SVM Classifier Optimization using Genetic Algorithm for Classification of Polarimetric Synthetic Aperture Radar Imagery

Satellite image classification is considered as one of the most common approach for information extraction from remotely sensed data. With the advent of microwave sensors and taking into account the advantage of distinctive characteristics of the microwave range in electromagnetic spectrum extraction of different information in comparison to optical sensors are provided. Polarimetric information has significant implications for identifying different phenomena and distinguishing between them. In synthetic aperture radar imagery unlike hyperspectral imagery, where in spectral bands provide required features for pattern recognition process, we need to construct such features. Nowadays we can extract a wide range of features from polarimetric images using target decomposition theorem and SAR descriptors. In this paper at the first stage we try to extract features in three categories including original data features, decomposition features and SAR parameters. Then SVM algorithm with RBF kernel is used to classify polarimetric image. Due to the binary nature of support vector machines algorithm, the one against all approach is used to perform a multi-class classification. In this approach for m class m binary classifier are considered. In this study the genetic algorithm is used in order to calculate kernel parameters, feature space dimension reduction and selection of optimal features. In this study the superior performance of SVMs achieved by simultaneously optimization of SVMs parameters and input feature subset on Polarimetric imagery are demonstrated. The other point of the paper is higher accuracy of SVM classifier by kernel parameter selection using genetic algorithm and considering all the features in relation to optimal feature selection using genetic algorithm and kernel parameter selection using grid search. In another section of this study object based image analysis is used to compare the performance of SVM classifier in conjunction with genetic algorithm with OBIA. In OBIA approach, at the first stage, an image to be analyzed is segmented into individual image objects in an object based approach. The image pixels from the image are grouped to form the objects in a segmentation process. The created image objects should represent the objects in reality. In this research, multiresolution segmentation algorithm was used to create the image objects. By delineating objects from images, object based image analysis enables the acquisition of a variety of additional textural and spatial features, which are helpful in improving the accuracy of polarimetric image classification and also reducing the effect of speckle in PolSAR images by implementing classification based on image objects, and the textural information extracted from image objects. To extract optimal features, it is essential to use an appropriate analysis tool. For this purpose, in this paper SEaTH analysis tool was used. In this method, by using Jeffries-Matusita’s measure, the features are extracted as the optimal features in an appropriate separation of the probability distribution function for the training samples belonging to different classes. The proposed method was applied to the RADARSAT-2 imagery of an urban area in fine quad polarimetric mode. This imagery was selected in order to include a variety of land cover categories i.e. urban, water, bare soil and vegetation. The results demonstrated that the accuracy of OBIA approach is higher than support vector classification using grid search and all input features. Finally, we demonstrate that classification accuracies are significantly higher by simultaneously optimization of SVMs parameters and input feature subset on Polarimetric imagery.