Abstract

Soil retention services are vital for preventing erosion and maintaining ecological stability. On the Loess Plateau, spatial zoning supports targeted ecological management and sustainable development, but previous studies have underemphasized the influence of statistical unit scale, limiting the scientific basis for differentiated strategies. Employing a Pressure--State--Response framework, we developed a zoning indicator system and applied Self-Organizing Feature Map, UMAP dimensionality reduction, and K-means clustering at grid, watershed, and county scales. We found significant differences in the number, distribution, and clustering patterns of zones across scales, driven by the varying dominance of natural versus socio-economic factors. Natural drivers prevailed at grid and watershed scales, whereas socio-economic factors dominated at the county scale. Multilevel geographical detector analysis identified the watershed scale as optimal, owing to its ecological coherence and hydrological integrity. Machine learning further revealed pronounced spatial heterogeneity in factor contributions and directions at the watershed scale. These findings highlight the scale dependence of soil retention mechanisms and the need for site-specific management strategies. Our framework offers a novel approach to ecosystem service zoning, informing research and policy on the Loess Plateau and analogous regions.