Evaluation and improvement of spatiotemporal estimation and transferability of multi-layer and profile soil moisture in the Qinghai Lake and Heihe River basins using multi-strategy constraints

The machine learning regression (MLR) algorithms have brought new insights into soil moisture (SM) estimation. However, few studies have explored the potential of MLR algorithms for multi-layer and profile SM modeling, as well as their spatiotemporal transferability, which are important for practical deployment and application. In this study, the dual-polarization C-band radar data was used as the core to construct a multi-layer and profile SM estimation framework, constrained by multi-source auxiliary data (optical vegetation descriptors, soil properties, and terrain factors). Validation was carried out in two new SM observation networks: the Qinghai Lake basin (QLB-NET) and Heihe River basin (WATERNET). The results shown that the multi-output and multi-input stacking strategy regression (SSR) model demonstrated excellent spatiotemporal extensibility (RMSE = 0.027–0.044 cm³/cm³) and interannual transferability (RMSE = 0.031–0.055 cm³/cm³) in multi-layer and profile SM estimation. However, the cross-spatial transfer accuracy of the SSR model was poor (RMSE > 0.060 cm³/cm³). To address this, two improvement schemes were proposed, focusing on the accessibility of in-situ observation data. The first involved introducing a small number of samples from the target domain to update the hyperparameters in the SSR model. The second method used initial estimates from a scattering model, namely the modified change detection model, to constrain the SSR model and improve cross-spatial transfer accuracy. Both schemes achieved satisfactory transfer accuracy. The former strategy reduced SM estimation errors by 40.8–72.8 % and 24.1–68.1 % across various soil depths for two study areas, while the latter strategy achieved reductions of 30.3–67.2 % and 22.4–68.8 %, respectively. Additionally, factors influencing SM estimation and transfer accuracy were identified, including station difference, SM variability, vegetation cover, soil properties, and imaging orbits. Surprisingly, due to relatively low temporal variability and sensitive to vegetation productivity, the spatiotemporal estimation and transfer accuracy of deeper SM (10–30 cm) was better than that of surface SM (0–10 cm) at most observation stations. The SSR model outperformed deep learning algorithms of different architectures in terms of spatiotemporal estimation and transfer accuracy, operating efficiency, and computational overhead. In conclusion, the framework proposed in this study offers new perspectives and application prospects for remote sensing estimation of multi-layer and profile SM.