Pansharpening of multispectral images based on nonlocal parameter optimization

High-quality pansharpened multispectral (MS) images are rarely obtained from fast, efficient, and robust algorithms. In most cases, effective pansharpening methods have huge computational complexity, as in the case of variational methods, or algorithms based on sparse representations. Moreover, injection models are often application dependent, not sufficiently general to be applied to different scenarios, and the resulting algorithm implementations cannot process large-size images. The proposed pansharpening method is accurate and fast and can be successfully applied to huge images. It also solves the problem of context-adaptive schemes that tune the spatial injection parameters on local statistics: Instabilities and blocky artifacts can be generated by pansharpening methods whose parameters are computed on local windows. The proposed method is an extension of the classical component-substitution algorithms: An optimal detail image (in the mmse sense) extracted from the panchromatic band is calculated for each MS band by evaluating band-dependent generalized intensities. It overcomes window-based local estimation of parameters by applying a nonlocal parameter optimization through K-means clustering. Very high quality scores, both at degraded and full scale, and excellent visual quality of the fused images demonstrate the validity of the method.