Continuous Learning for AI-Informed Landmine Risk Assessment

Humanitarian demining efforts aim to assist 60 million people living in landmine-contaminated areas across 70 countries, with complete clearance requiring over 1,100 years at current rates. Recent AI-informed risk assessment systems seek to improve resource allocation but operate statically, providing fixed contamination maps that remain unchanged throughout operations and overlooking the dynamic nature of mine action, where field surveys continuously reveal ground truth that could refine risk assessments. We address this gap by conducting the first systematic evaluation of continuous learning strategies for landmine risk assessment using real operational data from Colombia. We model demining operations as a sequential resource allocation problem under budget constraints and evaluate six learning strategies, comparing passive baselines against retraining approaches, active learning, and active search methods. Our evaluation reveals that straightforward greedy retraining achieves 98.7% average recall under 50% query budget, substantially outperforming competing approaches, while active search methods achieve competitive performance when no historical training data is available. Our results demonstrate the value of model adaptation in humanitarian AI applications under geographic distribution shifts and inform ongoing deployment efforts for continuous learning systems in operational mine action contexts.