Probabilistic mapping of acid rock drainage in glacial retreat zones using multi-source remote sensing and machine learning

The accelerated retreat of Andean glaciers is one of the most evident impacts of climate change, with direct implications for environmental stability and water security. In this context, the progressive exposure of sulfide‑rich lithological units promotes the generation of Acid Rock Drainage (ARD), a process that degrades water quality and poses a threat to downstream ecosystems and water uses. Despite its environmental relevance, the study of ARD in high-mountain regions remains limited due to terrain inaccessibility and the site-specific and high-cost nature of traditional methods, which are based exclusively on field sampling and laboratory analyses.

This study presents an innovative methodological framework, implemented in Google Earth Engine, for the probabilistic mapping of ARD in glacial retreat zones of the Cordillera Blanca (Áncash, Peru). We integrated Sentinel‑2 surface reflectance imagery, spectral ratios sensitive to iron oxides, topographic variables derived from a 12.5m ALOS PALSAR digital elevation model, and an ordinal geological classification based on ARD generation potential. In addition, field spectral signatures resampled to the satellite sensor were incorporated through the Spectral Angle Mapper (SAM), providing independent physical information on mineralogical alteration processes.

Variable selection was performed through correlation analysis and multicollinearity diagnostics (VIF ≤ 5), ensuring a parsimonious and physically interpretable set of predictors. The performance of three nonlinear algorithms (Random Forest, SVM, and XGBoost) was evaluated under a spatial cross-validation scheme using 5 km hexagonal blocks, designed to minimize biases associated with spatial autocorrelation. Results showed that Random Forest achieved the best performance, with an AUC of 0.96 and an F1‑score of 0.90 under spatial validation, demonstrating strong generalization capability. Model interpretability analysis using SHAP revealed that the ferric iron index and SAM spectral similarity were the most influential predictors, confirming the importance of integrating field data into remote‑sensing‑based approaches.

The resulting probabilistic map identifies ARD hotspots concentrated in recently exposed periglacial zones, consistent with field observations based on physicochemical parameters and heavy‑metal analyses in water. This study demonstrates the effectiveness of combining remote sensing, machine learning, and geological knowledge for monitoring ARD in glaciated mountain ranges, providing a cost-effective and scalable tool that contributes to environmental risk management in a changing climate.