Generation of long-term and seamless 1 km surface soil moisture dataset within the area of in-situ network

Surface soil moisture (SSM) plays a significant role in the energy exchanges and the complex interaction within the air–soil–water–plant-human nexus. To better evaluate and utilize the microwave remote sensing (RS) SSM products at coarse scale (e.g., 0.25°) and the retrieved SSM data at fine-scale (e.g., 1 km), a pixel-scale reference dataset should be generated within the area of in-situ network. However, in the Tibetan Plateau (TP), where in-situ SSM data is sparse and limited, the current fine-scale SSM datasets generated using machine learning (ML) methods face certain limitations in terms of spatial extrapolation capability. In this study, we developed a framework that integrated ML method with geostatistical spatiotemporal fusion method to generate long-term and seamless 1 km SSM dataset with higher spatial extrapolation accuracy. The study area included five ground observation network regions (Shiquanhe, Pali, Naqu, Heihe and Maqu). Firstly, the incomplete 1 km scale SSM was retrieved by upscaling the in-situ SSM using the Residual Dense Network (RDN) model. Then, the Bayesian maximum entropy (BME) method, considering the uncertainties of the upscaled SSM, was employed to spatiotemporally fuse upscaled and in-situ SSM to improve the accuracy of spatial extrapolation. Validation based test sites shows that the accuracy of the fused SSM data was improved across all five regions, with the improvement in ubRMSE ranging from 3.33% to 21.28%, resulting in an overall increase of 8.2%. The fused SSM can more effectively capture the temporal variability of the measurements of test stations. The results demonstrate that the proposed framework effectively generates a reference SSM dataset within the ground observation network area.