Downscaling of soil moisture in the Wujiang River Basin based on integrated decision tree-based machine learning

Soil moisture (SM) is a crucial variable for agriculture, drought monitoring and hydrological modelling. While microwave remote sensing provides large-scale SM data, its coarse spatial resolution limits fine-scale applications. This study compares three decision tree-based ensemble machine learning algorithms: random forest (RF), light gradient boosting machine (LightGBM) and extreme gradient boosting (XGBoost), to downscale the 25-km resolution European Space Agency Climate Change Initiative Soil Moisture (ESA CCI SM) to a 1-km resolution in the Wujiang River Basin, China. Models were evaluated using randomly and temporally stratified test sets and validated against in situ SM observations. The results revealed that the R2 values of the three SM prediction models ranged from 0.6694 to 0.7163, with RMSE 0.0165–0.0178 m3/m3 and Bias 0–0.0001 m3/m3, with LightGBM performing relatively better. Evaluation using independent year-stratified test sets revealed seasonal and spatial differences among the three models. Before and after downscaling, SM's spatial distribution remained largely consistent and responded well to precipitation. The RF model achieved the highest accuracy (RMSE = 0.0189 m3/m3, Bias = 0.003 m3/m3) and the strongest correlation with the site SM observations (R = 0.6714). The downscaled SM aligns better with the measured data and the proposed downscaling framework supports hydrological, agricultural and environmental applications.