Machine Learning-Based Root-Zone Soil Moisture Estimation Using Satellite-Derived Surface Soil Moisture

Root-zone Soil Moisture (RZSM; 10–102 cm) is a critical variable for land–atmosphere interactions, plant water availability, groundwater recharge, and hydrological extremes; however, its reliable estimation at deeper layers over large spatial scales remains challenging. Ground-based monitoring networks such as the International Soil Moisture Network (ISMN) provide accurate multi-depth soil moisture observations, but their utility is constrained by sparse station distribution, high installation and maintenance costs, and limited spatial coverage (Dorigo et al. 2011). In contrast, microwave remote sensing based satellite missions, including Soil Moisture Active Passive (SMAP), Soil Moisture and Ocean Salinity (SMOS), and Sentinel-1, offer frequent and spatially continuous SM observations but are sensitive only to near-surface conditions (top ~5 cm), leaving deeper soil layers unobserved. This disparity between depth-limited in-situ observations and surface-focused satellite measurements motivates the present study to develop a machine learning based framework to estimate RZSM from satellite-derived surface SM by incorporating temporal memory and forcing. This approach effectively captures persistence effects and vertical moisture transfer, which are essential for accurate prediction of deeper SM layers (Pal &Maity, 2019). Multi-depth SM observations from 5 to 102 cm, obtained from ISMN stations and categorized according to USDA Hydrologic Soil Groups (HSG A–D; four stations per HSG), account for differences in soil water movement and retention behaviour (Ross et al. 2018). For each soil group, Support Vector Regression (SVR) and Random Forest (RF) models were trained using a sequential, depth-wise prediction strategy comprising four depth transitions: 5–10 cm, 10–20 cm, 20–51 cm, and 51–102 cm. Model evaluation demonstrates strong predictive performance across all depth intervals (R² = 0.85–0.95 for RF and 0.63–0.95 for SVR at validation sites), indicating that HSG classification effectively captures soil-specific SM dynamics. The trained models successfully generate comprehensive RZSM profiles using satellite-derived SM from the SMAP mission.These profiles are rigorously validated against ground-based observations and demonstrate strong applicability across diverse landscapes lacking direct subsurface measurements.