Harnessing Variational Autoencoders and self-organising maps for groundwater contamination assessment in peri-urban Ghana

Abstract

Although advanced machine learning models have demonstrated considerable potential for environmental monitoring, their application to assessing groundwater contamination in Ghana's peri-urban areas remains inadequately explored and poorly understood. To bridge this gap, this study aimed to apply advanced non-linear machine learning techniques, specifically Variational Autoencoders (VAEs) and Self-Organising Maps (SOMs), to analyse groundwater contaminants in south-eastern Ghana. The study examines intricate relationships and patterns among various pollutants to provide a comprehensive evaluation of groundwater quality. All the physicochemical parameters evaluated fell within the WHO guideline values. The VAE and SOM analyses confirm dual-source controls on groundwater chemistry in the Birimian terrains, involving both natural geogenic inputs from silicate and mafic lithologies and anthropogenic impacts from settlements. Inverse loadings across latent dimensions captured spatial heterogeneity, separating lithology-driven variables (e.g., Na⁺, Ca²⁺, EC) from pollution markers (e.g., NO₃⁻, Cl⁻). SOM clustering further distinguished zones of minimal human influence from areas with localised contamination, such as Pb hotspots and elevated EC and salinity linked to mineralisation or saline intrusion. Scattered peaks in F⁻ and Cl⁻ suggested episodic anthropogenic inputs. The results reveal notable disparities in machine learning model performance based on target variable features; the Nitrate Pollution Index (NPI) yielded a Test R² of 0.983, indicating superior predictive accuracy. Conversely, challenges with the Fluoride Pollution Index (FPI) and Pollution Index of Groundwater (PIG) exposed limitations due to unmeasured geological factors and low variability. We propose a data-driven, scalable diagnostic tool for monitoring water quality that can be integrated into national frameworks. This tool has implications for Sub-Saharan Africa and other regions similarly affected.