Comparative Analysis of Machine Learning Approaches for Spatio-Temporal Mapping and Migration of Groundwater Contamination in Northern Ontario

Groundwater contamination from legacy industrial activities poses serious environmental and public health risks, particularly in Indigenous communities reliant on untreated groundwater sources. This study compares three machine learning (ML) approaches—Sparse Principal Component Analysis (SPCA), Principal Component Analysis with Universal Kriging (PCA + UK), and Sparse Generative Topographic Mapping (SGTM)—to assess their effectiveness in spatio-temporal mapping of groundwater contamination in a First Nation community in Northern Ontario, Canada. Using hydrogeochemical monitoring data from 2005 to 2022 collected across 18 wells in the upper aquifer, each method was assessed for spatial resolution, interpretability, and robustness to data sparsity. SPCA identified dominant contaminants (e.g., arsenic, nitrate, sulfate) and persistent contamination zones, offering high interpretability and validation against geoelectrical survey data. PCA+UK achieved high-resolution spatial interpolation, capturing directional migration trends and evolving contamination plumes over a two-decade period. SGTM provided nuanced detection of nonlinear and heterogeneous patterns, revealing clustered contamination regimes with distinct geochemical profiles, including arsenic hotspots near residential areas. All three ML approaches consistently identified critical contamination sources and migration patterns, though each showed strengths for different objectives. These findings improve methodological approaches for environmental risk assessment and support evidence-based decision-making in areas affected by legacy pollution, providing insights for future applications of machine learning in hydrogeochemical modelling.