Bridging Expert and Data-Driven Models: AHP and Scaled Machine Learning for Groundwater Potential Mapping in the FCT, Nigeria

Abstract

Groundwater is the principal source of potable water in the crystalline basement terrain of the Federal Capital Territory (FCT), Abuja, Nigeria, yet borehole failure remains common because aquifers are discontinuous and fracture controlled. Groundwater potential zone (GPZ) mapping supports the strategic siting of productive boreholes, and both expert-based multi-criteria decision analysis (MCDA) and data-driven machine learning (ML) are now widely applied to this problem. This study asks two questions that are specific to the FCT rather than novel in principle: how the choice of data-scaling method affects ML-based GPZ predictions in this terrain, and how closely an expert-driven MCDA delineation and a data-driven ML delineation built from the same thematic layers agree. Fourteen candidate thematic layers derived from remote sensing, a digital elevation model, and geological and soil data were compiled, of which twelve were retained as ML predictors, and records from 184 boreholes across the six area councils of the FCT (84 productive and 100 low-yield) supplied the labels. An analytic hierarchy process (AHP) weighted MCDA model was implemented in ArcGIS Pro, and 34 supervised classifiers were trained in MATLAB under four scaling schemes (standardization, maximum absolute scaling, maximum-minimum scaling, and normalization). Averaged across the 34 classifiers, the mean area under the ROC curve (AUC) ranged from 0.86 (standardization) to 0.91 (maximum absolute scaling); because tree ensembles are scale-invariant by construction, the scheme-level differences are expected to concentrate in scale-sensitive learners such as k-nearest neighbours and support vector machines. Neural networks and tree-based ensembles achieved the highest performance, with bagged trees giving the best test accuracy (88.9 %) and the leading neural networks AUCs near 0.95. Feature-importance analysis ranked lithology and drainage density as the leading controls. The AHP model returned a consistency ratio of 0.06 and a five-class GPZ map (about 25 % very good to 26 % very poor) that agrees, in broad spatial pattern, with the ML delineation, both indicating higher potential in the wetter, gently sloping south and lower potential in the north. The results suggest that data-scaling choice should be reported and justified in ML-based GPZ mapping, and that expert-based and data-driven maps can serve as mutual cross-checks for groundwater development in basement terrains.