A Physically Based Neural Network for Fast Infrared Atmospheric Transmittance Simulation

The atmospheric radiation transfer model (RTM) is the foundation and core of the physical retrieval of atmospheric and surface parameters in remote sensing as well as the assimilation of satellite observation data. Infrared RTM is widely used in the retrieval of Earth’s surface temperature, cloud detection, and water vapor remote sensing. This letter is committed to exploring the fast and accurate simulation of atmospheric radiation transfer over the clear-sky ocean using deep learning algorithms, with the key issue being the fast calculation of atmospheric transmittance. We have constructed a neural infrared transmittance model (NITM) for atmospheric radiation transfer simulation from thermal infrared channels and applied it to the visible infrared imaging radiometer suite (VIIRS) M15 and M16 channels. In addition, to improve the model’s performance, we conducted a sensitivity analysis on the inputs and selected predictors with high sensitivity to transmittance. The comparison results with the line-by-line radiative transfer model (LBLRTM) indicate that the algorithm achieves fast and precise simulation of atmospheric radiation transfer, with the simulated brightness temperature (BT) accuracy better than 0.1 K.