Robust spatially regularized sparse unmixing of hyperspectral remote sensing images with spectral library pruning

Abstract

With the widespread use of endmember spectral libraries, sparse regression techniques have become crucial in hyperspectral image unmixing. Recently, considering spatial information in sparse unmixing frameworks has become increasingly important, as it enhances the accuracy of mixed pixel decomposition. However, challenges such as spectral mismatch due to high correlation among endmember spectra and susceptibility to complex noise, like Gaussian and sparse noise, affect unmixing accuracy. To address these issues, this paper introduces the robust spatially regularized sparse unmixing algorithm with spectral library pruning (RSUSLP). The algorithm decomposes the unmixing process into multiple layers and prunes the spectral library at each layer to alleviate spectral mismatch. It models sparse noise within the unmixing framework and utilizes spectral weighting in conjunction with spatial weighting to increase row sparsity and spatial correlation, thus improving robustness. The optimization problem described in the algorithm is solved using the alternating direction method of multipliers (ADMM). As illustrated by experimental results from both simulated and real hyperspectral data, RSUSLP significantly surpasses current sparse unmixing methods by reducing spectral library interference and effectively handling noise, thereby enhancing the accuracy and performance of mixed pixel decomposition.