Nonlinear Soil Moisture Retrieval from Sentinel-1 SAR Using Ensemble Machine Learning

Accurate soil moisture estimation plays a vital role in applications such as agricultural management, hydrological modeling, and climate analysis. Synthetic Aperture Radar (SAR) observations provide a reliable source of information for soil moisture retrieval due to their high spatial resolution and capability to operate independently of weather and day conditions. This study investigates the performance of two gradient boosting based machine learning algorithms, Extreme Gradient Boosting (XGBoost) and Categorical Boosting (CatBoost), for soil moisture prediction using SAR-derived and auxiliary environmental features. Sentinel-1 SAR imagery is employed as the primary data source, while reference soil moisture measurements are obtained from the NASA Soil Moisture Active Passive (SMAP) mission, together with auxiliary variables derived using Google Earth Engine. Feature engineering and regularization strategies are applied to enhance model robustness and reduce overfitting. Model performance is evaluated using the Pearson correlation coefficient (r), Root Mean Square Error (RMSE), and coefficient of determination (R²). The results indicate that both models achieve reliable predictive accuracy; however, XGBoost exhibits stronger fitting capability during training, whereas CatBoost demonstrates improved generalization performance, reduced bias, and enhanced robustness on unseen data, highlighting its suitability for operational SAR-based soil moisture estimation.