Validation and ensemble-based layer-wise correction of soil moisture observations from automatic stations

Soil moisture is one of the fundamental variables in land–atmosphere interactions, hydrological processes and vegetation dynamics. Accurate soil moisture information is of great significance for environmental and climate studies. In recent years, automatic monitoring stations have been increasingly deployed owing to their advantages of high-frequency and standardized measurements. However, their measurements suffer from non-negligible biases, limiting the reliability of automatic soil moisture datasets. To address these challenges, this study conducts a comprehensive validation and machine-learning-based correction of three automatic soil moisture systems: the automatic soil moisture station, the CRS-2000C regional soil moisture measurement system, and the soil temperature–moisture monitoring system at the Xilinhot National Climate Observatory, China. Using manual measurements as reference, the observations from the three automatic stations were validated. All the automatic soil moisture measurements revealed substantial biases, especially in deeper soil layers. To reduce the biases, an ensemble correction framework that employed generalized additive model to integrate Cubist, Random Forest, XGBoost, and CatBoost models was developed for layer-wise soil moisture correction. Five-fold cross-validation was applied to evaluate correction performance. After correction, the accuracy of all stations improved significantly, with R 2 increasing by 0.075–0.289, mean absolute error decreasing by 0.012–0.057 m 3 /m 3 , and root mean square error decreasing by 0.013–0.075 m 3 /m 3 , with particularly pronounced improvements in deeper layers. This study highlights the necessity of correcting automatic soil moisture observations and provides an effective framework for the correction.