An ensemble model of knowledge- and data-driven geospatial methods for mapping groundwater potential in a data-scarce, semi-arid fractured rock region

Abstract In remote arid regions of South Australia, local industries, agriculture, mining, and households rely on limited groundwater resources. Data scarcity often leads to drilling unproductive wells when siting new bores. This study introduces an innovative geospatial method for groundwater exploration using an ensemble mapping approach. It combines knowledge- and data-driven machine learning methods: fuzzy analytic hierarchy process (FAHP), multi-influencing factor (MIF), frequency ratio (FR), random forest (RF) and maximum entropy (MaxEnt) to map groundwater potential. The approach leverages the strengths of each method without relying on the bias of a single approach. Morris sensitivity analysis filters areas of higher uncertainty, enhancing knowledge-driven methods before ensemble integration. Spatial representation shortcomings are addressed for key parameters, including drainage density weighted by stream order, terrain curvature integrated into slope models, yield-distance analysis for lineament density, and combining underlying lithology with surface geology to represent water- and non-water-bearing formations at depth. Each groundwater potential model’s performance was evaluated using receiver operating characteristic (ROC) curves and area under the curve (AUC), with the MIF model producing the lowest AUC of 85.41%. Although the study focuses on the arid township of Leigh Creek in the northern Flinders Ranges, the methodology is applicable to other regions with minimal well datasets worldwide. This research also contributes to addressing the scarcity of geospatial groundwater potential studies in Australia.