GIS-based evaluation of advanced supervised learning methods for groundwater spring potential modeling

Abstract

Groundwater, a critical resource for environmental sustainability and socio-economic development, is spatially governed by geological, topographic, and climatic factors. This study developed a GIS-based groundwater spring potential modeling method in the Zhangjiamao area, China, based on the transition zone between the Loess Plateau and the desert. By integrating 93 spring data and 12 multi-source heterogeneous factors, including terrain, hydrology, geology, and landuse data, the predictive performance of six supervised learning models Quadratic Discriminant Analysis (QDA), Linear Discriminant Analysis (LDA), Fisher’s Linear Discriminant Analysis (FLDA), Fuzzy Unordered Rule Induction Algorithm (FURIA), Random Forest (RF), and Bayesian Network (BN) was systematically compared, and corresponding groundwater spring potential zoning maps for the Zhangjiamao area were generated. Factor selection involved multicollinearity diagnostics, correlation analysis, and importance ranking. The most influential factors were distance to rivers (MDA = 8.83), elevation (MDA = 7.44), slope angle (MDA = 6.19), and lithology (MDA = 6.73). Models’ validation showed that all models performed well (AUC > 0.8), with the RF model performing best with AUC values of 0.904 (training) and 0.969 (validation). The standard errors were relatively small (0.0295/training, 0.0192/validation), indicating stable and reliable results. This study clarifies the mechanism of spring potential formation under geohydrological coupling, and offers a methodological framework to support sustainable groundwater development and management in arid and semi-arid areas.