Pansharpening using data-centric optimization approach

Earth's observation satellites provide simultaneously both multispectral (XS) and panchromatic (pan) images but XS image has a lower spatial resolution when compared to pan image. Pansharpening is a pixel-level fusion technique resulting in a high-resolution multispectral image in terms of both spatial and spectral resolution. The problem lies in maintaining the spectral characteristics of each channel of the XS image when pan image is used to estimate the high spatial XS image. Many techniques have been proposed to address the problem. A popular method involves a sensor-based approach where correlation among the XS channels and correlation between the pan and spectral channels are incorporated. In this paper, we take a wholesome approach based on the reflectance data irrespective of the sensor physics. A linear regression model is formulated between the XS channel and the panchromatic data. We formulate an optimization problem in terms of Lagrange multiplier to maximise the spectral consistency of the fused data with respect to the original XS data, and to minimise the error in variance between the reference data and the computed data. We validate and compare our method with IHS and Brovey methods based on evaluation metrics such as Chi-square test and the R-2 test. The implementation is done and presented using IKONOS satellite data.