Key points are not available for this paper at this time.
Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) enables high temporal resolution soil moisture (SM) monitoring, but it faces challenges from complex signal decoupling and nonlinear relationships. This study develops a machine-learning retrieval framework that integrates multi-feature inputs to address these issues. The framework uses the Discrete Wavelet Transform (DWT) to adaptively decompose signal-to-noise ratio (SNR) trends and applies Total Least Squares (TLS) to ensure robust feature extraction. By incorporating interference observables from multiple GPS satellites, we implement a Bayesian-Optimized Random Forest (BO-RF) model. This model captures the intricate mapping between the observables and the dynamic patterns of SM. Validation with data from the Plate Boundary Observatory (PBO) network shows remarkable performance: at stations P037, P041, and P043, the model achieves R of 0.937, 0.959, and 0.890, respectively, with corresponding RMSE of 0.033, 0.021, and 0.044 cm³·cm–3. When compared with the RF, BO-XGBoost, and BO-CNN, the BO-RF model reduces errors by 5.8% to 43.2%. Furthermore, the integration of SHAP (SHapley Additive exPlanations) analysis boosts the framework’s interpretability and uncertainty quantification, creating a robust and scalable paradigm for GNSS-IR SM estimation.
He et al. (Thu,) studied this question.