Mapping of groundwater protection zones using expert-driven and machine learning methods: a case study of Yulin City, China

Groundwater protection is critical for sustainable water resource management, particularly in arid regions. However, current zoning methods show challenges such as data bias of expert-driven models and limited interpretability of machine learning models. To address these issues, using 16 hydrological datasets from Yulin City in northwest China, two methodological frameworks were constructed: one combining the traditional Analytic Hierarchy Process (AHP) with Geographic Information System (GIS), and the other combining machine learning methods with Principal Component Analysis (PCA) and Self-Organizing Map (SOM). Rather than proposing a novel hybrid model, this study establishes a comparative framework that serves as a prescriptive decision protocol: AHP-GIS provides a transparent, defensible basis for regulatory implementation, while PCA-SOM with SHAP analysis offers interpretable insights into data-driven patterns. The zoning results of these methods show high spatial consistency (81.10%) with some differences (18.90%). Both methods effectively captured medium to key protection zones, particularly in areas characterized by high groundwater yield, good water quality, and ecological sensitivity. SHAP analysis further explained methodological divergences: pollution resistance and mining intensity were the primary drivers of key protection zone in PCA-SOM (12.89%), contrasting with the expert-assigned priority to functional zone and water quality in AHP-GIS (20.06%). This dual-framework approach overcomes the limitations of individual methods by using AHP-GIS to address the black-box nature of machine learning for policy applications, while using PCA-SOM to counteract the subjective bias inherent in expert weighting. Comparisons reveal fundamental trade-offs between transparency and objectivity, pattern sensitivity, regulatory consistency and adaptability to complex spatial relationships. By providing a decision protocol for method selection based on specific management contexts, our findings offer actionable guidance for overcoming the limitations of current approaches.