Cubist-Inspired Deep Learning with Sonar for UXO Detection and Classification
David Williams | NATO STO Centre for Maritime Research and Experimentation (CMRE)
An unfavorable legacy of former military activities at sites designated for base realignment and closure (BRAC) and at Formerly Used Defense Sites (FUDS) is the contamination of aquatic environments with military munitions. In the United States, more than 400 underwater sites, spanning an area in excess of 10 million acres, potentially contain such munitions. The presence of these munitions is a serious threat to both humans and the environment, so remediation is necessary. But the return of these contaminated waters to public use is contingent upon the analysis and assessment of wide-area and detailed underwater surveys. Therefore, the Department of Defense (DoD) has an express need for the development of technologies that will enable the detection and classification, at high probability, of military munitions found at underwater sites. The objective of this project is to develop a novel detection and classification framework for unexploded ordnance (UXO) that exploits synthetic aperture sonar (SAS) data. The new algorithms will be based on deep-learning techniques, specifically deep convolutional neural networks (CNNs). The successful development of this approach should enable the attainment of higher probabilities of detection and classification, at much lower false alarm rates, than is possible with existing approaches. As a result, the application of these machine-learning algorithms to sonar data collected at potentially contaminated underwater sites can guide remediation efforts to effect savings. Specifically, because fewer resources will be spent investigating harmless clutter, the cost of remediation should decrease substantially.
“Deep learning” is the generic umbrella term used to denote classification algorithms with architectures characterized by a nested functional structure that engenders highly nonlinear decision surfaces. The great capacity of deep-learning algorithms, such as deep CNNs, when paired with vast amounts of data and sufficient computational resources, has translated into state-of-the-art performance in diverse domains. In this project, new deep CNNs will be developed for UXO detection and classification. The inspiration for this work, perhaps surprisingly, is the avant-garde movement known as Cubism that revolutionized the art world in the early 1900s. A hallmark of this style of painting and sculpture was the depiction of multiple perspectives simultaneously. In an analogous manner, the new deep CNNs to be developed will incorporate multiple representations of the data – e.g., SAS imagery, wavenumber domain data, acoustic color plots – simultaneously. The key is that these alternative representations make certain relevant information accessible and therefore exploitable by a CNN. In the standard image domain, many of the salient clues for classification would effectively remain hidden. To leverage the prior deep-learning research in mine countermeasures (MCM), a significant component of this project will also involve exploring the feasibility of two types of transfer learning for CNNs.
This project will address the DoD’s need for robust detection and classification approaches for underwater UXO. The scientific community will benefit from the development of a deep-learning framework that, as a by-product, also automatically uncovers valuable classification features (via the learned CNN filters). The result of this work also has the potential to form a foundation for follow-on efforts that would seek to unify high-frequency and low-frequency sonar-data-based classification approaches.