CLSS BSED RTIOIG EFFECT O SUB-PIXEL SIGLE LD COVER UTOMTIC MPPIG nil Kumar *a, Suresh Saggar b, a Inian Institute of Remote Sensing, Dehraun, anil@iirs.gov.in b Birla Institute of Technology & Science, Pilani (Rajasthan), suresh.saggar@gmail.com WG VII/4 KEY WORDS: Sub-pixel, Possibilistic c-means (PCM), Image Ratioing, Fuzzy Error Matrix (FERM). BSTRCT: Sub-pixel base igital image classification outputs from coarse spatial resolution remote sensing images can be closer to groun information as compare to har classification outputs. This has been proven from work publishe by ifferent researchers in last ecae. In sub-pixel base classification output, each pixel represents a class in the form of en member or membership value. Fuzzy classification approach may be beneficial where a mixe pixel may be assigne multiple class memberships. Fuzzy approaches may be applie as supervise an unsupervise classification approaches. With ifferent isasters coming aroun the worl, isaster management professionals may be intereste to extract only one lan cover class of interest affecte in isaster from remote sensing ata. In conventional approach for extracting lan cover classes; all the classes present in the area have to be ientifie first, then single class of interest can be taken out from the classifie ata. The conventional classification approaches will give error if all the classes present in the area have not been ientifie properly. eural network classifier can be traine to extract single class of interest from remote sensing ata. The successful implementation of neural network classifier epens on a range of parameters relate to the esign of the neural network architecture. Moreover, neural networks are very slow in the learning phase of the classification, which is a serious rawback when ealing with images of very large size. Further, besies their complexity, neural networks are restricte to the analysis of ata that is inherently in numerical form. In this work it has been trie to apply fuzzy set theory base sub-pixel classifiers for extracting single class of interest. The use of fuzzy set base classification methos in remote sensing have evoke growing interest for their particular value in situations where the geographical phenomena are inherently fuzzy. While stuying Fuzzy c-means (FCM) as well as Possibilistic c-means (PCM) for extraction of single class of interest, it has been foun that PCM can be use for single class extraction of interest. In this work water class has been ientifie for single class extraction. well known problem of merging of shaow pixels with water class was encountere in this work. To overcome this problem, remote sensing sensor inepenent class base ratioing ata has been generate. Class base ratioing ata has been taken as input in PCM classifier, an water class has been ientifie at sub-pixel level. The input ata was use from IRS-P6 LISS-III sensor an testing ata as fraction image generate from IRS-P6 LISS-IV sensor. The output from this work was evaluate using fuzzy error matrix (FERM) an foun overall accuracies 93.4% an 95.3% respectively for both sub scenes. For this work, a moule in inhouse SMIC package was evelope in JV environment by the authors. It has been observe while incorporating class base ratioing ata in PCM base sub-pixel classifier, effect of pixels having varying illuminations ue to shaow within class was minimize (Figure ). It was also observe from this approach that shaow pixels were not mixing with water class pixels.. ITRODUCTIO Digital image classification is a funamental image processing operation to extract lan cover information from remote sensing ata an it assigns a class membership for each pixel in an image. Lan cover information can be extracte using crisp as well as fuzzy classification approaches. In a crisp classification, each image pixel is assume to be pure an is classifie to one class. Often, particularly in coarse spatial resolution images, the pixels may be mixe containing two or more classes. Fuzzy classifications may be beneficial where a mixe pixel may be assigne multiple class memberships. Fuzzy approaches may be applie as supervise an unsupervise classification approaches. To ientify a feature of interest not only have to classify iniviual pixels as belonging to a specific class such as soil, vegetation or water but also ientify a set of such pixels as a part of the feature. In number of applications there may be requirement for extracting only one lan cover class from remote sensing multi-spectral images, at sub-pixel level. In the past many learning algorithms le neural network has been use for performing classification to extract lan cover class from remote sensing image an it can be use for single class extraction at sub-pixel level (ziz, 004). For example, feature extraction for multi-source ata classification with artificial neural networks (Benetsson an Sveinsson, 997), the Hopfiel neural network as a tool for feature tracking an recognition from satellite sensor images (Côté an Tatnall, 997), remote sensing image analysis using a neural network an knowlege-base processing (Murai an Omatu 997). The successful implementation of neural network classifier epens on a range of parameters relate to the esign of the neural network architecture (rora an Fooy, 997). Moreover, neural networks are very slow in the learning phase of the classification, which is a serious rawback when ealing with images of very large size. Further, apart from their complexities, * Corresponing uthor 57
The International rchives of the Photogrammetry, Remote Sensing an Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 008 neural network are restricte to the analysis of ata that is inherently in numerical form. In this paper fuzzy logic base algorithm, which is inepenent of statistical istribution assumption of ata, has been stuie to extract single lan cover class from remote sensing multispectral images. s the images are affecte, by unesirable effects on the recore raiances cause by variations in topography has been remove by class base sensor inepenent ratioing at preprocessing stage. Possibilistic c- Means algorithm for this work has been implemente in such a manner that remote sensing image from any sensor can be use for single class extraction. For this work LCM: utomatic Lan Cover Mapping moule has been use from JV base in-house image processing package, name as SMIC: Sub-Pixel Multi-Spectral Image Classifier (Kumar et al., 006).. IMGE RTIOIG D CLSSIFICTIO PPROCHES The process of iviing the pixels in one image by the corresponing pixels in a secon image is known as ratioing. It is one of the most commonly use transformations applie to remotely sense images. There are two reasons why this is so. One is that certain aspects of the shapes of spectral reflectance curves of ifferent earth-surfaces cover types can be brought out by ratioing. The secon is that unesirable effects on the recore raiances, such as that resulting from variable illumination cause by variations in topography, can be reuce. In this work class base sensor inepenent ratioing has been applie by the following function; Maximum ( I ) Ratio Im age ( I ) = Minimum ( I ) Where I is class of interest; Class base ratio image of each class was use as an input in fuzzy set theory base classifiers. From the escription of the fuzzy logic base sub-pixel classification algorithms, Fuzzy c- Means (FCM) an Possibilistic c-means (PCM) (Krishnapuram an Keller, 993) approaches have been stuie. Fuzzy c-means (FCM) is basically a clustering technique where each ata point belongs to a cluster to some egree that is specifie by a membership grae, an that the sum of the memberships for each pixel must be unity (Bezek, 98). This can be achive by minimizing the generalize least - square error objective function; = c m J m ( U, V ) ( μ X v i = j = ij ) i j () subject to constraints; c μ = j = ij for all i μ > i = ij 0 for all j 0 μ ij for all i, j where Xi is the vector enoting spectral response of a pixel i, vj is the collection of vector of cluster centers of a class j, μij is class membership values of a pixel, c an are number of clusters an pixels respectively, m is a weighting exponent (<m< ), which controls the egree of fuzziness, Xi v j is the square istance ( ij ) between X i an v j, an is given by; T ij = X i v j = ( X i v j ) ( X i v j ) () where is the weight matrix. mongst a number of -norms, three namely Eucliean, Diagonal an Mahalonobis norm, each inuce by specific weight matrix, are wiely use. The formulations of each norm are given as (Bezek, 98), Eucliean orm = I Diagonal orm = D j Mahalonobis orm = C j where I is the ientity matrix, Dj is the iagonal matrix having iagonal elements as the eigen values of the variance covariance matrix, Cj given by; T C j = ( X v X v i i = j )( i j ) (3) In this research work value of weighting exponent m has been taken as. an Eucliean orm of weight matrix has been taken, as it givens maximum classification accuracy compare to other weighte norms (ziz, 004). The class membership matrix μij is obtaine by; μ ij = (4) ( m ) c ij k = where, = c ; j = ij 58
The International rchives of the Photogrammetry, Remote Sensing an Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 008 The original FCM formulation minimizes the objective function as given in equation (). subject to c μ = for all i j = ij But in PCM one woul le the memberships for representative feature points to be as high as possible, while unrepresentative points shoul have low membership in all clusters (Krishnapuram an Keller, 993). The objective function, which satisfies this requirement, may be formulate as; c m c m J m ( U, V ) = ( μ X v + i = j = ij ) i j η j = j ( μ i = ij ) (5) Subject to constraints; max μij > 0 for all i j μ > i = ij 0 for all j vectors of lan cover classes ( = c ). When j = ij extracting only one lan cover class, in that case = ij an than μ ij for all features becomes one. This conclues that all features in a remote sensing multi-spectral image belongs to one class, which is not the case. While working with PCM algorithm for extracting single lan cover class it behaves as follows; = ij while extracting single lan cover class; an μ ij = for class features from equation (4) an m = i = μ ij in equation (6) i ij So, η j = K = from equation (6). n now from equation (7) μ ij will be calculate; This inicates that possibilistic view of the membership of a feature vector in a class has nothing to o with its membership in other classes (Krishnapuram an Keller, 993). 0 μ ij for all i, j here μ ij is calculate from equation (4). In equation (5) where ηj is the suitable positive number, first term emans that the istances from the feature vectors to the prototypes be as low as possible, whereas the secon term forces the μij to be as large as possible, thus avoiing the trivial solution. Generally, ηj epens on the shape an average size of the cluster j an its value may be compute as; m μ = i = ij ij η K j (6) m μ i = ij where K is a constant an is generally kept as one. fter this, class memberships, μij are obtaine as; μ ij = (7) ( m ) + ij η j 3. SUB-PIXEL UTOMTIC SIGLE LD COVER CLSS EXTRCTIO For extracting lan cover classes, FCM is epene upon number of lan cover classes to be extracte from remote sensing multi-spectral image. This can be seen from membership values generate from equation 4, are epene upon summation of istances of unknown feature to mean 4. TEST DT USED Test ata sets for this work have been acquire from WIFS sensor of IRS-P6 satellite over two ifferent sites of Inia (figure ) le; san reservoir (Dehraun District) (Center coorinates 77 O 40 05.36 E, 30 O 6 08.00, cquisition ate 4 th ovember 003) an Rana Pratap Sagar Dam (Kota Distirct, Rajasthan state) (Center coorinates 4 O 47 30.8, 75 O 34 4.77 E, cquisition ate 8 th February 004). ll the ata sets were geo-reference an all the four bans of these ata sets were use with spatial resolution of 56 m. Major lan cover classes present in these area are water, forest, fallow lan, san, crop lan an settlement. The one class of interest, namely, water was stuie for iscrimination with backgroun. Separate sample ata sets were use as training as well as testing stage. The size of training ata use for supervise subpixel classification approach was approximately equal to 0n (Jensen, 996), were n is imension (number of bans) of ata use. In this work it has been trie to stuy the performance of PCM algorithm for extraction of single lan cover class at subpixel classification with small training ata set. The numbers of pixels use at training stage are given in table. These pixels were collecte from Rana Pratap Sagar Dam image only an same training ata set was use for other image also. For testing the classification accuracy reference fraction images were use from LISS-III sensor of IRS-P6 satellite for both the ata sets of same ates as of WIFS ata (figure ). total of 500 testing pixels for class water were ranomly selecte, from corresponing outputs an their reference images respectively, which are significantly larger than the sample size of 75 to 00 pixels per class as recommene by Congalton (99) for accuracy assessment purpose. The accuracy assessment of subpixel classification output has been one using Fuzzy Error Matrix (FERM) (Binaghi et al., 999). 59
The International rchives of the Photogrammetry, Remote Sensing an Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 008 5. DISCUSSIO D COCLUSIO a) san reservoir b) Rana Pratap Sagar Dam Figure. FCC images of test sites from WIFS sensor Sl. o.. Type of ata use o. of pixel use Training ata Water 4 Table. Details of training ata set a) san reservoir In this paper, we have presente a special form of PCM algorithm while extracting single lan cover class from multispectral remote sensing images. For this the LCM moule from, SMIC: Sub-Pixel Multi-Spectral Image Classifier package (Kumar et al., 006) has been use. The LCM moule has capability to process multiple multi-spectral images for single lan cover class extraction at sub-pixel level using supervise approach. The moule etects multi-spectral images of same bans by ecoing the heaer files in generic BIL format an then reas all the images. ll these images can be processe using PCM single lan cover class extraction algorithm using training ata. t the output, two options are available, first, where a threshol is provie to ientify pixels having membership greater than some specific membership value of that single lan cover class or secon option is without a threshol. While using PCM approach for single lan cover class extraction, its variables appear ifferently in comparison of multiple lan cover class extraction. In single lan cover class extraction case; = ij ; an μ = for class features from equation (4), ij i ij so, η j = K = from equation (7). Which is not the case while extracting multiple lan cover class extraction using PCM metho. From table, it is clear that user, proucer as well as over accuracy of single lan cover class water were foun very high. s in figure 3, upper stream of san reservoir membership values were foun in the range of 0.890 to 0.996. In own stream of river membership values for lan cover class water were of the range 0.84. In another ata set; Rana Pratap Sagar Dam scene, upper stream of water boy have membership value in the range 0.996. Thus, it can be conclue that water quality in both the two sites is more or less same. Membership values of the fraction images may be correlate with groun ata to extract turbiity of water, lentic an lotic water boies, etc. Due to sub-pixel classification water component in some of the other lan cover classes were foun of small membership values. This inicate that sub-pixel classification is very much goo to classify remote sensing multi-spectral ata having vagueness in the ata sets. While incorporating class base ratioing ata shaow effect has been minimize in classification output. Due to class base rationing ata shaow areas i not merge with class water. Further, LCM moule approach will be very much helpful were there is an urgent nee of processing multiple multi-spectral images for single class extraction at sub-pixel level, specifically in floo prone areas an risk analysis. Presently this moule has limitation to incorporate training ata from ifferent input multiple multi-spectral images. To incorporate training ata from ifferent input multiple multispectral images will efinitely improve the sub-pixel classification accuracy of single lan cover class extraction, which nees to be examine further. b) Rana Pratap Sagar Dam Figure. Reference FCC images of test sites from LISS-III sensor 530
The International rchives of the Photogrammetry, Remote Sensing an Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 008 Benetsson J.. an Sveinsson J. R.,997, Feature extraction for multisource ata classification with artificial neural networks. International Journal of Remote Sensing, Vol. 8, no. 4, 77 740. Bezek, J. C.,98, Pattern Recognition with Fuzzy Objective Function lgorithms, Plenum, ew York, US. a) san reservoir b) Rana Pratap Sagar Dam Binaghi, E., Brivio, P.., Chessi, P. an Rampini,.999. fuzzy Set base ccuracy ssessment of Soft Classification. Pattern Recognition letters, 0, 935-948. 0 μ μ = Membership Value Congalton, R. G.99, review of assessing the accuracy of classifications of remotely sense ata. Remote Sensing of Figure 3. Single class water extracte from test ata sets using Environment, 37, 35-47. WIFS Images Côté S. an Tatnall. R. L.,997, The Hopfiel neural network Sub-Pixel Classification ccuracy Data of san reservoir User s ccuracy (%) Water 93.3 95.4 Proucer s ccuracy Water 93.7 97. (%) Overall ccuracy (%) 93.7 97. Data of Rana Pratap Sagar Dam Table. User s, Proucer s an Overall ccuracies for ifferent ata sets. REFERECES rora M. K. an Fooy G. M.,997, Log-linear moeling for the evaluation of the variable affecting the accuracy of probabilistic, fuzzy an neural network classifications. International Journal of Remote Sensing 8(4): 785-798. ziz, M..004. Evaluation of soft classifiers for remote sensing ata, unpublishe Ph.D thesis, IIT Roorkee, Roorkee, Inia. as a tool for feature tracking an recognition from satellite sensor images. International Journal of Remote Sensing, Vol. 8, no. 4, 87 885. Jensen, J. R.996. Introuctory Digital Image Processing: Remote Sensing Perspective n eition (US: Prentice Hall PTR). Krishnapuram R. an Keller J. M.,993, possiblistic approach to clustering, IEEE Transactions on Fuzzy Systems,, 98 08. Kumar,., Ghosh, S. K., an Dahwal, V. K.,006, Sub-Pixel Lan Cover Mapping: SMIC System. ISPRS International Symposium on "Geospatial Databases for Sustainable Development" Goa, Inia, September 7-30, 006. Murai H. an Omatu S.,997, Remote sensing image analysis using a neural network an knowlege-base processing. International Journal of Remote Sensing, Vol. 8, no. 4, 8 88. a) Raw Image b) Single class without class base ratioing input c) Single class with class base ratioing input Figure : Single class water extracte using PCM classifier 53
The International rchives of the Photogrammetry, Remote Sensing an Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 008 53