Pansharpening through an improved sparsity-based fusion method

Abstract

ABSTRACTIn this paper, a patch-wise manner based on the sparsity is proposed to fuse a panchromatic (PAN) image and a low resolution multispectral (LMS) image. In the sparsity-based pansharpening methods, improving the training process of the dictionaries and sparse coefficients of the fused image, which is the main goal of this paper, have a significant impact on the fused results. In this paper, the fused image is obtained by minimizing the cost function which is obtained from incorporating a Markov random field (MRF)-based prior model into the maximum a posteriori (MAP) estimation. The contribution of this paper is twofold derived from our proposed prior model. 1) The prior model only involves the parts of the PAN information related to a considered band of the high-resolution multispectral image in the training process of the dictionary of the considered band. Not only does it improve the training of the dictionaries, but it also leads to finding more accurate sparse coefficients. 2) The high-frequency information of the PAN image is also involved in the training process as a separate term. This term decreases the spectral distortion by relieving the adverse effect of dissimilarity of the grey levels between the PAN and multispectral images on the fused image. The visual and quantitative comparison between the performance of the proposed method and eight well-known fusion methods on the Pleiades, QuickBird, and DEIMOS-2 data demonstrate the superiority of the proposed method.