Advanced Processing, Inversion and Uncertainty Quantification of EM31 HCP/VCP Profiles for Mapping Surface–Subsurface Interaction along the Hamiz River

Abstract

Alluvial aquifers underlying the Mitidja Plain in northern Algeria are vital freshwater sources but remain vulnerable to contamination from Oued El Hamiz, a river carrying substantial industrial, domestic, and agricultural pollutant loads. Because groundwater abstraction has intensified river infiltration along discrete, spatially heterogeneous pathways, conventional point-based monitoring (piezometers, sampling wells) cannot adequately delineate where contaminated river water enters the aquifer. This study presents an integrated geophysical–machine learning framework to map river infiltration vulnerability along a 3 km reach of the Oued El Hamiz corridor. A total of 476 EM31 electromagnetic induction stations were surveyed across three transects parallel to the river bank, with measurements acquired in both horizontal dipole (~3 m depth) and vertical dipole (~6 m depth) configurations. Two-dimensional apparent conductivity images were constructed to visualize subsurface heterogeneity, and a rule-based vulnerability index was developed by integrating hydraulic gradient, sediment texture, and proximity to the river. Unsupervised (K-Means) and supervised (Random Forest classification and regression) machine learning models were then applied to delineate hydrogeological zones, predict infiltration risk, and quantify controlling variables through feature importance analysis. Conductivity imaging revealed discrete low-conductivity "windows" corresponding to permeable sand and gravel deposits that act as preferential infiltration pathways, contrasted with higher-conductivity, clay-dominated zones that impede recharge. K-Means clustering identified four distinct hydrogeological zones, while the Random Forest classifier achieved 97.9% accuracy in predicting vulnerability categories, with the VD/HD conductivity ratio emerging as the dominant predictive feature—indicating that vertical stratification, rather than shallow conductivity alone, governs infiltration potential. Rule-based and machine learning approaches showed strong agreement (99.4%), and 125 critical infiltration zones were identified for priority field verification, including ten stations recommended for immediate monitoring well installation. These findings demonstrate that coupling EM31 geophysical surveying with machine learning provides an efficient, non-invasive methodology for characterizing river–aquifer interaction, offering actionable guidance for groundwater protection and management in the Mitidja Plain and comparable intensively exploited alluvial systems.