Sparse Tensor Model-Based Spectral Angle Detector for Hyperspectral Target Detection

Abstract

In recent years, the hyperspectral target detection technique has received widespread attention, and various methods have been proposed. However, most of these methods suffer mainly from two problems. First, the methods perform detection based on the original image, and thus, the targets are easily contaminated by the complex background. Second, a large amount of prior information is usually required, but difficult to obtain. To solve these problems, we proposed a Sparse Tensor model-based Spectral Angle (STSA) detector for hyperspectral target detection. First, with a tensor decomposition model, we obtained a sparse tensor component separated from the original image. Based on the sparse tensor, the detection is less contaminated by the background, and the sparse tensor retains the spatial structure using the developed 3D tensor-based decomposition model. Second, to further enhance the performance, the projected spectral angle method was developed, which only requires a single target spectrum as prior information. Finally, to increase the separation ability between the target and background, we applied a statistical strategy to highlight the target and suppress the background. The experiments were carried out based on four public hyperspectral datasets. The results showed that the proposed STSA method is more accurate than 11 benchmark methods.