Satellite Super-Resolution: on the Role of Multi-Date and Band-Shift

Super-resolution for satellite imagery has gained significant attention due to its potential to improve analytical products without requiring high-resolution imagery. In this work, we aim to study the multi-and single-image super-resolution problem, focusing on the interplay between multi-date information and band-shift characteristics of the imagery. For this study we propose a simulated dataset based on Landsat-8 imagery, aimed at replicating the band-shift properties of Sentinel-2 or their absence, while contemplating 20 multi-date frames for each scene. Our experiments show that the multi-date information can yield gains in both scenarios over single-date super-resolution. Furthermore, we also highlight the diminishing impact of adding frames in both scenarios and the importance of auxiliary frame selection strategies, where temporally closer frames tend to yield better reconstruction results. These findings provide valuable insights into optimizing super-resolution techniques for remote sensing, with implications for high-revisit multi-spectral satellites.