STUDY AND ANALYSIS OF SAR IMAGES USING UNSUPERVISED CLASSIFICATION

Abstract

Land cover and land use classification on a huge scale, e.g. national or continental scale, have become more and more important in wide range of applications like urban planning, farming, study of human impacts on environment, and study of damages caused by natural disasters like landslides, hurricanes, erosion or earthquakes. Earth observation using satellite imagery is a challenging tool for land cover monitoring that may offer a wide coverage efficiently. Automated classification of image into various land covers like water, urban and agriculture is an imminent necessity for large scale landscape monitoring. Land cover classification of the image is the process of segmenting the image into clusters according to desired parameters and labeling each cluster as various land covers according to characteristics of pixels within clusters. Therefore, there is a need to analyze satellite images and explore possibilities of using these images for classification purposes more efficiently. Radar sensors have specific advantages for land cover classification due to their operating frequency range: (1) Sensitive to surface characteristics like surface roughness and dielectric property, (2) Independent of sun illumination, and (3) Capable of providing all time (day and night) and all weather acquisitions. It's another advantage is its relatively high spatial resolution. Therefore, by seeing the advantages of Synthetic Aperture Radar (SAR) image, it is required to explore its applications for various land cover classification. Land cover classification of SAR image can be approached as (i) Clustering the pixels and labeling clusters into various land covers like water, urban and agriculture, and (ii) classifying land cover in to 'change' and 'no change' areas by comparing SAR image that are obtained at different time with quantitative assessment of changes. For classification of single polarized SAR image, backscattering coefficient and texture are the only information available. Hence, analyzing single polarized SAR image is still a big challenge. The present dissertation work is an effort to obtain unsupervised land cover classification in one hand and on the other hand, to classify 'change' and 'no change' areas with quantitative measurement of changes. The main aim of this dissertation is n to study and develop suitable algorithm for unsupervised classification of single polarized SAR image by (i) identifying textural features that should be most effective for land cover discrimination and devising techniques to measure features, (ii) defining algorithm to integrate the features for clustering, (iii) labeling of clusters with minimum apriori information and (iv) application of image analysis and interferometry for quantitative and qualitative analysis of changes on land surfaces (i.e., subsidence). To test and critically analyze the various land cover classification algorithms, three imagesets of ERS-2 SARC-band images (5.3 GHz frequency andVVpolarization) at a spatial resolution of -12.5 m are used. The test site selected for this study is Haridwar region, India (Latitude: 77° 51'E to 77°55'12"E; Longitude: 29°54'N and 29°50'50" N). New Orleans city of USA (Latitude: 30°3'27"N to 29°54'43"N; Longitude: 90°15'12"W to 89°53'36"W) is taken as the study area for application of image analysis and interferometric techniques to observe qualitative and quantitative changes on land surface (i.e., subsidence). For this purpose, 84 Single look complex (SLC) RADARSAT-1 images were acquired from 15 April, 2002 to 15 March, 2007 with 24 day repeat interval. First chapter presents the introduction of the thesis which includes motivation, major research gaps, and details about the study area and satellite image used. In chapter II of thesis, brief review of related works is presented. Chapter III explores the analysis of role of various intensity and textural measures for their discriminative ability for land cover classification (i.e., water, urban and agriculture areas) of SAR data. In this chapter, suitable algorithm by which various texture measures can be combined for unsupervised SAR data classification was proposed and the study of their combined effect on classification accuracy was presented. Texture, being an important property for describing various land covers in SAR data has led to analysis of various texture measures for its classification. Till now, texture measures have been applied individually or combined by K-means for classification purposes but gave less accuracy (approximately 50%, Dekker, 2003). There is a need to study the role of various texture measures in classification and to improvise the algorithm of combining texture measures. In this chapter, various texture measures namely mean, variance, semivariogram, lacunarity, weighted rank

fill ratio and wavelet components, and their combined effects for classification, were analyzed by applying them to synthetic images as well as to SAR data. The focus of this chapter was to increase classification accuracy by fusion of maximum information obtained by single polarized SAR data. Individual texture measures have been analyzed for improving classification accuracy in one hand, whereas in another hand, to obtain combined effect of these textural measures on classification accuracy, PCA has been applied and selected principal components were used for further classification. Because it is well known that PCA has its own advantage of fusing information from number of various input features, and giving output in terms of eigen vectors which are orthogonal, capturing information from the input features (Bajwa and Hyder, 2005; Joliffe, 2002). It was observed that feature set comprising of mean, variance, wavelet components, semivariogram, lacunarity and weighted rank fill ratio provided good classification accuracy up to 90.4% than by using individual textural measures (with individual texture measure, maximum classification accuracy achievable is approximately 76%) and this increased accuracy justified the complexity involved in the process (Chamundeeswari et al., 2006a; 2006b; 2009). An effort of devising adaptive unsupervised classification technique with the help of wavelets for multi resolution analysis of SAR image has been explored in chapter IV. Till now, most of the developed research works requires either ancillary image such as elevation model, ground truth or involves user interaction to decide parameters involved in classification algorithm. Hence, there is a need to develop user independent algorithm with minimum apriori information. For multi resolution analysis, wavelet transform is an excellent tool (Mallat, 1989) and its ability to capture different textures in SAR image has not yet been fully explored (Leporini and Pesquet, 1997). Therefore, in this task, an attempt to critically analyze multi resolution texture by wavelet decomposition has been carried out. Four band wavelet decomposition is applied on SAR image and textural features are extracted from wavelet coefficients corresponding to each of these sub-bands. Integrated feature vector corresponding to every sub-band chosen, is obtained by integrating textural features corresponding to that band and backscattering coefficient. Adaptive neuro-fuzzy algorithm ranks and selects important feature vectors for classification process. This algorithm helps in removing redundant feature bands and choosing relevant feature vectors for classification. Then, K-means classification is applied on

chosen feature vectors. The proposed classification process involves three user defined parameters namely (i) window size of local estimator 1, (ii) window size of local estimator 2 (Local estimator 1 & 2 are used to compute textural features from sub-bands obtained by wavelet decomposition at level 1 and 2 respectively), and (iii) number of feature vectors to be chosen for K-means classification from adaptive neuro-fuzzy algorithm. To make the algorithm adaptive and optimum, user dependency has to be removed and above mentioned user defined parameters are to be identified for maximum achievable classification accuracy. For this purpose, an algorithm is proposed for classification accuracy in terms of the parameters involved in segmentation process. This will be very helpful to develop automated land cover monitoring system with SAR image, where optimized parameters are to be developed only once and these parameters can be applied to SAR image of the scene year after year for a particular region. Single polarized SAR image is classified into water, and urban areas using the proposed method and overall classification accuracy is obtained in the range of 85.92%-93.70% by comparing with ground truth image (Chamundeeswari et al., 2007a; Chamundeeswari and Singh, 2006). Labeling of different clusters on SAR image is quite a challenging task. Therefore in chapter V, the task of how to label these clusters is critically analyzed and studied. Earlier researchers have used backscattering coefficient for labeling the different clusters. But it is difficult to get unique value of backscattering coefficient for different clusters (i.e., major land covers namely water, urban and agriculture areas). Therefore, roughness parameters, along with backscattering coefficient may be used for labeling of clusters in SAR image. In this chapter, we have proposed an algorithm that includes surface roughness as one of the criteria along with backscattering coefficient, for labeling various clusters. Surface roughness measures RMS height,'^' (i.e., vertical roughness) and correlation length 7' (i.e., horizontal roughness). This surface roughness may directly indicate the various major land covers like water, urban and agriculture areas. First to check the sensitivity of textural measures on roughness, effect of eight different textural measures namely of mean, variance, wavelet components, semivariogram, lacunarity and weighted rank fill ratio on surface roughness is critically analyzed. This analysis has been carried out on developed synthetic images (approximately 300 images for various combinations of s and I). Semivariogram, weighted rank fill ratio and wavelet components are found to be suitable texture measures for retrieving surface roughness parameters. An empirical relation has been proposed for retrieving surface roughness from these texture measures. With the help of backscattering coefficient and surface roughness parameters retrieved from texture measures, clusters can be labeled. The proposed algorithm is applied on SAR image to label land covers namely water, urban and agriculture areas. Land cover labels obtained by proposed method are found to be in good agreement with topographic information. Unsupervised classification of 'change' and 'no change' pixels and quantitative analysis of changes, i.e., subsidence are taken as next task in chapter VI. For this purpose, various image analysis approaches have been studied and critically analyzed for classifying 'change' and 'no change' pixels, whereas, for quantitative analysis of changes, i.e., subsidence, interferometric approach has been applied. To classify the areas of 'change' and 'no-change', various image analysis approaches like image differencing, minimum ratio detector, Correlation coefficient technique and integrated intensity, texture and orientation difference maps have been applied. Each approach has its own advantages and disadvantages. Therefore, it is difficult to get same area of 'change' and 'no-change' with all these approaches. To get the common area of 'change' and 'no-change', an intersection operator method is proposed. This method is used to obtain more reliable set of 'change' and 'no-change' pixels. By studying the properties of pixels belonging to such change and no-change areas, Weibull probability distribution model is proposed. This model includes location, shape and scale parameter. It is found that this model best suit compared to others like lognormal, exponential or Gaussian distribution for the distribution of pixels on the basis of chi square test (Chamundeeswari et al., 2008a; 2008b). The results obtained by proposed methodology are quite encouraging. Repeat pass SAR interferometry is potentially a unique tool for precise generation of DEM and large coverage deformation tool. Selection of InSAR image pairs is very crucial and requires analysis of large image sets to identify suitable image pairs. The baseline distance, caused by drift in orbit between passes, provides different viewing angles required for getting interferogram. But if baselines are too large, the accuracy of D-InSAR will decrease since the removal of the topographic phase term can not be performed very accurately. Care should be taken that baselines between image pairs are not too large (<100m). Spatial overlap (>50%) and azimuth spectra overlap

(>90%) of the two images involved in generation of interferogram are also to be considered to get good quality interferogram and hence accurate deformation. In addition, SAR interferometry requires digital elevation models to map deformations occurred. Hence, in this study, D-InSAR using three pass interferometry, that does not require elevation model or any apriori information is used for subsidence measurement. The subsidence for major change areas in study site are calculated and found to be in the range of (0-16) mm per year. These results are found to be in agreement with earlier findings (Dixon et al., 2006). Dixon has used permanent scatterers to measure deformation which require DEM, while we have used differential InSAR approach avoiding the need for such additional information like elevation model of the study site. The quantitative analysis of subsidence with image analysis approach is very difficult. Therefore, an algorithm is proposed in this chapter by which more information about the changes like nature of subsidence (low, medium or high) could be retrieved from image analysis approach. From the Subsidence map obtained by DInSAR, pixels are categorized as low, medium and high subsidence pixels. To avoid cumbersome method of DInSAR, image analysis approach of image ratioing has been attempted to classify the image into low, medium and high subsidence pixels. For identifying low, medium and high subsidence areas, pdf is labeled for each type of subsidence (Chamundeeswari et al., 2008c, Singh et al., 2008a; 2008b). Results are quite in good agreement with the results obtainedby DInSAR. Only quantitative analysis cannot be done by this method and it only highlights low, mediumand high subsidence region. The results are validated with the ground truth survey undertaken in and around Haridwar region, India for unsupervised land cover classification. Ground truth survey of agriculture areas reported by Said (2006) is also taken as reference. For classifying 'change' and 'no change' areas from multi temporal SAR images, New Orleans city of USA is considered as study site and comparison of the results is done with the results published by Dixon et al., (2006) using interferometry based subsidence for the years 2002-2005. Finally in chapter VII, the contributions made in the thesis are summarized and scope of future work is outlined. The obtained analysis and results may give to various users to design monitoring system for land cover classification as well as classification of vii 'change' and 'no change' areas. This type of study and proposed algorithms will certainly enhance the analytical capability of applications of radar images for end users.