资源环境科技发展态势分析平台

Inversions of the Earth Observation Satellite (EOS) Advanced Microwave Scanning Radiometer (AMSR-E) brightness temperatures (T-bs) to derive the land surface temperature (T-s) are presented based on building a global transfer function by neural networks trained with AMSR-E T-bs and retrieved microwave T-s*. The only required inputs are the T-bs and monthly climatological emissivities, minimizing the dependence on ancillary data. The inversions are accompanied by a coarse estimation of retrieval uncertainty, an estimate of the quality of the retrieval, and a series of flags to signal difficult inversion situations. For approximate to 75% of the land surface the root-mean-square difference (RMSD) between the training target T-s* and the neural network retrieved T-s is below 2.8 K. The RMSD when comparing with the Moderate Resolution Imaging Spectroradiometer (MODIS) clear-sky T-s is below 3.9 K for the same conditions. Over 10 ground stations, AMSR-E and MODIS T-s were compared with the in situ data. Overall, MODIS agrees better with the station T-s than AMSR-E (all-station mean RMSD of 2.4 K for MODIS and 4.0 for AMSR-E), but AMSR-E provides a larger number of T-s estimates by being able to measure under cloudy conditions, with an approximated ratio of 3 to 1 over the analyzed stations. At many stations the RMSD of the AMSR-E clear and cloudy sky are comparable, highlighting the ability of the microwave inversions to provide T-s under most atmospheric conditions. Closest agreement with the in situ T-s happens for stations with dense vegetation, where AMSR-E emissivity is less varying.