The objective of this project was to demonstrate a proof-of-concept regarding the feasibility of using convolutional neural networks (CNNs) for unexploded ordnance (UXO) classification. Within this context, one main strand of work focused on assessing the applicability of two forms of transfer learning for the task of underwater object classification: target-concept transfer and sensor transfer. The use of transfer learning would allow data collected during mine countermeasures operations, and from different but similar sensors, to be leveraged for the UXO problem. The other main task of the project was to develop a CNN framework that could exploit multiple representations of sonar data simultaneously. The idea underlying the use of multiple representations (derived from the same raw data) is that complementary classification clues would be made accessible in different representations.

Using real measured sonar data collected at sea, the feasibility of the two forms of transfer learning were successfully demonstrated. To address the second task, a novel CNN-based classification framework for multiple-representation data was developed. More specifically, a flexible approach that naturally extends to any number of representations was made by fusing the disparate representations at the penultimate dense layer of the CNN. Importantly, it was demonstrated that successful classification could be achieved with notoriously “data-hungry” CNNs – even when only limited training data were available – by leveraging domain expertise in the design of the networks. Employing CNNs with orders of magnitude fewer parameters than are typically used resulted in robust classifiers that generalized well even to objects not seen during training. The benefit of drawing on an ensemble of predictions obtained via various forms of multiple representations – in terms of input data representations, isometries, frequency sub-bands, aspect sub-apertures, multiple object views, and unique CNN architectures – was also illustrated. In particular, it was shown how this approach can greatly reduce the false alarm rate while maintaining the requisite high detection probability that UXO remediation demands.

Collectively, these results show that the proof-of-concept can be deemed to have been fully achieved. That is, the groundwork has now been laid for a comprehensive classification framework that could incorporate multiple data representations, from both high-frequency and low frequency sonar data, as well as from different sensors, in order to improve underwater UXO remediation efforts. Additionally, the research has opened another promising avenue for future follow-on work.

The main thrust of the proposed follow-on work is to leverage the developed CNN framework in conjunction with specially controlled experiments to learn principled, explainable classification features that can be tied directly to the wave phenomena of the physics involved. This plan envisions a symbiotic feedback loop in which features uncovered by the CNN – and related to specific object characteristics – help inform the understanding of the problem’s physics, and vice versa. Both components would contribute to enabling more efficient UXO remediation strategies from a more rigorous scientific perspective.