Citation

Baur J, Steinberg G, Nikulin A, Chiu K, de Smet T. J. Conv. Weapons Destr. 2021; 25(1): e29.

Copyright

(Copyright © 2021, James Madison University, Center for International Stabilization and Recovery)

DOI

unavailable

PMID

unavailable

Abstract

Two rapidly emerging technologies revolutionizing scientific problem solving are unpiloted aerial systems (UAS), commonly referred to as drones, and deep learning algorithms.1 Our study combines these two technologies to provide a powerful auxiliary tool for scatterable landmine detection. These munitions are traditionally challenging for clearance operations due to their wide area of impact upon deployment, small size, and random minefield orientation. Our past work focused on developing a reliable UAS capable of detecting and identifying individual elements of PFM-1 minefields to rapidly assess wide areas for landmine contamination, minefield orientation, and possible minefield overlap. In our most recent proof-of-concept study we designed and deployed a machine learning workflow involving a region-based convolutional neural network (R-CNN) to automate the detection and classification process, achieving a 71.5% rate of successful detection.2 In subsequent trials, we expanded our dataset and improved the accuracy of the CNN to detect PFM-1 anti-personnel mines from visual (RGB) UAS-based imagery to 91.8%. In this paper, we intend to familiarize the demining community with the strengths and limitations of UAS and machine learning and suggest a fit of this technology as a key auxiliary first look area reduction technique in humanitarian demining operations. As part of this effort, we seek to provide detailed guidance on how to implement this technique for non-technical survey (NTS) support and area reduction of confirmed and suspected hazardous areas with minimal resources and funding.

Language: en