Specific Spectral Target Detection for Multispectral Images via Target-Focused Spectral Super-Resolution

Abstract

Spectral target detection using spectral information provided by hyperspectral (HS) images has been deeply studied. However, due to low spatial resolution and difficulty in obtaining HS images, spectral target detection based on it faces extremely serious problems of small-scale and mixed spectra. To address this problem, taking the more easily obtained high-spatial-resolution multispectral (HMS) image as an appropriate input, this article proposes a specific spectral target detection method through target-focused spectral super-resolution (SSR). Specifically, by taking the given target spectrum and the spectral library as priors, a target-focused SSR model under the sparse representation framework is proposed first, to enrich the spectral information of the HMS image, and to accurately reconstruct the corresponding high-spatial-resolution HS image, especially for the target area. Then, a target-specific band selection strategy is designed, to extract the most distinguishable spectral bands against background, which can enhance the separation between the target and background and help to reduce the false alarm rate of the detection. Finally, a background separation-based spectral target detection method for the selected bands is proposed, to locate the spectral targets directly by using the optimized target sparse coefficient matrix. Experimental results on four different datasets show that, our proposed method achieves the best target detection performance in comparison to other relative state-of-the-art methods, and can even efficiently handle the detection of subpixel-level spectral targets through this unmixing-like spectral dictionary expression.