A novel method for multispectral image pansharpening based on high dimensional model representation

Abstract

Pansharpening methods are used to enhance the spatial resolution of a low resolutional multispectral (MS) image by fusing with a high resolutional panchromatic image (PAN). The main difficulty of pansharpening is avoiding spectral distortion while getting a sharpened MS image with high spatial resolution. Intensity-Hue-Saturation (IHS) based methods are applied to transform from color space to IHS and provide equalization of a PAN component with an MS image to eliminate distortion problems. However, most of the modified IHS methods still cause spectral distortion. To overcome this problem, a novel pansharpening method, based on Adaptive High Dimensional Model Representation is proposed in this article. HDMR is a well-known decomposition method for multivariate functions and data sets. The algorithm we propose includes three stages: the first stage is to obtain HDMR components of the MS image using the HDMR decomposition and then to use scaling factors to optimize the effects of the information the components hold. The second stage requires the calculation of some weighting factors in each band to minimize the spectral distortion. Computing the spatial details obtained from the difference between the PAN image and the Adaptive HDMR expansion of the MS image, and adding the difference to the MS image constitutes the third stage. Our proposed algorithm is easy to implement in pansharpening similar to component substitution (CS) based methods, HDMR terms are calculated once and then used adaptively by employing scaling and weighting factors which are determined through a straightforward methodology. The method also provides greater spectral fidelity than the traditional CS based methods as a result of the scaling factors. The proposed method has been tested on different MS images and compared with state-of-the-art pansharpening methods. The results are given both in terms of visual quality and numerical assessments.