Assessing the Feasibility and Uncertainty of Estimating Multidepth Soil Moisture Over the Qinghai–Tibet Plateau Using Microwave Scattering and Regression Models

Abstract

This study assessed the feasibility and uncertainty associated with integrating multisource data, including active microwave observations, optical vegetation factors, terrain parameters, and soil properties, to estimate multidepth soil moisture (SM) across the Qinghai–Tibet Plateau (QTP) at four specific soil depths and over three profile soil depths (0–40 cm). First, the change detection model (CDM) was refined through vegetation scattering removal, solution boundary constraints, and water-balance constraints, which improved estimation accuracy at various depths. Second, multidepth SM estimation was regarded as a multi-input and multi-output (MIMO) regression task. Two advanced machine learning regression (MLR) models were used to evaluate the spatiotemporal extensibility and transferability of SM estimation. These methods were validated in four study regions with varying soil conditions and climatic conditions: NaQu, MaQu, Ngari, and Pagri. The results indicated that the modified CDM using temporal Sentinel-1 C-band synthetic aperture radar (SAR) data from 2015 to 2019 achieved satisfactory estimation accuracy. Across the four study regions, the root-mean-square error (RMSE) values for specific depths and profile depths were 0.038–0.057 and 0.038–0.052 cm3/cm3, respectively. Surface SM (SSM) estimates satisfied the RMSE threshold of <0.060 cm3/cm3, whereas deep SM estimates achieved RMSE <0.040 cm3/cm3. The regression models achieved higher spatiotemporal estimation accuracy (0.023–0.042 and 0.025–0.038 cm3/cm3). However, the regression models exhibited limited cross-spatial transferability. The introduction of scattering model and sample constraints significantly improved the cross-spatial transferability of the SM regression models. Moreover, both the microwave scattering and regression models demonstrated interannual transferability. The estimation errors for multidepth SM at different observation stations were influenced by their geographical environments (e.g., SM heterogeneity, surface vegetation state, and soil properties) and showed strong linear relationships with SM temporal variability. This study explored the complex interrelationships among in situ multidepth SM. By incorporating multi-source data sensitive to the spatiotemporal variability of SM, this study enabled the indirect estimation of deeper SM beyond the direct sensing depth of C-band SAR, achieving relatively high accuracy. The proposed framework shows particular promise for estimating near-surface to root-zone SM over plateau grasslands and bare-soil areas of the QTP during the nonfrozen period.