Deep learning-based multiple-image super-resolution for Sentinel-2 data

Enhance the capacities of super-resolution reconstruction

The general goal of the DeepSent project was to enhance the capacities of super-resolution reconstruction applied to multispectral Sentinel-2 images, especially if multiple images of the same region, captured at a different time, are available. The existing networks applied to super-resolved the Sentinel-2 images in a band-wise manner (each band will be treated independently), and then attempted to exploit the correlation among the multiple bands. Such an approach enhanced so that the developed network presented with a multispectral image at the input to generate a high-resolution output. The output encompasses a panchromatic image, as well as an RGB image, and a multispectral image of higher resolution than the one presented at the input. An important goal of the DeepSent project was to develop techniques addressed at raising the quality of the data used for training deep models. This was achieved by selecting the degradation technique (or a set of techniques) employed to generate the low-resolution images that can be coupled with the original images (treated as a high-resolution reference).