ADVANCED MATHEMATICAL TOOLS FOR SAR IMAGE ANALYSIS

Abstract

scaling varies with the variation of pixels, and consequently, the satellite images follow the multifractal behavior. The multifractal dimension has been estimated for 3 generalized dimensions. These images are classified with the overall classification accuracy upto 71% whereas that with using fractal dimension alone comes to be nearly 64%. The objective of Chapter 4 is to analyse the fractal dimension with roughness parameters for labeling different clusters for the classes. We have proposed a scheme to label the clusters. Since it is difficult to get the real SAR data in large amount, a set of 7 simulated SAR images have been generated for different 1 and s values. The _value_of_t_spans between 0.5 to 15.0 cm. and that for s between 0.1 and 5.0 cm. The scattering coefficient is computed for various 1 and s values which is used to compute the D of the simulated images. In order to simulate the SAR images, various sensor parameters and the surface roughness parameters have been used. The backscattering coefficient (e) has been estimated using well known theoretical models and based on the a° values, SAR images are generated. The simulated images have been analysed for textural measures using fractal dimension (D) and lacunarity (L) . Based on these two measures, the images are clustered and this information is used for labeling. The developed scheme has been successfully applied on real SAR images. The overall classification accuracy for simulated images is obtained up to 85% whereas for SAR images up to 78% overall accuracy is achieved. Finally, a comparison of fractal feature based classification with traditional contextual classifier, i.e., majority filter has been done. Chapter 5 deals the study of soft computing approach for classification of SAR images using textural features. Particle Swarm Optiinization (PSO), Self-Organizing Maps (SOM) and Support Vector Machines (SVM) techniques have been tested for obtaining the classification accuracy with D maps. A PSO based clustering has been performed on D maps which is helpful in achieving higher accuracy of textural classification. SOM is unsupervised neural network approach of clustering, which is based on competitive learning. SVM is a supervised method, which is also used for clustering of fractal maps. The overall accuracy using PSO is 66.88%, using SVM is 68.84% and with SOM is 65.36% whereas for K-means accuracy is approximately 60%. ix A study of retrieval of surface parameters with the use of fractal dimension is implemented in chapter 6. The objective of this chapter is retrieval of surface roughness parameters using fractal dimension. In this chapter the relation among D, I and s has been analysed and an empirical relation has been proposed to retrieve 1 and s using D. In order to devise a direct relationship among D, 1 and s, various synthetic surfaces have been generated using 1 and s and for these surfaces, D is estimated. Using a non-linear multiple regression, the relation in D, 1 and s is established. The procedure is repeated for noisy surfaces, where Gaussian and speckle noise are added to the generated surfaces and then for noisy surfaces the value of D is estimated. The relation has been devised using the generated surfaces without noise and tested on the noisy images and it is observed that the relation represents both noisy and non-noisy images very well. In order to retrieve the roughness parameters (1, s), given the value of D, a look up table approach is followed because an analytic unique solution is not available. Chapter 7 concludes the thesis with its contributions and future