A Lightweight Deep Neural Network for Sea Surface Wind Speed Retrievals From the FY-3D/MWRI

Sea surface wind speed (SSWS) is an important oceanic dynamical parameter, extensively utilized in numerical simulations of climate change and weather forecasts, as well as in storm intensity assessment. Microwave radiometers onboard sun-synchronous satellites can provide a large amount of SSWS data globally. Atmospheric attenuation caused by large raindrops and rainwater contamination can both lead to estimation errors, especially for the sensors without L-band and C-band channels such as the Microwave Radiation Imager (MWRI) sensor onboard the Fengyun-3D (FY3D) satellite. To investigate the potential of MWRI in SSWS detection under bad weather conditions, a lightweight deep neural network (LWDNN) is used based on the Global Change Observation Mission First-Water (GCOM-W1) Advanced Microwave Scanning Radiometer 2 (AMSR2) all-weather SSWS product in 2021. The SSWS product from AMSR2, soil moisture active passive (SMAP), buoys, and the ERA5 reanalysis data have been utilized to validate the LWDNN under all weather conditions. The overall root mean square error (RMSE) of MWRI SSWS is less than 2.0 m/s under all weather conditions and less than 1.5 m/s in the clear-sky region. In order to test the retrieval effectiveness of LWDNN in cyclone regions, the scenes of cyclones are collected. Results show that the RMSEs of the MWRI maximum wind speed (VMAX) product relative to the AMSR2 and SMAP data are 6.31 and 6.66 m/s, respectively, in the wind speed range of 20–70 m/s, and there are no systematic biases. The RMSEs of the MWRI SSWS relative to the Stepped-Frequency Microwave Radiometer (SFMR) is 5.21 m/s in the wind speed range of 6–56 m/s, and the SSWS of MWRI and SFMR is generally consistent.