Top-of-Atmosphere Radiative Flux Retrievals from CERES Measurement Using Artificial Neural Networks

The Clouds and the Earth's Radiant Energy System (CERES) on Terra satellite provides accurate top-of-atmosphere (TOA) radiative flux values by applying Anisotropic Models (AM) to its broadband radiance measurement. For scene types defined by coincident imager-based cloud retrievals these models are developed empirically as functions of the surface and atmosphere condition over CERES field-of-view. However, the CERES/Terra data has 5.6% of the CERES footprints with imager information missing or not sufficient for a complex scene identification. To avoid any systematic bias in the data it is very important to provide TOA fluxes as precisely as it is possible for such footprints. In this study, we demonstrate an application of a partially connected feed-forward error-backpropagation Artificial Neural Network (ANN) simulation to the TOA radiative flux retrievals from CERES measurements in the absence of imager information. All-sky ANN-based anisotropic models are developed for ten different surface types. To maximize the ANN performance we used a partially connected first hidden neuron layer, and low noise ANN training sets. We demonstrate performance of the ANN-based AMs in comparison with the original empirical models. For all scene types, the ANN-derived mean TOA fluxes deviate from original CERES/Terra TOA fluxes by less than 0.3 W/m2 for shortwave and 0.5 W/m2 for longwave channels. Instantaneous ANN-derived TOA fluxes are consistent in average within 32.3 W/m2 for shortwave, 9.5 W/m2 and 7.1 W/m2 longwave day- and night-time channels, respectively. Applying similar ANN technique we attempt to retrieve cloud properties over field-of-view using CERES measurement alone.