Machine learning for optimal unexploded ordnance mitigation

Abstract

Offshore infrastructure projects in the UK must map and, ideally, distinguish all WWI and WWII munitions from non-threatening objects, a task which is complicated by the challenges of underwater data acquisition and the diversity of munitions. Magnetic data is the only commercially viable geophysical data type capable of both detecting buried UXO and covering the large (and increasing) survey areas necessary for modern offshore infrastructure projects. In many projects, however, the analysis will yield hundreds or even thousands of targets, many of which will be false alarms due to the overly simplistic nature of the current methods and compounded by misconceptions about the relationship between the target and the corresponding magnetic signal.
As such, there are significant opportunities for improving the timescales and cost effectiveness of commercial site survey projects if the number of magnetic targets that require detailed inspection and/or mitigation can be reduced. To this end, within this thesis classical Machine Learning (ML) and Deep Learning (DL) approaches are successfully implemented to categorize targets into potential UXO or non-threatening objects via classification and to extract information about their position (X, Y, Z [where Z is depth plus sensor altitude]), volume and shape via regression.
Synthetic magnetic data, which mimics the positional noise and acquisition related signal aliasing, novel physics-based data augmentation technique, is used to create the large and heterogeneous dataset required to train the ML and DL. Various horizontal array acquisition configurations are created and compared to determine the minimum and optimal sampling required for detection and use within the ML and DL. Performance of the ML and DL regressions were measured through comparison with novel non-ML proxies for position and volume, whereas, due to the complexity of the task, the shape and the classification had no non-ML counterparts. All methods performed best with large arrays and significant overlap between lines, but the ML and, to a greater degree, the DL outperformed the non-ML methods, even with larger (i.e., less ideal) line and sensor spacing.
The significance of these results lies in the potential for considerable cost reduction in site surveys. The ability to successfully differentiate potential UXO from non-threatening objects, even for a portion of targets, could reduce the amount of time spent investigating objects. Additionally, the ability of the ML and DL to perform very well with fewer lines and larger sensor spacings could reduce the amount of vessel time to complete a survey. Given the current drive towards leveraging offshore energy resources for a low-carbon future, these advances are both timely and, potentially, highly impactful.

More information

Identifiers

ORCID for Sara Lise Macdonald Underhay: orcid.org/0000-0002-2833-4366

ORCID for Timothy Henstock: orcid.org/0000-0002-2132-2514

Catalogue record

Export record