example, the composton of the vegetaton reflectvty s a non-lnear combnaton of LSMA therefore can t be used here. One labor problems of n classfyng urban areas, an area dffcult belongng to a class. Because sometmes a partcular class of spatal features appear n dfferent places. For example, vegetaton cover are splt n two categores lght and dark. The problem n ths thess have appeared n the smulaton. The feature vectors of dfferent areas sometmes became clear that the area of the fabrc vares by regon, whch has close feature vector., a soluton was used to overcome ths problem and t was consder a number of addtonal features. [7] Normalzed to a range of classes used to reduce classes overlap. Neural networks are another popular method of categorzaton. However, untl now t has ntroduced a large number of neural network MLP neural network has the hghest usage. The MLP neural network has many defects. One of the problems wth neural network MLP, consderng many hdden layers and hdden layer neuron. So far, great thngs have been done n ths area, but stll have not 246

3 Internatonal Journal of Electroncs and Electrcal Engneerng Vol. 1, No. 4, December, 2013 found a good way to do t s more tral and error wth the number of layers and hdden layer neuron s determned. Another dsadvantage of the MLP, the output of whch s used for tranng should nclude both desrable and undesrable output s Network to become famlar wth all types of data, but sometmes produce undesrable data s not possble. Many segmentaton methods have been proposed. Tradtonal methods, hstogram and cluster threshold settng and dentfy the edge of the area to be extracted. One dsadvantage of these methods s that most of these technques are not sutable for nosy envronments. Furthermore, each of these has ts own problems. For example, the nvestment threshold Hstogram of the hstogram of data only takes nto account the space detals of the matter wll not be vsble. The cluster has a bg problem and t s to determne number of clusters. Edge detecton s stll vulnerable aganst nose. Edge of the mage wthout nose s measured wth hgh accuracy, but the actual mage, the nose s nevtable. Regon-based methods, are nsenstve to nose and spatal data and attrbute data are consdered Although ths method has a problem and t s the choce of the startng area, and much depends on the choce of [8], [9]. In Table I the dfferences between classfcaton methods and ther features has been compared. located n a specfc offset, Matrx C wll be defned over a n m mage wth offset (Δx, Δy) parameters as follows: )1) Co-occurrence matrx s a statstcal method that can extract the second-order statstcs n a textural mage. Another defnton that can be offered for GLCM s that: GLCM s a 2D hstogram n whch the (,) element refers to the co-occurrences of and. Co-occurrence matrx s dentfed n two pxels by relatve frequences of P(,,d, θ ), f these two pxels are located n the dstance of d and drecton of θ, one wth the brghtness of and the other wth the brghtness of. Hence, GLCM s a functon of the dstance of r and angle of θ and the brghtness of and. We should look for approprate r and θ, when dentfyng the letters. When GLCM matrx s estmated, the features should be extracted from the matrx. As ths matrx cannot be drectly gven to neural network classfer as a feature vector, some mathematcal operatons have to be performed on GLCM. These mathematcal operatons are known as feature extracton from GLCM. Some of these features are mentoned here. 1) Mean GLCM: f1 TABLE I. A COMPARISON OF CLASSIFICATION METHODS Advantages Settng threshold: there s no need for any data pre-computaton complexty s.low Cluster-out: t s easy to.mplement Tme consumng, the growth s related to the choce of the.startng area Resstant to nose Sometmes the calculaton s.very heavy Fuzzy belongng functons can not be used to express some.properte. Usng math functons and fuzzy rules and segmentaton Usng fuzzy Parallel characterstcs of neural network s.fully utlzed Usng neural networks for segmentaton Neural network Slow learnng and more learnng Feautures Settng threshold s requred that Hystogram have a number of peak. Cluster by assumng that each regon n the mage has a separate cluster n.feature space Technque Settng threshold, the cluster Defects The threshold Settng: the dentfed peaks n the hstogram does not work well and there s no guarantee that the tem be always.contnuous Cluster problem: to determne the number of clusters p (, ) p(, ) 2) Contrast: f2 2 p(, ) 3) Entropy: p(, ) f 3 log p (, ) 4) Angular Second Moment: f 4 p(, ) 2 5) Homogenety s based on Dscontnuty detecton Edge method p(, ) f 5 1 6) Dssmlarty f 6 p(, ) 7) Correlaton f 8 ( x )( y ) p(, ) x y 8) Energy fg = p2(,) IV. DEFINITION OF GLCM =0 =0 Co-occurrence matrx s a matrx that s defned over an mage. If the dstrbuton of co-occurrence values are 247 (2)

4 Internatonal Journal of Electroncs and Electrcal Engneerng Vol. 1, No. 4, December, 2013 These features, as mpled by ther relatons, yeld a number of a GLCM and they are approprate for the producton of feature vector [10]. GLCM, as mentoned earler, can express the way of dstrbuton of pxels values, through approprate selecton of d and θ. It s more common n GLCM works to use the vector form nstead of d and θ. For example nstead of d=1, and θ =45, we wll wrte offset=[1 1]. Fg. 1 shows ths concept schematcally. Fgure 2. Fgure 1. Vector Expresson of d and Fg. 2 llustrates one of the mages avalable n the database. We wll use ths mage here to show the results obtaned from the algorthm. Frst, as t was mentoned earler, the vsual sample wll be collected from each class. The number of samples of each class s shown n Table III. The neural network used n ths secton s obtaned from MLP smulaton. There are 60 nput neurons n the neural network and there are 35 hdden layer neurons and 4 output layer neurons. Tangent-sgmod transfer functon s used to transfer several layers [12]. n the defnton of GLCM Therefore, the selecton of offset parameter wll sgnfcantly nfluence the results obtaned from the applcaton of GLCM, because t shows the way of ts formaton. In ths research, the mage was frst dvded nto 7 7 wndows. In other words, we transfer the nput mage matrx to 7 7 matrces. Then, 12 matrces wll be estmated for four dstances d=2,3,4,5 and angles of θ=0, 45, 90. All offsets are shown n Table II. TABLE III. THE NUMBER OF SAMPLES OF EACH CLASS TABLE II. ALL OFFSETS USED IN THIS RESEARCH FOR GLCM ]2,2[ ]3,3[ ]4,4[ ]5,5[ ]0,2[ ]0,3[ ]0,4[ ]0,5[ An mage from the database. ]2,0[ ]3,0[ [4,0] ]5,0[ Class Tree Street Number Hgh way 44 Resdent al houses 44 After tranng the neural network, the system should be tested by an mage. Fg. 2 s an mage that s used to test a system. After tranng the neural network, the error crteron value s changed as follows and t wll tend toward zero (Fg. 3). when 12 GLCM of 7 7 wndows are estmated, the feature extracton wll be followed. 5 features of correlaton, angular second moment, dssmlarty, homogenety and energy wll be obtaned from each matrx. Therefore, 15 features wll be acheved for every lne of Table I and the total number of 60 features wll be obtaned for Table II. V. SIMULATION RESULT Here, the results of mplementaton of the algorthm, that was defned n the earler sectons, dscussed. The database where the mages were collected from s called IKONOS [11]. Ths database s one of the most well-known databases used n most essays on Images. One of the most sgnfcant features of the mages of ths database s the hgh resoluton of mages. Hgh qualty s not avalable n all databases. Ths feature has made the database dstngushable. Another fact related to the mages of ths database s the aeral mages of urban areas. Fgure 3. Change of error crteron durng neural network tutoral The blue color n the followng mages llustrates the resdental houses, orange shows hghways, red shows 248

5 Internatonal Journal of Electroncs and Electrcal Engneerng Vol. 1, No. 4, December, 2013 street, and green shows green spaces. In ths secton we wll see the results of the applcaton of neural networks. A. Neural Network MLP Fgure 7. Fgure 4. The result of the test mage n Fg. 6 Pcture of the overall system test Fgure 8. Pcture of the overall system test for hdden layer neurons 50 Fgure 5. The result of the test mage n Fg. 4 The second test mage, shown n Fg. 6 and the result s shown n Fg. 7. Fgure 9. The result of the test mage n Fg. 8 Fgure 6. In another experment, neuron number of hdden layer neurons s altered and results close to 50 canddates. Fg. 8 shows the test nput mage and Fg. 9 shows the result of t. In the next experment, the number of hdden layer neurons to 20 neurons changed, the results are shown n Fg. 10. Test Image No

6 Internatonal Journal of Electroncs and Electrcal Engneerng Vol. 1, No. 4, December, 2013 Do fuzzy system called ANFIS s also known. The system combnes neural networks and fuzzy systems are used n many applcatons for classfcaton tasks. The ntal fuzzy system s the Takag-Sugeno type. Fgure 10. The result of hdden layer neurons aganst 20 B. Neural Network SVM: In ths secton the results of the SVM Neural network wll be presented. SVM s a powerful Neural network, on of the neural network, can be defned as the kernel of demarcaton between the classes and do not need to be straght lne. Lnear kernel functon s used n ths smulaton. Fgure 13. The result of the ANFIS VI. CONCLUSION The textural analyss plays a sgnfcant role n dgtal mage processng and ts expresson, and t can provde us wth the extra data for workng on the satellte mages. Usng the concept of texture, we came to ths concept that the area of dfferent zones on satellte mages has dfferent textures. Then, the co-occurrence matrx was used to obtan the features of these regons. The extracted features nclude: dssmlarty, angular second moment, correlaton, energy and homogenety. Four classes of resdental houses, hghways, street, green space were consdered n ths study. The results reveal that SVM neural network has the best effcency. REFERENCES [1] Fgure 11. [2] Results obtaned usng the SVM neural network [3] [4] [5] [6] [7] [8] Fgure 12. Image nput system testng ANFIS 250 A. Nazf, Y. Vural, and T. Fatos, An overvew of character recognton based focused on off-lne handwrtng, IEEE Transactons on Systems, Man, and Cybernetcs-Part C: Applcatons and Revews, vol. 31, no. 2, May Ø. D. Trer and A. K. Jan, Goal drected evaluaton of bnarzaton methods, IEEE Trans. Pattern Anal. Machne Intell., vol. 17, pp , Dec L. Lam, S. W. Lee, and C. Y. Suen, Thnnng methodologes A comprehensve survey, IEEE Trans. Pattern Anal. Machne Intell., vol. 14, pp , Sept B. S. Manunath and W. Ma, Texture features for browsng and retreval of mage data, IEEE Transactons on Pattern Analyss and Machne Intellgence, vol. 18, no. 8, pp , August A. P. Carleer and E. Wolff, Urban land covermult-level regon-based classfcaton of VHR data by selectng relevant features, Internatonal Journal of Remote Sensng, vol. 27, pp , M. Tuceryan and A. K. Jan, (1993) Texture analyss, n Handbook of Pattern Recognton and Computer Vson Sngapore: World Scentfc,.C. Chen, L. Pau, & P. Wang (Eds K. S. Shanmugan, V. Narayanan, V. Frost, S. Stles, J. A., and J. C. Holtzman, Textural features for Dadar mage analyss, IEEE Transactons on Geoscence and Remote Sensng, vol. 19, pp , D. A. Claus, and B. Yue, Comparng co-occurrence probabltes and Markov random felds for texture analyss of SAR sea ce magery, IEEE Transacton on Geoscence and

7 Internatonal Journal of Electroncs and Electrcal Engneerng Vol. 1, No. 4, December, 2013 Remote Sensng, vol. 42, pp , [9] Informaton from Imagery. [Onlne]. Avalable: [10] R. R. Jensen, J. R. Boulton, and B. T. Harper. The relatonshp between urban leaf area and summertme household energy use, n Geo-spatal Technologes n Urban Envronments, R. R. Jensen, J. D. Gatrell, and D. McLean Eds. Sprnger, [11] E. G. Irwn, N. E. Bockstael, and H. J. Cho, Measurng and modelng urban sprawl: Data, scale, and spatal dependences, n Proc. 53 rd Annual North Amercan Regonal Scence Assocaton Meetngs of the Regonal Scence Assocaton Internatonal, Toronto, Canada, [12] V. C. Radeloff, R. B. Hammer, and S. I. Stewart, Rural and suburban sprawl n the U.S. Mdwest from 1940 to 2000 and ts relaton to forest fragmentaton, Conservaton Bology, vol. 19, pp , Gholam Reza Shahryar receved a B.S. and a M.S. n Electronc and Electrcal of Engneerng from Bushehr Unversty, Iran n 2007 and 2011 respectvely. From 2011 to 2012, he was a senor researcher n the Ol Termnal Company Insttue (O.T.C.I). Also, he has been workng as a senor researcher n the Power Converson and System for Renewable Energy Center of the Young Researchers Club (BPJ), Iran. Hs man research nterests are Analyss and Smulaton of Power Regulaton and ntegrated crcuts (cmos). Abbas Gharb receved B.S and M.S degrees n Electronc and Electrcal of Engneerng from Bushehr Unversty, Iran n 2007 and 2011, respectvely. From 2011 to 2012, he was a senor researcher n the the Department of Electronc and Electrcal Engneerng Bushehr & Borazan Unversty, Bushehr, Iran, He oned the Department of Electrcal Engneerng of Borazan Unversty n 2011, where he has been actve n research on the dgtal sgnal processng systems and ntellgent systems usng ntegrated blocks and crcutshe s also have many done researchers on am modulators, neural network, and automatc gan control and CMOS technology. Azm Rezae Motlagh receved B.S and M.S degrees. n Electronc and Electrcal of Engneerng from Bushehr Unversty, Iran n 2004 and 2012,respectvely. From 2010 to 2012, he was Teachng n the the Department of Electronc and Electrcal Engneerng of Borazan Unversty, Bushehr, Iran, where he has been actve n research on the analog and dgtal multplers and nonlnearty crcuts and systems he s already focous at desgn of ntellgent dgtal multpler usng dgtal processng systems. 251