Pansharpening Based on Low-Rank and Sparse Decomposition

This paper explores the low-rank and sparse (LRS) decomposition to solve the problem of pansharpening. By exploiting the significant correlation among the multispectral (MS) image bands, the LRS decomposition is employed as a decorrelation tool, from which the spectral and spatial informations in MS images can be separated. Based on Go Decomposition (GoDec), we provide two contributions. An LRS-based pansharpening method (i.e., ImPCA) which is designed in terms of component substitution (CS) concept is given. In order to improve the performance of ImPCA by reducing the spectral distortion which is characterized by the color or radiometric changes in the pansharpened images, the local dissimilarity between MS and panchromatic (PAN) images is taken into account by exploiting the context-based decision (CBD) model. Experimental results with both simulated and real data demonstrate that after the local dissimilarity is considered, the quality of the pansharpened images is significantly improved. The improved version of ImPCA is comparable with other popular methods.