Deep residual fully connected network for GNSS-R wind speed retrieval and its interpretation

Global navigation satellite system reflectometry (GNSS-R) has emerged as a new technique to provide L-band bistatic measurements for ocean wind speed retrieval, in which traditional geophysical model functions (GMFs) or shallow neural networks (NNs) are normally used. However, it is still challenging to identify and consider all relevant parameters in the GMF. Meanwhile, NN models face limitations due to the degradation problem, which restricts their depth and consequently their performance. Furthermore, the interpretation of NN models for GNSS-R wind retrieval is another issue. To this end, we propose a residual fully connected network (RFCN) fusing auxiliary information such as geometry, receiver gain, significant wave height, and current speed with track-wise corrected σ0. Referred to the European Centre for Medium-Range Weather Forecast (ECMWF) ERA5 wind product, the root mean square error (RMSE) and bias of RFCN winds are 1.031 m/s and -0.0003 m/s, respectively, with a 6% improvement in RMSE compared to debiased NOAA Cyclone Global Navigation Satellite System (CYGNSS) Version 1.2 (V1.2) wind speed retrieval. Moreover, in an intertropical convergence zone (ITCZ) area with large current speeds, the RMSE and bias are 1.006 m/s and -0.022 m/s: an improvement of 11.6% and 87.9% compared to debiased NOAA CYGNSS V1.2 winds. The bias ‘strips’ in these areas are nearly eliminated. Daily averaged error analyses also demonstrate that RFCN winds are more robust and consistent with ECMWF winds. For wind speeds larger than 20 m/s, referred to Soil Moisture Active Passive (SMAP) Level 3 final wind products, the RMSE and bias of fine-tuning RFCN (FT_RFCN) winds are reduced by 25.7% and 91.5% compared to NOAA winds. Finally, the RMSE and bias of retrievals in tropical cyclones, measured by Stepped Frequency Microwave Radiometer (SFMR) during 2021-2022, reveal an improvement of 3.5% and 21.2% compared to NOAA winds. Through SHapley Additive exPlanations (SHAP) models developed for RFCN and FT_RFCN, the contribution of each feature is quantitatively evaluated, while providing insights into their interactions within the ‘black-box’ NN models with clear physical meanings.