Sparse-SpatialCEM for Hyperspectral Target Detection

Abstract

The constrained energy minimization (CEM) algorithm is widely used for target detection in hyperspectral imagery. This method, as well as most target detection algorithms, focuses on the use of spectral information and neglects the spatial information embedded in images. In real hyperspectral images, it is usual that targets of interest only occupy a minor portion of the pixels, and an object may consist of multiple consecutive pixels in space. Considering these facts, we propose a novel constrained detection algorithm, referred to as Sparse-SpatialCEM, to simultaneously force the sparsity and spatial correlation of the detection output via proper regularizations. Several algorithms, including the CEM, SparseCEM, and constrained magnitude minimization algorithms, are limiting cases of the proposed framework. The formulated problems are solved by using the alternating direction method of multipliers. We validate the proposed algorithms and illustrate its advantages via both synthetic and real hyperspectral data.