Methodological Trends and Challenges in Deep Learning based Super-Resolution for Sentinel-5P Data

Super-resolution (SR) is the reconstruction of a higher resolution (HR) image from one or more low resolution (LR) images. In remote sensing, SR is particularly useful because it lets us enhance spatial detail beyond what is provided by satellite sensors originally. Satellite-based air quality monitoring plays a crucial role in evaluating and managing human-induced emissions. Sentinel-5P has provides data related to atmospheric pollutant measurements with a spatial resolution of 3.5x5.5km². It is one of the best available spatial resolution however it is limited in detecting fine-scale sources of NO_x emissions, particularly in densely populated urban regions and maritime corridors. This study highlights the relatively underexplored class of super-resolution frameworks that employ deep learning techniques to enhance the spatial resolution of Sentinel-5P radiance data. The deep learning based method developed specifically for enhancing the spatial resolution of Sentinel-5P radiance data are outperforming in super-resolution of Sentinel-5P NO2 data. The state of the art approaches integrated a physical degradation model based on the point spread function (PSF) using an anisotropic Gaussian kernel and a modified lightweight U-net to reconstruct high resolution outputs. With this setting, the models were able to achieve the best performance according to the evaluation metrices. Such a deep learning super-resolution techniques offer an advantage for further detailed analysis of Sentinel-5P data by enhancing its spatial resolution. The effectiveness of the super-resolution depends heavily on accurately modeling the sensor-specific degradation process and it needs fine-tuning for robutness. Deep neural networks requires substantial computational resources for training and inference, which limits their deployment in real-time or resource constrained environments. Although the model accounts for sensor degradation, it still faces challenges when dealing with unforeseen real-world artifacts such as atmospheric interference, measurement noise, and other distortions not captured by the model. A significant limitation found was lack of higher resolution benchmark in the current state of research in this field. Large scale super-resolved dataset would be useful for local analysis such emissions from ships. These findings highlight the need for broader participation from the research community to validate, extend, and independently assess the proposed methods. Future experiments will include comparisons against advance GANs and other transformer-based models, and cross-validation with CAMS reanalysis data and ground-based stations.