Machine Learning–Driven Remote Sensing Framework for Predicting Groundwater Pollution under Climate Change Scenarios

Abstract

:  Pollution poses a significant threat to global water security, a challenge that is increasingly intensified by climate variability, which alters recharge dynamics, contaminant transport pathways, and aquifer vulnerability. Conventional monitoring approaches, constrained by sparse well networks and limited spatial coverage, are often inadequate for capturing the heterogeneity of contamination plumes across complex hydrogeological systems. This study develops a machine learning–driven framework integrating multi-sensor remote sensing data and climate projections to predict groundwater contamination under changing environmental conditions. Supervised learning algorithms—including Random Forest (RF), Support Vector Regression (SVR), Gradient Boosting Machines (GBM), and Artificial Neural Networks (ANN)—were trained using satellite datasets (Landsat, Sentinel-1/2, and MODIS), meteorological reanalysis (ERA5), and in-situ groundwater quality observations (>23,000 samples) across three hydrogeological regions. The models incorporated spectral indices, land surface temperature, soil moisture proxies, terrain attributes, and climate variables within a spatiotemporal prediction framework. Results indicate that RF and GBM models achieved the highest predictive performance, with test-set coefficients of determination (R²) ranging from 0.79 to 0.87 for nitrate and 0.74 to 0.81 for salinity, and root mean square errors (RMSE) of 3.2–4.8 mg/L and 180–240 µS/cm, respectively. Heavy metal predictions showed moderate performance (R² up to 0.76), reflecting stronger dependence on localized geochemical controls. Ensemble modeling approaches improved prediction accuracy by 10–15% compared to single-model implementations, with overall classification accuracies exceeding 85% for contamination threshold exceedance.

Climate-informed models demonstrated enhanced generalization under extreme conditions, improving predictive skill by 0.04–0.07 in R² during drought years. Future projections under SSP2-4.5 and SSP5-8.5 scenarios indicate potential increases in groundwater nitrate concentrations of 8–28% by mid- to late-century, driven by reduced recharge and increased evapotranspiration. However, spatial transferability remains limited, with inter-regional model performance declining by up to 0.30 in R² due to hydrogeological heterogeneity and data scarcity. The study highlights the potential of integrating remote sensing, machine learning, and climate modeling for large-scale groundwater quality assessment and early warning systems. It further emphasizes the need for hybrid physics-informed approaches and targeted data acquisition to improve model robustness and operational applicability in data-sparse regions.