An AI Derived Forward Radiative Transfer Emulator for Global All-sky Conditions

With the development of high spatial, spectral, and temporal resolution sensors to meet performance requirements in numerical weather prediction (NWP) systems, demand for efficiency of the Radiative Transfer Models (RTMs) continues to increase in NWP. A fully connected “deep” neural network algorithm with the Community RTM (CRTM) as the reference (FCDN_CRTM) has been developed to explore the efficiency, accuracy and stability of reproducing the Advanced Radiative Transfer Modeling System (ATMS) brightness temperatures (BTs) in 22 microwave bands under global ocean clear-sky conditions. However, the data assimilation step of NWP systems requires the RTM to operate under all-sky conditions over both ocean and land surfaces. Therefore, in this study, we introduced cloud parameters into the FCDN_CRTM and extend the model training and evaluation to include land surfaces. Emissivity spectra were generated randomly rather than using the CRTM surface type-based default emissivity and were input into FCDN_CRTM. This was done to meet the NOAA Microwave Integrated Retrieval System (MiRS) requirement, where emissivity is a directly-retrieved parameter and is updated in each retrieval iteration so that emissivity can vary spatially and temporally. Similar to earlier research, the cloud parameters were extracted from the European Centre for Medium-Range Weather Forecasts (ECMWF) with 0.25ᵒ horizontal resolution and 91 vertical layers. The three cloud types - water cloud, rain cloud and graupel - were defined and their profiles were added as new features to retrain the FCDN_CRTM. However, when using the well-trained FCDN_CRTM in global all-sky conditions to test the ECMWF data-simulated CRTM BTs with the CRTM default emissivity, it was found that a significant cold bias persisted in the ocean for all surface window bands. After the issue was fixed, the mean biases of the FCDN_CRTM prediction minus CRTM simulation are within 0.02K and the standard deviation (SD) are less than 0.3 K for all 22 bands, indicating that prediction accuracy for all-sky global is also close to that for clear-sky ocean cases, especially for temperature sounding bands (Channels 5-15). Furthermore, the FCDN_CRTM Jacobians in term of surface parameters and atmosphere profiles under all sky conditions were also analyzed by comparing them with CRTM Jacobian for data assimilation or inversion algorithm applications.