A Study on the Fusion of Multiple Linear Statistical Regression and Deep Learning Models for Sea Surface Wind Speed Retrieval in Typhoon Regions Based on FY-3D_MWRI Data

The Level 1 brightness temperature data from the FY-3D (Fengyun-3D) Microwave Radiation Imager (MWRI) provide an important data source for retrieving high wind speeds over ocean surfaces in typhoon-affected regions. This study proposes a wind speed inversion method, MLF-GA, which fuses a GRU-ATTENTION deep learning model with results from multiple linear statistical regression., This method achieves fine-scale inversion of sea surface wind speeds in typhoon regions by integrating 145-dimensional features, including multi-channel brightness temperatures from FY-3D MWRI, wind speed analyses from the National Centers for Environmental Prediction (NCEP), and environmental parameters from the European Centre for Medium-Range Weather Forecasts (ECMWF). The model incorporates a label-aware attention mechanism to dynamically weigh the features of high wind speed samples, significantly enhancing its ability to predict extreme wind speeds. A large number of all-weather high wind speed samples from 2020 to 2023 are added, and quality control of the experimental results is conducted through statistical data-driven methods. The study further demonstrates the effectiveness of using a polarization difference threshold (37V-37H > 5 K) in brightness temperatures for noise suppression and reveals the significant impact of cloud and precipitation interference on inversion accuracy. The experiments select typhoons “SAOLA”, “YAGI” and “BEBINCA” from the Eastern Hemisphere, and “BERYL” and “IDALIA” from the Western Hemisphere as test set. Results show that in high wind speed zones (>20 m/s), the model achieves root mean square errors (RMSE) of 3.08 m/s, 2.81 m/s, and 2.10 m/s in the Eastern Hemisphere, and 2.46 m/s and 1.88 m/s in the Western Hemisphere, representing an approximate 30% reduction in error compared to traditional multiple linear regression models. Validation using observational data from NDBC buoys shows that the RMSE between the model-retrieved wind speeds and actual measurements ranges from 2.01 to 2.47 m/s, outperforming the NCEP reanalysis data (RMSE of 2.07-2.31 m/s). This method offers a high-precision solution for the dynamic monitoring of high wind speeds in typhoon-affected areas and holds significant potential for applications in meteorological forecasting and disaster prevention and mitigation.