Diurnal Ammonia Mapping based on Deep Learning from Geostationary Hyperspectral Infrared Sounder Observations

Atmospheric ammonia (NH₃) is a key air pollutant with high spatiotemporal variability, challenging the observation of its diurnal cycle. The Fengyun-4B Geostationary Interferometric Infrared Sounder (FY-4B/GIIRS) offers high-frequency measurements that capture this variability. We introduce a novel Multi-modal Fusion Transformer (MF-Transformer) to retrieve NH₃ total columns directly from hyperspectral radiances, meteorology, and ancillary data, circumventing costly radiative transfer simulations. Our retrievals are consistent with the IASI (Infrared Atmospheric Sounding Interferometer) NH₃ product (correlation coefficient, R=0.79) and Optimal Estimation (OE) retrievals (R=0.75), outperform benchmark machine learning models by ~20% in accuracy, and eliminate unphysical negative values. The method is orders of magnitude faster than OE approach, enabling global full-disk processing in tens of seconds. This advance allows the resolution of rapid NH₃ variations, demonstrating a transformative capability for operational monitoring.