Soil hydraulic properties (SHPs) are essential for hydrological modeling, yet their large-scale measurement remains challenging. Pedotransfer functions (PTFs) provide an alternative for estimating SHPs at broader scales. Integrating geospatial data into machine learning (ML) frameworks can significantly enhance the predictive performance and generalizability of PTFs, highlighting their potential for large-scale SHP estimation. In this study, we developed spatially explicit PTFs for China's drylands using data from 4,382 sites (12,306 soil samples) combined with more than ten ML algorithms. Results demonstrate that ML models accounting for geospatial heterogeneity outperformed simpler ML algorithms substantially. By integrating soil, vegetation, climate, and topographic factors, the models improved the prediction accuracy of various SHPs, including saturated soil hydraulic conductivity (Ks) and van Genuchten parameters (8s, 8r, a, n), by 31 %-79 % compared to existing PTFs, exhibiting strong robustness for applications in China's drylands. The optimal PTFs were applied to a 500 m x 500 m regional map of soil and environmental variables, generating maps of five SHPs (Ks, 8s, 8r, a, n) across six soil depths (0-5, 5-10, 25-30, 55-60, 95-100, and 195-200 cm) in the study region. From these maps, four additional SHPs, i.e., field capacity (8fc), wilting point (8wp), plant available water (8pa), and soil macroporosity ((bm), were derived at the same depths and resolution. The SHP dataset reveals distinct spatial distribution patterns and vertical heterogeneity of SHPs across China's drylands. This high-resolution, deep-profile dataset provides a robust foundation for large-scale hydrological and land surface modeling in China's drylands.