Spatial Structural Priors for Sparse Unmixing of Remotely Sensed Hyperspectral Images

Abstract

As spectral libraries continue to expand, sparse unmixing has become essential for effectively interpreting mixed pixels in remotely sensed hyperspectral data. Integrating spatial information into sparse unmixing is very important to enhance unmixing performance. However, hyperspectral images are often affected by noise interference (encompassing both Gaussian and sparse noise). Such interference may lead to pixel degradation, resulting in decreased signal-to-noise ratio (SNR). Existing spatially enhanced sparse unmixing algorithms struggle to interpret low SNR hyperspectral data. To address this challenge, we propose a new sparse unmixing algorithm based on spatial structure prior features (SSPSU) that introduces a multidimensional gradient total variation regularization to explore the inherent low-rank structure, local smooth features, and sparse modules within the hyperspectral image. In addition, by incorporating spatial weighting factors at the superpixel scale, it further promotes spatial correlation among local pixels. The joint application of these two regularization terms effectively mines the spatial structural prior information within the image, mitigating the effects of various kinds of noise on unmixing and enhancing the algorithm's capacity for estimating land cover abundances. Our experimental validation with simulated and real data sets reveals that the proposed SSPSU algorithm can effectively counteract the interference of complex mixed noise and outperform other state-of-the-art unmixing algorithms in terms of accuracy and stability of abundance estimation, particularly under low SNR conditions.