A Cross-Scale Loss for CNN-Based Pansharpening

To cope with the lack of input-output training samples, deep learning (DL) methods for pansharpening usually resort to Wald's protocol or other similar downscaling processes. By doing so, the scaled versions of the multispectral (MS) and panchromatic (PAN) components serve as input while the original MS plays as output during the training phase. As a side effect, the informational gap between reduced and full scales causes a mismatch between the training and test phases. In fact, DL methods typically provide a pretty good performance at reduced scale, with a good margin over traditional solutions that tends to vanish in the full-resolution framework. In this work, we propose a training framework that involves both the reduced and the full scale versions of the multiresolution image samples. This is achieved thanks to a suitably defined loss which comprises costs for both scales. Our numerical and visual experimental results confirm that the proposed approach provides an improved performance in the full-resolution case.