Predictive Analysis of Landmine Risk

Abstract

Demining is a highly impactful but complex problem which requires considerable resources and time. Land mine detection is the most hazardous and time consuming of the tasks in the demining pipeline. Currently, the risk of landmines being present in an area is estimated on the basis of non-technical surveys which are expensive and slow. This paper presents a novel spatial landmine risk prediction model to help and improve the allocation of resources in demining operations and even predict future areas of interest. Our approach is based on training predictive models on geographical and social development data for areas recorded to have been demined in the past. We then use this model to predict areas with high chance of mine presence in the vicinity of the demined area so as to progressively expand the area of operations. We explore weighted classification and biased scoring methods to improve the performance of our base logistic regression and support vector machine models. Refinement of conventional models allows us to tackle the problem of unbalanced datasets in our application. The resulting pipeline is then characterized in terms of various performance metrics. The results show that the pipeline has a potential to provide reliable predictive information based on historic demining data, which can help organizations plan their resource allocations in future demining operations.