Machine Learning for Optimal Unexploded Ordnance Mitigation

Unexploded Ordnance (UXO) pose a significant threat for onshore and offshore infrastructure projects throughout the world, but especially in north-west Europe. Current UK/EU legislation requires all infrastructure projects to conduct a UXO site investigation and mitigate the risk of any suspect items identified, through avoidance or removal. Offshore mitigation can be hugely time consuming and expensive, often involving the inspection and avoidance/removal of 1000s of targets using divers for shallow water and remote submersibles for deeper water. The weakest link in this process is our inability to differentiate during site investigation a potential UXO (which requires mitigating) from a false positive (e.g., boulder, which does not require mitigating unless it is a piling/drilling hazard). Multiple geophysical survey techniques (multibeam echosounder, sidescan sonar, transverse gradiometer, and decimetre-resolution 3D seismic reflection) can each characterise the bed and shallow subsurface in different ways at high resolution, but with significant ambiguity in target characterisation. Historically, these different data sets are treated in isolation, preventing effective reduction in ambiguity. However, during a recent major infrastructure project on the Thames, a joint qualitative integration of multiple data sources produced a 60% reduction in remediation operations.