Multi-layer high-resolution soil moisture estimation using machine learning over the United States

The lack of proper understanding of multi-layer soil moisture (SM) profile (signals) remains a persistent challenge in sustainable agricultural water management and food security, especially during drought conditions. We develop a machine-learning algorithm using the concept of learning from patterns to estimate the multi-layer SM information in ungauged locations firmly based on local knowledge of the climatic and landscape controls. The Contiguous United States (CONUS) is clustered into homogeneous regions based on the association between SM and climate and landscape controls. Extreme Gradient Boosting (XGBoost) algorithm is applied to homogenous regions to capture the complex relationship between appropriate predictor variables and in-situ SM at multiple layers over the CONUS. Soil Moisture Active Passive (SMAP) Level 4 (L4) surface (0–5 cm) and rootzone (0–100 cm) SM along with climate and landscape datasets are used as predictor variables. In-situ multi-layer SM recorded by Soil Climate Analysis Network (SCAN), Snow Telemetry (SNOTEL), and U.S. Climate Reference Network (USCRN) networks are utilized as predictands. XGBoost models are then trained region-wise and layer-wise to estimate multi-layer SM information at 5, 10, 20, 50, and 100 cm depths (five layers) at 1-km spatial resolution. Results indicate that the predictor variables have varying levels of influence on SM with changing soil depth, and meteorological variables have the least importance. Validation at 79 independent locations indicates the multi-layer SM estimates successfully capture temporal dynamics of SM, with most locations achieving ubRMSE less than 0.04 m3/m3. The high-resolution SM estimates offer spatial sub-grid heterogeneity compared to SMAP L4 SM.