Machine Learning Assisted Calibration of pyWBM Using In-Situ, Satellite, and Reanalysis Soil Moisture Data for High Resolution Soil Moisture Ensemble Projections

Reliable soil moisture estimation is challenged by sparse in-situ networks, inconsistencies across satellite products, and structural limitations in simplified land-surface models. This study develops a machine learning assisted calibration framework for pyWBM, a Python implementation of the University of New Hampshire Water Balance Model, to generate improved historical reconstructions and ensemble projections of root-zone soil moisture for counties across Illinois. We integrate in-situ observations from nine Illinois State Water Survey stations with satellite and reanalysis soil moisture estimates from Soil Moisture Active Passive Level 4 Carbon Product Version 7 (SMAP L4C Version 7) and North American Land Data Assimilation System Phase 2 (NLDAS-2) model outputs (VIC, NOAH, MOSAIC). Meteorological forcing is obtained from Gridded Surface Meteorological Dataset (GRIDMET) for calibration and Localized Constructed Analogs Version 2 (LOCA2) for future projections. Calibration targets multiple key parameters that control storage dynamics and partitioning processes including available water capacity, wilting point, drying coefficient, runoff shape factor, and Potential Evapotranspiration (PET) scaling coefficients. Using JAX-based automatic differentiation, we evaluate thirteen loss functions and identify three, Root Mean Square Error (RMSE), Outer 50 Percent Root Mean Square Error (Outer50RMSE), and Kiling-Gupta Efficiency (KGE), as the most informative based on performance over the full record, the driest five days per year, and the wettest five days per year. Parameter comparisons reveal robust differences between calibration sources: wilting point is systematically higher when calibrated with in-situ data, even when the ensemble is expanded across alternative loss functions. In contrast, available water capacity does not show a consistent separation between satellite- and in-situ-based estimates. Residuals exhibit slight seasonality, with the Outer50RMSE trained models showing the largest variance. To assess ensemble coverage, we introduce an ensemble coverage metric defined as the ratio between the intersection of ensemble spread and observed soil moisture relative to the observed range. In 6 of 9 counties, satellite-based calibrations produce higher coverage, indicating that multi-source calibration can better represent the overall distribution of soil moisture despite the limited temporal record of in-situ data. Projection ensembles generated using seven-year versus twenty-year calibration windows exhibit consistent drying signals across counties, and longer calibration periods reduce the spread of extreme projections while stabilizing parameter distributions. Overall, the results show that integrating in-situ, satellite, and reanalysis datasets with machine learning–enabled calibration improves model performance, enhances ensemble robustness, and provides more defensible future projections. However, the model still struggles to capture abrupt soil moisture declines and seasonal transitions, highlighting ongoing limitations in simplified water balance models when confronted with extreme hydrologic variability. The framework developed here offers a scalable pathway for generating county-scale soil moisture projections to support drought monitoring, agricultural decision-making, and climate resilience planning.