SMAP Soil Moisture Simulation via GRU-ViT Model and Residual Learning

Obtaining high-precision soil moisture data is paramount for climate research, early warning of hydrological events, and safeguarding food security. However, achieving high-quality, large-scale soil moisture monitoring remains challenging due to the scarcity of in-situ measurement data. To address this, we propose a two-stage learning framework. In the first stage, we constructed a GRU-ViT model utilizing 12 environmental factors and SMAP L4 SM to generate daily soil moisture time-series products for the Loess Plateau spanning 2017 to 2023. In the second stage, a residual learning method based on in-situ data was employed to correct systematic biases in the initial products. Results indicate that the final soil moisture dataset, derived from GRU-ViT modeling and subsequent residual learning, achieves superior accuracy and reliability, with the R value improving to 0.792 and the error reducing to 0.024 m 3 / m 3 . This proposed two-stage simulation strategy not only offers an effective solution for regions with sparse ground stations but also provides a scientific basis for refined regional water resource management and disaster prevention and mitigation.