SSAT: Sensor-Satellite Auto-Correlation Transformer for Enhanced Aerosol Optical Depth Prediction

Satellite-derived aerosol optical depth (AOD) observations are highly valuable to describe the horizontal distribution of aerosols, but are hampered by spatial data gaps and limited temporal coverage (typically once a day). The synergism of these measurements with chemistry-transport models (CTM) may be used to overcome these limitations. Although physically constrained methods (e.g. data assimilation) are a common practice for addressing this synergism, these methods may be difficult to implement and are highly computationally demanding. On the other hand, statistical or learning-based techniques can still offer flexible and effective solutions. In this work, we present SSAT, a transformer-based approach that fuses satellite and model outputs to refine AOD predictions without modifying the underlying CTM itself. Our design leverages the auto-correlation mechanism from Autoformer and introduces GridMSE, a specialized loss function aimed at improving spatial coherence and handling imbalanced data. Extensive experiments show that SSAT has higher accuracy and better visual correspondence compared to existing tree-based and deep learning baselines, particularly in underrepresented aerosol load regimes. While it does not replace physically rigorous data assimilation, SSAT can serve as a complementary tool for quickly refining AOD maps, ultimately offering a more complete basis for monitoring aerosol distributions and informing environmental analyses.