What question did this study set out to answer?

The study aims to create a high-resolution surface soil moisture dataset using machine learning methods for better drought and flood forecasting.

April 17, 2026Advances in Atmospheric Sciences0 citationsOpen Access

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

YDYifan Dong HYHuiling Yuan

Key Points

The study aims to create a high-resolution surface soil moisture dataset using machine learning methods for better drought and flood forecasting.
Developed a machine learning framework for soil moisture data fusion.
Trained four machine learning models: Random Forest, XGBoost, LightGBM, and CatBoost.
Used Recursive Feature Elimination (RFE) for feature selection and Optuna for hyperparameter tuning.
Analyzed model interpretability using SHAP.
The fusion model improved soil moisture estimation with lower root-mean-square error compared to CLDAS.
RFE reduced features by 57% while maintaining predictive accuracy.
SHAP identified high-accuracy inputs as the most significant predictors, followed by terrain, soil, and meteorological features.

Abstract

Abstract High-resolution soil moisture (SM) data are critical for drought monitoring and flood forecasting. This paper describes the establishment of an interpretable machine learning (ML)-based framework for SM data fusion and generates a daily-scale, 1-km resolution, surface SM (0–10 cm) dataset over China (2000–2025). Four state-of-the-art ML models—Random Forest, XGBoost, LightGBM, and CatBoost—were trained based on in situ SM data from 2371 automatic observation stations across China. Model performance was optimized via Recursive Feature Elimination (RFE) and automated hyperparameter tuning using Optuna, while SHapley Additive exPlanations (SHAP) provided mechanistic interpretability of the ML models. The key findings of this study are as follows: (1) The fusion model primarily enhances SM estimation, exhibiting lower root-mean-square error than CLDAS (China Meteorological Administration Land Data Assimilation System) SM, despite marginally weaker daily temporal correlation; (2) RFE eliminated 57% of features while preserving predictive accuracy; (3) SHAP analysis revealed high-accuracy SM inputs as the most influential predictors, followed by static (terrain and soil properties) and meteorological variables. The SM fusion method developed in this study is transferable to multi-source satellite SM fusion and downscaling. The dataset is publicly available at https://doi.org/10.11888/Terre.tpdc.302923 .

AIに質問

Bookmark

View Full Paper