How much hydrology is embedded in low-cost sensors? Machine-learning emulation of the mesoscale hydrological model from citizen soil moisture observations

Estimating groundwater recharge, actual evapotranspiration, and soil moisture is essential for agriculture, groundwater management, and drought prediction. Since these variables cannot be directly measured at large scales, they are typically derived using indirect approaches such as hydrological modeling. Models like the Mesoscale Hydrological Model (mHM) can simulate these variables but require specific inputs and substantial computational resources, while generally operating at regional rather than field scales. Using a dense citizen science network of low-cost soil moisture sensors across Flanders, we developed a LightGBM-based machine learning framework to emulate mHM-derived estimates of groundwater recharge, actual evapotranspiration, and soil saturation. The framework uses engineered features from point-scale observations to estimate regional-scale mHM outputs with sensor data alone. We analyzed predictor importance and spatial error patterns and three numerical experiments were conducted: (1) estimation of daily differences versus direct values, (2) assessment of the effects of increasing sensor density and spatial distribution, and (3) assessing model transferability using an independent field dataset, including sensor measurements and groundwater levels. The results indicated that LightGBM successfully emulated mHM outputs at most sites for all three variables, capturing seasonal patterns and differences between wet and dry years with low bias and error. Predicting daily differences improved model performance, except for recharge, and sensor density influenced bias under random resampling. Recharge predictions aligned with periods of rising groundwater levels, although winter signals were less clear due to sampling issues. Overall, machine learning models combined with sensor networks can approximate hydrological processes by capturing the timing, relative magnitude, and temporal patterns of hydrological responses. However, they do not replace physically based hydrological models and cannot provide direct hydrological measurements. Future work could optimize sensor density and distribution for scalable drought monitoring and water resource management.