A proposal for full-scale processing and assessment of MS pansharpening

In this study, we propose to perform pansharpening fusion using multispectral (MS) data that have been harmonized with the encompassing panchromatic (PAN) image and to measure the quality using the harmonized data as a reference for consistency. Harmonization is a fast and robust procedure that removes mismatches, mainly aliasing artifacts, and shifts towards PAN. The results, presented on GeoEye-1 and WorldView-2 data with 10 pansharpening methods covering all possible approaches of the last twenty years, including neural ones, highlight: 1) that on harmonized data all methods work reasonably well in visual terms, while without harmonization, methods based on multiresolution analysis (MRA) produce less sharp results than those of component substitution (CS); 2) that the full-scale quality indexes, which penalized MRA methods with respect to CS in the absence of harmonization, are comparable and match visual quality. In this way, the fullscale tests are in accordance with the degraded-scale ones, where the mismatches between the datasets are reduced by the spatial downsampling. Unfortunately, degraded scale tests are no longer acceptable, since neural methods exist that can be trained on ground truth (GT), which is the original undegraded image, and is not available in practical cases. The intrinsic inconsistency of full-scale simulations had been blamed on the inadequacy of quality indexes. Instead, it depends on the mismatches of the data, which can be corrected. We believe that our study will open new horizons in the development of better and better pansharpening methods.