High-resolution soil moisture retrieval in heterogeneous terrain: A comparison of ensemble learning approaches using landsat 8/9

The study area is located in Tianjun County, eastern Qinghai Province, China, near Qinghai Lake on the northwestern Qinghai-Tibet Plateau, covering approximately 40 × 36 km. Using more than 20 soil moisture–related indices derived from Landsat data together with elevation information, this study develops a 30 m resolution soil moisture retrieval framework. Four ensemble learning models, Random Forest (RF), Extremely Randomized Trees (ERT), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost), were trained and evaluated using QLB-NET in situ soil moisture observations to assess their spatial and temporal predictive performance. All models showed relatively stable performance, with CatBoost and XGBoost significantly outperforming ERT and RF. CatBoost achieved the best test performance, with average values of R² = 0.79, RMSE = 0.060 m³ /m³ , Bias = –0.0009, and MAE = 0.0457 across multiple runs. The resulting high-resolution soil moisture maps reveal pronounced spatial heterogeneity and clear month-to-month variability. SHAP analysis indicates that elevation is the most influential predictor of soil moisture, and soil moisture variance increases markedly with terrain complexity at a spatial scale of 3 km. These results highlight the value of integrating high-resolution remote sensing data with ensemble learning methods for soil moisture estimation in complex mountainous regions. • Framework developed for 30 m soil moisture using Landsat 8/9 and elevation. • CatBoost showed superior accuracy and stability among four ensemble models. • SHAP analysis identified elevation, SWCI, and TVSDI as key SM predictors. • Terrain complexity significantly enhances SM heterogeneity at the 3 km scale. • High-resolution maps reveal fine-scale spatial and seasonal SM variability.