Efficient Hyperspectral Super-Resolution of Sentinel-5P Data via Dynamic Multidirectional Cascade Fine-Tuning

Sentinel-5P is a valuable resource for academics and policymakers. The ability of the satellite’s equipment to span the electromagnetic spectrum from ultraviolet (UV) to short-wave infrared (SWIR) frequencies is vital in determining the distribution of important gaseous pollutants on a global scale, a significant turning point for air quality monitoring. In technical terms, Sentinel-5P provides an excellent balance between spatial and spectral resolutions; however, physical limitations keep hindering the quality of its products. S5Net is the first deep-learning-based (DL-based) approach designed to super-resolve Sentinel-5P radiance images. Despite its simplicity, this neural network has showed excellent performance when applied to monochromatic images, particularly when compared to more complex deep neural networks. Yet, this groundbreaking study has a significant limitation: the computational inefficiency of the fine-tuning employed, which must be adequately extended to numerous channels. We hence propose a novel dynamic multidirectional cascade fine-tuning procedure, whose routine is fully governed by the correlation between consecutive spectral channels. Our study is accordingly successful in striking a remarkable balance between spectral coherence and spatial resolution improvement, as well as substantially optimizing computing efficiency. The code is available at
https://github.com/alcarbone/S5P_SISR_Toolbox
.