China’s 1 km Daily Surface Soil Moisture Fusion Dataset (2000–2025) Based on Explainable Machine Learning

It employs the CatBoost ensemble learning model, utilizing in-situ measurements from over 2,000 automatic soil moisture stations across China as training labels (ground truth). A diverse array of input data serves as predictors, including reanalysis products (ERA5, ERA5-Land, GLDAS-Noah, GLDAS-CLSM), meteorological variables, topographic data, and soil properties.

The fusion framework emphasizes efficiency, automation, and explainability: During model training, an XGBoost-SHAP integrated Recursive Feature Elimination (XGB-SHAP-RFE) method is applied for feature selection, and the Optuna framework is leveraged for hyperparameter optimization, enhancing predictive accuracy. Additionally, SHAP (SHapley Additive exPlanations) methods provide interpretable, visual explanations of the model outputs.