Superresolution for multispectral satellite images

Optical remote sensing satellites, like ESA's Sentinel-2, measure reflected energy in a series of spectral band from the visible to infrared part of the electromagnetic spectrum. Sentinel-2 MSI has 12 spectral channels, and spatial resolution ranging from 10 to 60. A high number of Sentinel-2 images are available.
Other satelllite sensors, like Spot, have both several multispectral bands of medium resolution and one panchromatic band of higher resolution. The goal of this thesis will be to generate superresolution multispectral images.

The  deep learning architecture we will use is diffusion models. We will combine supervised and unsupervised learning. A possibility will be to train supevised on a limited set of images with both high and medium resolution available, and unsupervised on Sentinel-2 MSI images.