Sea Surface Wind Speed Retrieval from GNSS-R Using Adaptive Interval Partitioning and Multi-Model Ensemble Approach

Sea surface wind speed is a crucial parameter for studying climate change and ocean dynamics. Accurate, real-time measurements are essential for meteorological and oceanographic observations. Global Navigation Satellite System Reflectometry (GNSS-R) is a key technology for sea surface wind speed retrieval. Existing wind speed retrieval models employ two primary approaches: unified modeling across the entire wind speed range and independent modeling for partitioned wind speed intervals. The former cannot effectively address physical property variations across wind speed ranges. The latter, while mitigating this issue, relies on empirical thresholds for interval partitioning that ignore actual data distribution and struggles to assign new samples to appropriate intervals during prediction. To address these limitations, this study employs the Gradient-Boosted Adaptive Multi-Objective Simulated Annealing (GAMSA) algorithm to construct a multi-objective optimization function and perform data-driven wind speed interval partitioning. Specialized XGBoost sub-models are then constructed for each partitioned interval, and their predictions are integrated through a stacking ensemble learning architecture. The experiments utilize a Cyclone Global Navigation Satellite System (CYGNSS) and ERA5 reanalysis data. The experimental results show that the proposed method reduces the root mean square error (RMSE) from 1.77 m/s to 1.43 m/s and increases the coefficient of determination (R2) from 0.6293 to 0.7770 compared with a global XGBoost model. It also exhibits enhanced accuracy under high wind speeds (>16 m/s) and, when independently validated with buoy data, achieves an RMSE of 1.52 m/s and R2 of 0.79. The proposed method improves retrieval accuracy across both overall and individual wind speed intervals, avoids the sample isolation problem inherent in traditional empirical partitioning methods, and resolves the issue of assigning new samples to appropriate sub-models during application.