Employing Machine Learning for Ocean Bottom Mapping: Reducing Waste and Increasing Efficiency

Underwater mines represent an asymmetric threat to US naval power. While Machine Learning is a powerful tool to enhance efficiency of mine detection, there are significant challenges to this methodology. This presentation discusses some existing solutions as well as proposed work to reduce the dependency on human involvement in mine countermeasures (MCM).

MCM efforts are divided into detection, classification, identification, and disposal. Our presentation focuses on how Machine Learning (ML) and Deep Learning (DL) can significantly enhance the classification stage, where targets are identified as benign or mine-like objects (MLO). Currently this process relies on computer-aided detection and classification based on texture, geometry, and spectral features, followed by human verification. This introduces the critical factor of human error which is inevitable given enough time and a large enough data set combined with fatigue and stress.

ML is a broad category of computer-based data detection in which the system detects and predicts patterns in provided images. One weakness of ML is that it must separately analyze and evaluate each component of the problem before combining them. DL, a subfield of ML, solves this by utilizing artificial neural networks (ANN). This allows for the rapid and accurate processing of vast amount of data. The drawback of DL is that before being implemented, it demands a large quantity of high-quality datasets to learn from. Unfortunately, there is a lack of publicly available datasets, and there is difficulty in gathering new data.

Possible solutions include sonar data simulation, data augmentation, and transfer learning. Sonar data simulation is a complex field as it is often difficult to manufacture data with the same levels of imperfections present in real data. This is because acoustic propagation is a complex process influenced by several factors affecting sound speed, all of which vary with the underwater environment. There is also limited amount of real data to compare and validate against. At higher levels of control, this presentation will describe the techniques used by this team to overcome these challenges.

In summary, ML and DL are the future of MCM, increasing efficiency and reducing potential errors associated with the human factor. If adopted we could greatly cut back the risk posed by mines to our navy’s ships and submarines, enabling the warfighter to be best prepared in combat situations.

PRESENTER

Lee, Andrew
andrew.m.lee37.civ@us.navy.mil
347-585-5484

NIWCLANT

CO-AUTHOR

Ryan Stefkovich
ryan.j.stefkovich.mil@us.navy.mil

CATEGORY

Usage of HPC to Support Modeling and Simulation that Informs Combat Decisions

SYSTEMS USED

Narwhal

SECRET

Yes