SAR image classification post-processing with multiscale complementary Gaussian kernel weighting

Abstract

Classification errors may occur both in pixel- and spatial-based classification methods at the pixel level for synthetic aperture radar (SAR) images. In this study, a classification post-processing method is proposed by utilizing complementary Gaussian kernel weighting (CGKW) for regularization of classification errors on the classified SAR images. To demonstrate the validity of the proposed method, the uniform kernel weighting (UKW), the Gaussian kernel weighting (GKW), Markov random fields (MRF) and total variation-L1 (TV) methods are also presented for comparison purposes. The proposed approach is a combination of filtering- and relearning-based classification post-processing method that can be applied to SAR images. In the proposed framework, class probabilities for every pixel are initially obtained by a convolutional neural network. Afterward, classification results are updated again using the selected weighted averaging method and neighboring classification probabilities, and then, the results of UKW, GKW, CGKW, MRF and TV are compared. Experimental results prove that the proposed CGKW method improves the final classification accuracy better than other comparison methods, and the resulting classification difference between the UKW, GKW, MRF and TV methods is statistically significant according to McNemar's statistical significance test.