Complex matrix and multi-feature collaborative learning for polarimetric SAR image classification

Abstract

Nearest-regularized subspace (NRS) algorithm is an effective tool to obtain both accuracy and speed for PolSAR image classification. However, existing NRS-based methods can only learn the polarimetric feature vector well, but ignore the PolSAR complex matrix structure and channel correlation. So, how to collaboratively learn the original complex matrix and multiple features to improve the classification accuracy is a key problem. Besides, speckle noises are also the main factor of causing misclassification. To address these limitations, two novel methods are proposed. Firstly, a superpixel-based Riemannian NRS(SRNRS) method is proposed, which can not only learn complex matrix structure by Riemannian metric, but also reduce the speckle noises and computing time by superpixels. Then, a superpixel-based collaborative learning method(CM_SJNRS) is proposed, which integrates the complex matrix and multiple features into the NRS classification framework for the first time. Coupled dictionaries and different metrics are designed for two kinds of feature spaces respectively, and then a collaborative learning model is developed to fuse them. Experimental results demonstrate the proposed SRNRS method can reduce both the speckle and computing time, and the proposed CM_SJNRS method can improve classification performance by fusing two types of features.