Real-time estimation of SMAP soil moisture in mountainous areas and its impact on rainfall-runoff simulation

The role of soil moisture (SM) in determining natural hazards such as flash floods, debris flows, and landslides in mountainous areas is critical. However, the operational capacity to monitor SM in real-time is constrained by the temporal resolution of low-Earth orbit satellite platforms. To address this, we developed a machine learning (ML) framework that fills Soil Moisture Active and Passive (SMAP) mission SM’s temporal gaps during 1 ∼ 3 days satellite revisit intervals by generating real-time SM estimates at 36 km resolution in the Hengduan Mountain region (HDMR). The accuracy of the algorithms was evaluated, and their impact on rainfall-runoff simulation was examined. The seven ML models have been found to perform significantly better overall correlation than the direct estimation method. Out of these models, the Long short-term memory (LSTM) network model has been identified as the top performer, exhibiting the lowest RMSE of 0.021 m³·m⁻³. The error spatial patterns of these models align with the recommended quality of the SMAP product, where sparse vegetation regions show higher accuracy of inversion compared to dense vegetation regions. Additionally, verification work in the Wake catchment indicates that SM can vary significantly over the estimation period of 1 ∼ 3 days, and the discharge hydrographs simulated using real-time and delayed SM values exhibit notable differences.