Underwater sites impacted by unexploded ordnance (UXO) or discarded military munitions may pose an unacceptable risk to human health and the environment. Due to the explosive properties of munitions, the risk to humans of exposure to underwater munitions may be immediate and debilitating, if not fatal. SERDP reported that the U.S. Army Corps of Engineers and the U.S. Navy have identified more than 400 underwater sites in the United States that potentially contain military munitions. The objective of this project is to develop a novel domain enriched deep learning framework that will detect and classify UXO by exploiting data from a recently developed sonar system capable of producing three-dimensional synthetic aperture sonar (SAS) imagery. The said sonar system was developed under the sediment volume search sonar (SVSS) program, also supported by SERDP. Machine learning and especially deep learning approaches are versatile across complex environments and have demonstrated some promise for detecting UXO – yet the success of existing learning frameworks is heavily dependent on the quantity and quality of available training imagery. This project – by leveraging domain specific insights and prior knowledge – will alleviate this crucial shortcoming and pave the way for practically deployable machine learning algorithms which can guide remediation efforts.

Technical Approach

Deep learning based image analysis and classification algorithms have quickly emerged to supplant the state of the art for many real-world problems. The modeling capacity of deep neural networks combined with vast amounts of training data and sufficient computation has shown to yield unprecedented performance gains. When training imagery is limited or noisy though – which is invariably the case with SAS, the performance drops sharply. In this work, the project team plans for new architectural innovations via a deep learning framework that designs a prior information network in conjunction with the classification network. The prior information network leverages observations made in the recent SVSS program – that elastic scattering phenomena visible in captured images vary by the particular ‘munition target’ and hence may be used to distinguish between them as well as separate man-made munitions from naturally occurring clutter. The elastic scattering prior provides a regularization term in the learning of the classification network, thereby presenting the potential of enhancing its performance when training is ample and enabling graceful (as opposed to sudden) decay in classification accuracy in the regime of limited/realistic training.


This project aims to take a first step in the direction of consummating the promise of learning frameworks for UXO detection and classification. The prior guided deep learning framework should be of strong interest to researchers engaged in employing machine learning for environmental-sensing problems. While the SVSS imagery will be the focus of this program, the project team expects that the techniques developed would be applicable to other similar acoustic survey systems. The project team intends to focus on two-dimensional slices (images) cut along the following axis: along-track, cross-track and depth. Exploitation of the three-dimensional image data cube is the ultimate goal and achieving the same would entail several non-trivial algorithmic extensions.