Kernel Extended Local Tangent Space Alignment for SAR Image Classification

Abstract

This study proposes a novel local tangent space alignment (LTSA) variant, kernel extended (KE)-LTSA for synthetic aperture radar (SAR) image classification. It attempts on one hand to extract local geometric structures embedded in local neighbourhoods and on the other hand to maximize global interclass separability characterized by the overall distances among different classes. Moreover, it is formulated with kernel technique to obtain better performance than linear counterparts. Experimental results on the MSTAR database demonstrate that the proposed method can significantly improve the classification performance. Results also indicate the robustness when taking into account target variability and neighbourhood size.