：Soil retention services are critical for mitigating erosion and maintaining ecological stability. In the Loess Plateau, spatial zoning of these services supports targeted ecological management and sustainable development. However, the effects of spatial statistical unit scale on zoning outcomes remain underexplored, limiting the accuracy of region-specific strategies. To address this, we developed a spatial zoning indicator system for soil retention services based on the Pressure-State-Response (PSR) framework. Integrating Self-Organizing Feature Map (SOFM), Uniform Manifold Approximation and Projection (UMAP), and K-means clustering, we conducted zoning across three spatial units—grid cells, watersheds, and county-level regions. Our analysis revealed that zoning outcomes differ significantly in zone number, distribution, and clustering patterns across scales. Natural factors dominate zoning at grid and watershed levels, while socio-economic factors are more influential at the county level. Using a multivariate stratified geographical detector, we found that the watershed scale achieved the highest zoning quality. Additionally, machine learning models indicated significant spatial heterogeneity in the contribution and direction of key influencing factors at the watershed scale. These findings highlight the scale-dependent nature of soil retention service mechanisms and the need for differentiated management strategies. Our framework offers a new approach for analyzing and managing ecosystem services in the Loess Plateau and similar ecologically sensitive regions.