RELand: Risk Estimation of Landmines via Interpretable Invariant Risk Minimization

What is it about?

We have designed and implemented a comprehensive pipeline for estimating landmine contamination risk to support surveys in humanitarian de-mining operations. The system (1) rectifies the labels in the literature and integrates different sources of information into a novel dataset for landmine presence with rich relevant features; (2) uses an innovative interpretable deep learning tabular model with Invariant Risk Minimization to generate risk prediction estimates, outperforming current practices experimentally; and (3) presents the model output through an interactive interface to support the allocation of demining resources and the prioritization of land release operations. We are currently conducting a field study to test the proposed system to support the clearance of two recently prioritized areas in Colombia. Looking ahead, we aim to leverage the insights gathered from the field study to refine the system in its three components, preparing it for deployment on a larger scale including in other nations. The presence of landmines and other explosive remnants of war in over 70 countries, affecting more than 60 million people globally, underscores the urgent need for innovative solutions. Responsible ML-informed systems hold the promise of widespread application and positive impact on communities affected by conflict worldwide.

Featured Image

Photo by Evgeniy Kozlov on Unsplash

Photo by Evgeniy Kozlov on Unsplash