SVM Enhancement with Application to SAR Imagery Classification

This paper investigates enhancing the performance of support vector machines (SVMs) in the application of synthetic aperture radar (SAR) imagery classification. The approach is to replace the conventional Euclidean distance in the SVM kernel with a new similarity measure that is less sensitive to perturbations. Same-target SAR images show perturbations, in part due to the presence of speckle and in part due to small variations in radar depression angle and target orientation. It is expected that SVMs with the proposed new kernel will outperform those with the conventional Euclidean kernel. Experimental results are presented to validate this expectation for both batch and iterative implementations of SVMs. The paper also argues that the proposed approach is well-founded theoretically by demonstrating that the new kernel is still a Mercer kernel