Well-constrained LDA and SVM classification of ERT/IP data for leachate mapping in urban landfills

Landfill leachate represents a major environmental concern due to its potential to contaminate groundwater. Geophysical methods such as electrical resistivity tomography and induced polarization are widely used to characterize landfill interiors and infer leachate plumes, as leachate typically shows distinct resistivity and chargeability relative to surrounding waste and natural soils. However, the interpretation of geophysical data is still largely qualitative or semi-quantitative, often leading to ambiguous results. To overcome these limitations, we propose a supervised machine-learning framework based on Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM), trained using resistivity (ρ), chargeability (m), and normalized chargeability (Mn) derived from well data. The classifiers applied to the full geophysical models obtain spatially continuous maps of leachate occurrence. Benchmark models (Logistic Regression and Random Forest) were tested to assess the role of model complexity under sparse training conditions. The results highlight the complementary strengths of LDA and SVM: LDA systematically achieves perfect recall (100%), providing a more inclusive reconstruction of leachate-affected zones, whereas SVM attains perfect precision (100%), enabling a more conservative delineation of leachate accumulations. In the three-parameter space (ρ, m, Mn), selected for both physical completeness and comparability with previous studies, both methods reach 95% accuracy, with F1-scores of 89% (LDA) and 86% (SVM), while Random Forest shows results comparable with SVM, and Logistic Regression yields lower performances. Overall, the proposed approach bridges the gap between point-scale well information and volumetric geophysical imaging, providing a more objective, reproducible, and operational tool for leachate mapping in complex urban landfills.