Prediction of Landmine Contamination Risk Using a Machine Learning Model Ensemble

Abstract

Landmine contamination remains one of the most serious humanitarian challenges in modern Ukraine. As a result of the full-scale war, vast areas have been mined, and according to international organizations, nearly one quarter of the country’s territory is considered potentially dangerous. Traditional demining methods require substantial time and resources, creating the need for intelligent systems capable of predicting landmine risk, automatically identifying high-hazard areas, and improving demining planning. This study proposes a method for predicting landmine contamination risk using machine learning and geospatial technologies. A multilayer dataset was constructed, incorporating geospatial, socio-economic, and military features. The spatial grid covers more than 0.55 million cells sized 500×500 m, ensuring a high level of detail in mapping contaminated areas. Four baseline algorithms: Logistic Regression, Random Forest, XGBoost, and Invariant Risk Minimization were employed for data analysis. Based on these models, an ensemble architecture using the Stacked Generalization approach was implemented, where a meta-model integrates the outputs of the base learners into a unified prediction. Model performance was assessed using AUC, Precision, Recall, and F1-score metrics. Experimental results show that the ensemble outperformed the best individual model (XGBoost), confirming the effectiveness of combining heterogeneous algorithms to improve prediction accuracy. The study demonstrates that ensemble machine learning methods enable accurate and automated prediction of landmine contamination risk, enhancing demining planning and contributing to improved civilian safety.