| description abstract | AbstractWe have investigated whether differences across Clouds and the Earth?s Radiant Energy System (CERES) top-of-atmosphere (TOA) clear-sky angular distribution models, estimated separately over regional (1° ? 1° longitude?latitude) and temporal (monthly) bins above land, can be explained by geophysical parameters from Max Planck Institute Aerosol Climatology, version 1 (MAC-v1), ECMWF twentieth-century reanalysis (ERA-20C), and a MODIS bidirectional reflectance distribution function (BRDF)/albedo/nadir BRDF-adjusted reflectance (NBAR) Climate Modeling Grid (CMG) gap-filled products (MCD43GF) climatology. Our research aimed to dissolve binning and to isolate inherent properties or indicators of such properties, which govern the TOA radiance-to-flux conversion in the absence of clouds. We collocated over seven million clear-sky footprints from CERES Single Scanner Footprint (SSF), edition 4, data with above geophysical auxiliary data. Looking at data per surface type and per scattering direction?as perceived by the broadband radiometer (BBR) on board Earth Clouds, Aerosol and Radiation Explorer (EarthCARE)?we identified optimal subsets of geophysical parameters using two different methods: random forest regression followed by a permutation test and multiple linear regression combined with the genetic algorithm. Using optimal subsets, we then trained artificial neural networks (ANNs). Flux error standard deviations on unseen test data were on average 2.7?4.0 W m?2, well below the 10 W m?2 flux accuracy threshold defined for the mission, with the exception of footprints containing fresh snow. Dynamic surface types (i.e., fresh snow and sea ice) required simpler ANN input sets to guarantee mission-worthy flux estimates, especially over footprints consisting of several surface types. | |
