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

摘要: 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 post-processing method is proposed by utilizing complementary Gaussian kernel weighting (CGKW) regularization of on classified SAR To demonstrate validity method, uniform (UKW), (GKW), Markov random fields (MRF) total variation-L1 (TV) are also presented comparison purposes. The approach combination filtering- relearning-based that can be applied to framework, class probabilities every initially obtained convolutional neural network. Afterward, results updated again using selected weighted averaging neighboring probabilities, then, UKW, GKW, CGKW, MRF TV compared. Experimental prove CGKW improves final accuracy better than other methods, resulting difference between statistically significant according McNemar’s statistical significance test.