Machine learning as an interpretive consistency approach for coupling geochemical and geophysical domains in data-limited landfills

Understanding landfill-related groundwater contamination is challenging in heterogeneous hydrogeological settings characterized by sparse monitoring networks and limited subsurface information, constraints common to many environmental and earth-science systems. Based on our previous investigations, we present an interpretive framework that synthesizes machine learning (ML), hydrogeochemistry, and subsurface geoelectrical imaging to characterize leachate plume dynamics and surface methane emissions. We examine two Canadian case studies: a closed, non-engineered municipal landfill and an industrial bark dump near an Indigenous community. At both sites, direct-current electrical resistivity (ER) and induced polarization (IP) surveys were conducted, and the resulting data were inverted to produce 2D and pseudo-3D tomographic images of leachate-related anomalies. In the municipal landfill, a supervised Adaptive Neuro-Fuzzy Inference System (ANFIS) employed geoelectrical proxies of waste stabilization to infer landfill-scale methane-emission trends, with localized misfits attributed to near-surface biogas attenuation and metallic-waste-related high IP responses. Gaussian Mixture Modelling (GMM) and stacked pollution index (SPI) mapping of hydrogeochemical data delineated contamination zones used to validate low-ER anomalies at well-screening depth. For the industrial landfill, hydrogeochemical facies were identified from water-quality parameters using unsupervised ML methods (GMM, Hierarchical Agglomerative Clustering, and Self-Organizing Maps + K-means), supporting the interpretation of ER and IP anomalies associated with plume migration, mixing interfaces, and matrix-controlled attenuation. These studies illustrate a methodological framework in which ML links geophysical proxies and environmental processes. By integrating geophysical and hydrogeochemical information and evaluating cross-domain consistency using multiple ML methods, this framework reduces interpretive ambiguity and strengthens process-based conceptual models in data-limited settings.