An Analysis and Early Warning of Toxic Gas Outbursts in Small-Scale Gem Mining: Model Evaluation and GIS-Based Risk Mapping in Sri Lanka

Abstract

Artisanal and small-scale gem mining in Sri Lanka, particularly within the Pelmadulla region, is highly susceptible to toxic gas accumulation due to inadequate ventilation and the absence of systematic early-warning mechanisms. This research aimed to develop a predictive–GIS integrated framework for the detection and spatial mapping of toxic gas hazards in small-scale mining environments. Utilizing globally available gas sensor datasets (UCI, IEEE, Mendeley) and localized geological and spatial data, multiple predictive algorithms—Random Forest (RF), XGBoost, LSTM, GRU, TCN, and IBWO-TCN—were trained and evaluated using precision, recall, F1-score, AUROC, and lead-time metrics. The Random Forest model exhibited the highest predictive performance (F1 = 0.93, AUROC = 0.97) and was subsequently integrated with GIS-based hazard mapping for the Pelmadulla study area. The spatial analysis indicated that approximately one-third of mining sites fall within high or very high-risk zones. The findings highlight the potential of integrating predictive analytics and geospatial modeling to establish a low-cost, data-driven early-warning system, thereby enhancing occupational safety and supporting sustainable mining practices in Sri Lanka.