Root Zone Soil Moisture Prediction over Nebraska Croplands

Accurate root zone soil moisture (RZSM) information is critical for agricultural water management. While satellites provide observations of Surface Soil Moisture (SSM), they are confined to the top few centimeters, necessitating models to infer RZSM from surface measurements. This is difficult in managed agricultural landscapes because of site specific inputs like irrigation. In this study, we evaluated a machine learning (Random Forest) framework against a recursive exponential filter to assess the predictability of RZSM using SSM. We assessed the model performance using in-situ data from three adjacent corn and soybean fields under two scenarios: i) temporal transferability (out of sample years) and ii) spatial transferability (out of sample sites). In temporal transfer, the ML model (R² = 0.74) outperformed the exponential filter (R² = 0.52), where the predictive skill was found to be predominantly determined by SSM variability. In the spatial transfer task, the exponential filter performed poorly (R² = -0.02), which resulted from the filter’s exclusive reliance on SSM, and inability to accept ancillary variables required to account for site-specific heterogeneity. In contrast, the ML model remained robust (R² = 0.63) due to the integration of ancillary remote-sensing data alongside SSM, specifically Alpha Earth embeddings and Sentinel-2 imagery that identified the irrigation regime and tracked crop phenology, respectively. This combination allowed the model to correct for the systematic underestimation in irrigated sites. Machine learning was able to maintain reliable performance in spatial and temporal transfers to out-of-sample validation cases, without elaborately improving the physics of the hydrology model.