An improved GNSS-R ocean scatterometry recalibration approach using gradient boosting decision tree

Spaceborne Global Navigation Satellite System Reflectometry (GNSS-R) has developed rapidly for ocean scatterometry over the past three decades. However, residual signal calibration errors, arising from inaccuracies in some key parameters and the complex space environment, degrade its performance. In this work, we propose a one-step signal recalibration framework based on Gradient Boosting Decision Tree (GBDT) in a scaling correction manner guided by qualitative analyses. Results show good alignment between the learned and theoretical correction factors, wherein the bin ratio (BR) exhibits the dominant contribution. The recalibrated σ 0 exhibits lower sensitivity to calibration-related input features than the Cyclone Global Navigation Satellite System (CYGNSS) Version 2.1 (V2.1) σ 0 , and for several parameters, also lower sensitivity than the Version 3.2 (V3.2) σ 0 produced using traditional step-wise corrections during the period from January to June 2019. The maximum deviation of σ 0 across different wind speeds is also substantially reduced, leading to root mean square error (RMSE) of 1.52 m/s for ocean wind speed retrievals over 0–27 m/s, an improvement of 17.4% and 16.0% compared to using the V2.1 and V3.2 σ 0 products, respectively. Validation against buoy wind speed data shows an RMSE of 1.29 m/s and a bias of 0.020 m/s, representing reductions of 17.3% and 87.1%, respectively, relative to retrievals using the V3.2 σ 0 product. Feature importance analysis using SHapley Additive exPlanations (SHAP) confirms consistency with theoretical expectations, especially regarding coupling effects between input features. The recalibrated σ 0 exhibits significantly reduced dependence on BR, GNSS-R satellites antennas, and Global Positioning System (GPS) transmitters, simultaneously. Furthermore, the spatial distribution of correction factors is strongly correlated with the noise floor, indicating that the proposed GBDT model identifies high-noise cases through BR and improves the robustness of σ 0 in these high-noise regions. As an alternative to step-wise empirical correction methods that could introduce artificial biases, the proposed recalibration method provides a new perspective and potential pathway for improving GNSS-R scatterometry.