A soil moisture estimation model based on Stacking of decision trees and its evaluation over China

Study region This study focuses on mainland China, spanning arid, semi-arid, semi-humid, and humid climatic zones. Study focus A Stacking ensemble model integrating CLDAS land data assimilation product, AMSR2 microwave soil moisture, GPM precipitation, and static environmental variables is developed to estimate pentad-scale relative soil moisture (RSM) at a 0.1° (~10 km) resolution. The model uses Cubist, Random Forest, and Bagging as base learners. The contributions of dynamic and static features are quantified via SHAP analysis. New hydrological insights for the region The estimations reveal pronounced spatial variations and systematic bias patterns across China's climatic gradients. In semi-arid regions, the model systematically underestimates moisture (BIAS = −3.35%), whereas humid regions exhibit overestimation (+1.81%). The semi-humid transitional zone shows negligible bias (−0.31%) but the highest random error (RMSE = 18.62%), driven by localized spatial heterogeneity. Furthermore, the spatial distribution of moisture is closely associated with static environmental features, particularly topography and soil texture, through their interactions. Under severe drought (RSM < 30%), the model substantially overestimates soil moisture (BIAS = +17.48%), posing a critical limitation for operational drought warning. These findings provide quantitative insights into the varying dominant controls on soil moisture across diverse climatic gradients.