Study on spatio-temporal simulation and prediction of regional deep soil moisture using machine learning

Deep soil moisture (SM) plays a crucial role in vegetation restoration, particularly in semi-arid areas. However, current SM products have limited access and do not meet the spatio-temporal scale and soil depth requirements in eco-hydrological research. Thus, this study constructs a random forest prediction model for SM at different depths by identifying driving factors and quantifying the correlation effect of vertical SM based on the international SM network dataset. Subsequently, the SMAP product is integrated into the model to expand SM from point scale to regional scale, yielding an SM data product with a suitable scale and continuous time and space. The results indicate that the correlation between precipitation and SM changes into the interaction between adjacent SM layers as the depth increases. The lag time of SM in the shallow surface layer (0-3 cm) to precipitation was 1 day, and there was no delay on the daily scale in the 3-20 cm layers of the three underlying surface types. The response time of 50 cm SM to 20 cm SM was 1-2 days in cropland and grassland and 2 days in forest. Slope, land use type, clay proportion, leaf area index, potential evapotranspiration, and land surface temperature were the key driving factors of SM in the Shandian River region. The random forest model established in this study demonstrated good prediction performance for SM at both site and regional scales. The obtained daily products had higher spatial fineness than CLDAS products and could describe the SM characteristics of different underlying surfaces. This study offers new ideas and technical support for acquiring deep SM data in arid and semi-arid areas of northern China.