Abstract: The eco-hydrological mismatch between artificial vegetation and available soil water resources in parts of the north-central Loess Plateau has resulted in widespread soil desiccation and the formation of dried soil layers (DSLs). However, limited knowledge of the spatial distribution of DSLs constrains the assessment of soil water sustainability. To quantify the spatial distribution of DSLs in the shrub-grassland regions of the Loess Plateau, identify the dominant environmental factors driving DSL formation, and develop predictive models for DSLs, this study conducted field investigations across 111 sample plots (including farmland, grassland, forestland, and shrubland) to analyze 0–5 m soil profiles. Three key indicators were evaluated: dried soil layer thickness, soil water content within the dried soil layer, and soil water deficit. A total of five types of environmental factors were collected to elucidate the spatial characteristics and driving factors of DSLs in the shrub–grassland regions. Traditional statistical methods and three classical machine learning algorithms were employed to construct predictive models for DSL spatial distribution. The results reveal widespread development of DSLs (0–500 cm depth) in the shrub-grassland regions, with an average DSL thickness of 312 cm, an average soil water content within the DSL of 9.05%, and a mean soil water deficit of 0.59. Vegetation status is identified as the decisive factor influencing the spatial distribution of DSLs, while field capacity emerges as the critical determinant of soil water content in the DSL. Machine learning methods demonstrate strong predictive performance for DSL spatial distribution, particularly the back propagation neural network (BPNN) (except for DSL thickness) and support vector machine (SVM) models. This study quantitatively characterized the spatial patterns of DSLs in the shrub-grassland regions of the Loess Plateau and established vegetation status, soil properties, and meteorological factors as key environmental drivers. Furthermore, the predictive capabilities of machine learning models offer potential for spatial analysis of these findings, enabling improved management and forecasting of soil desiccation across broader regions.