Soil moisture (SM) is crucial for agricultural production and ecological restoration, and high-resolution SM data can be efficiently acquired through downscaling remote sensing products. Although deep learning is widely used for downscaling, its mechanisms often lack interpretability. In this study, a downscaling model based on the architecture of the Kolmogorov–Arnold Network (KAN) was constructed using multi-source remote sensing data to generate SM data at 500 m resolution in the Tibetan Plateau. In addition, the interpretability of the model was effectively improved by fitting mathematical equations between SM and simplified variables. The results showed that (1) the downscaled SM of KANs demonstrated superior accuracy compared to ANN, significantly reducing the RMSE and ubRMSE of the SMAP product while closely aligning with station observations in temporal change trends; (2) correlation and SHAP analysis results showed that NDVI was the most important environmental factor affecting SM changes, followed by DEM and ET; (3) the symbolic equations of SM and simplified variables for each month were obtained by fitting the KAN model, whose downscaling accuracy was less than that of the KAN model but better than that of the SMAP product. Overall, this study explored the relationship between SM and environmental factors, and the obtained high-resolution SM data have good accuracy, which can meet the needs of small- or medium-scale research and provide a reference for subsequent downscaling studies.

As the contradiction between vegetation growth and soil moisture (SM) demand in arid zones gradually expands, accurately obtaining SM data is crucial for ecological construction. Remote sensing products limit small-scale studies due to the low resolution, and the emergence of downscaling solves this problem. This study proposes an improved semi-physical SM downscaling method. The effects of environmental factors on SM in different geographical zones (Windy Sand Hills, Flood Plains, Loess Yuan, Hilly Loess, Earth-rock Hills and Rocky Mountain) were analyzed using Random Forests. Vegetation and topographic factors were incorporated into the traditional downscaling algorithm based on the Mualem–van Genuchten model by setting weights, yielding 250 m resolution SM data for the Loess Plateau. This study found the following: (1) The Normalized Difference Vegetation Index (NDVI) was the most important environmental factor in all divisions except the Flood Plain, and the Digital Elevation Model (DEM) was second only to the NDVI in the overall importance evaluation, both of which positively influenced SM. (2) SM variability increased and then decreased when SM was below 0.4 cm³/cm³, but showed a quadratic growth trend when exceeding this threshold. The Rocky Mountain division exhibited the highest variability under the same SM. (3) Validation showed that the improved algorithm, based on geographic divisions to analyze factors importance and interpolation of coarse-scale SM and variability, had the highest accuracy, with an average R of 0.753 and an average ubRMSE of 0.042 cm³/cm³. The improved algorithm produced higher resolution, more accurate SM data, and offered insights for downscaling studies in arid regions, meeting the region’s high-resolution SM needs.