Multi-sensor UAV platform with machine learning for landmine detection

Abstract

There are approximately 110 million landmines buried worldwide across diverse ecosystems, made from a variety of materials including metal, plastic, and ceramics. These conditions drive the need for a cost-efficient, easily deployed, all-terrain, and automated approach to detection and remediation beyond the antiquated use of handheld metal detectors. One proposed solution is to combine geophysical remote sensing with UAV (Unmanned Aerial Vehicle) technology but current remote sensing designs are limited by weather and environmental conditions. GPR (Ground Penetrating Radar), for example, is best suited for environments where soil moisture content is low, and the substrate is homogeneous, which is not common for many minefields. Airborne landmine detection with UAVs currently use single remote sensing methods and often yield false flags, while multi-sensor systems are land-based, expensive to manufacture, and cumbersome to transport to remote sites. Our proposed solution is an integrated multi-sensor package mounted on a custom Airgility quad-copter UAV that uses machine learning to evaluate weather and terrain conditions, and chooses which on-board geophysical instrument or instruments will yield most accurate landmine detection results. Using a combination of novel and commercially available sensors, the UAV performs an initial survey to measure various environmental conditions and machine learning tools determine which on-board remote sensing method is appropriate, then the UAV scans for landmines as unexploded ordnance (UXO), preferentially weighing data from the chosen geophysical method and transmits information about potential landmine detection back to the remote operator station. Here we present the designs for the UAV platform, sensor package, remote sensing package, and information on how machine learning is incorporated. Future iterations of the design will explore landmine removal procedures. We gratefully acknowledge the support of NSF Grant Nos. EAR1417148 and EAR1909055.