Low-Rank Decomposition Model for Adaptive Identification of Similar Neighboring Pixels in Hyperspectral Images

Abstract

In many hyperspectral image processing tasks such as classification, unmixing, and target detection, jointly using the pixels in an image patch can generally improve the performance. In this letter, we propose using a low-rank decomposition model to analyze the image patch. The image patch is decomposed into the sum of a low-rank matrix, a sparse matrix, and a bounded matrix using a convex optimization technique. In the obtained low-rank matrix, the pixels that are similar to the central pixel in the image patch would be codirectional to it. We applied the proposed model in the collaborative hyperspectral image classification to evaluate its performance. Experimental results on a real hyperspectral scene demonstrate that using only the similar neighboring pixels in the collaborative hyperspectral image classification can effectively improve the performance, and the classification performance is not sensitive to the image patch size.