A Deep Learning Approach to Estimate CERES Broadband Flux at 1km Resolution with MODIS Radiances

Direct observations of outgoing radiance by satellite broad-band bolometers including NASA Earth Observing System (EOS) Clouds and Earth's Energy Budget (CERES) instruments are critical for understanding the terrestrial energy budget and for validating climate models. Unfortunately, engineering constraints typically restrict the instruments to coarse spatial resolutions, which limits their ability to characterize the influence of kilometer-scale features like cities, mountain ridges, and smaller clouds on outgoing flux. This project utilizes a novel machine learning architecture to estimate the outgoing shortwave and longwave fluxes on a 1km grid with narrow-band Moderate Resolution Imaging Spectroradiometer (MODIS) data, and validates them against the simultaneous point spread function weighted ~16x24km CERES observations over a wide variety of scenes collected globally throughout EOS' 22 year data record. The network is incentivized to maintain closure with the coarse flux observations while learning their nonlinear statistical relationship to surface properties resolved by 11 MODIS L1b radiance bands in the visible, near-infrared, and infrared ranges, and their viewing geometry. Results offer reasonable flux estimates over many different scenes, and can serve as rich image enhancements which emphasize land surface emissivity, separate snow from clouds, and resist sun glint artifacts.