A Comparative Assessment of Machine Learning Techniques for UXO Detection using Geophysical Sensors

The global unexploded ordnance (UXO) crisis is a $200 billion dollar problem that causes 15,000-20,000 deaths every year. A geophysics-based, cost-efficient, easily deployed, all-terrain, and automated approach to detection and remediation is needed. One proposed solution is to combine geophysical remote sensing with UAV (Unmanned Aerial Vehicle) technology, but current designs are limited by weather and environmental conditions. Single sensor UAV methods often yield false flags, while current multi-sensor systems are land-based, expensive to manufacture, and cumbersome to transport to remote sites. We present a pilot study to detect analog UXO and explore environmental condition dependencies using ground penetrating radar (GPR), a proton precession magnetometer, and a 3-component fluxgate magnetometer. The survey area was chosen to be sufficiently removed from roadways and overhead power lines in order to minimize cultural noise sources. Three landmine analogs, made of (1) metal, (2) plastic, and (3) ceramic were buried at shallow depths (15-20 cm) and surveys were conducted in dry conditions (defined as a period without rainfall for at least 3 weeks) and wet conditions (defined as a period with consistent rainfall for at least 1 week). Environmental sensors measured the temperature, humidity, and soil moisture of the field area during each survey. After preprocessing sensor data for each environmental condition, we use them to train a variety of neural networks in order to detect the analog UXO. Here we compare the performance of multiple machine learning algorithm architectures for the processed GPR and magnetometer data, and discuss how environmental conditions impact the efficacy of each machine learning technique. Finally, we explore implications of these findings for the success of proposed geophysical sensor-fused, non-invasive, airborne platforms.