As reflected in SERDP Statement of Need MRSON-20-C1 and the referenced 2017 Workshop on Acoustic Detection and Classification of Unexploded Ordnance in the Underwater Environment, automated detection and identification of abandoned munitions through sonar interrogation require that system algorithms be robust to a wide variety of environmental and target conditions, with particular focus on the critical effects of burial state and grazing angle on the signatures. This project team will incorporate the physics insight gained from their current effort into the classification system being developed for a multi-sensor towbody featuring a down-looking sonar configuration. This new classification architecture will produce orientation estimates and appropriately adjust the decision algorithm, and will incorporate physics knowledge for specific target types as they are studied. This effort will be applicable to both the automatic classification system and the tools being developed for post-mission analysis.

The approach has three main focus areas:

Classification based on orientation estimation: Preliminary analysis of the existing classification experiments suggests that well-performing black-box classifiers are doing implicit orientation estimation and applying different analysis to returns based on incident angle at the target rather than simply identifying some universal characteristics of returns that can be observed across orientation. The project team will develop a system which subdivides the problem explicitly by estimating target orientation; they will consider image-processing techniques based on synthetic-aperture sonar outputs, but the focus will be on matched-filter processing using free-field return models. The system will produce a confidence measure for the orientation estimate, as well as a potential recommendation on travel path for maximally informative follow-up scans. The deliverable will be a classifier which is specialized for particular target orientations when they can be accurately estimated, but will revert to the baseline system when orientation is uncertain.

Use of physics-based features for specific target classes: The current work is focused on the use of finite element (FE) models to relate the observables in the scattered return to the classes of underlying physics behaviors. The value of this connection lies in the ability to take existing basic physics knowledge, compounded with modeling of how such behaviors are affected by the expected sources of variability such as sediment type and burial state, to develop a feature representation that is robust to the predicted range of observables. The project team is developing a classification architecture that can incorporate such physics knowledge for individual target types as they are developed.

Physics-based assist for post-mission analysis: For ConOps in which potential targets have been identified and are being considered for remediation or further scans, a graphical user interface being developed at the Applied Physics Laboratory, University of Washington (APL-UW) provides a framework for a user to drill down into the collected data and explore a variety of data products. The project team will develop an interface to enable the operator to analyze and view the collected returns in the context of known physics behaviors of the potential target type being considered.

Both automatic classification and post-mission assist will be improved by this work, which will be explicitly incorporated into the down-looking sonar tow-body being developed at APL-UW. The result will be a system better able to handle buried targets, more robust against previously unstudied clutter types, and more adaptable to improvement as new target classes are added to the mission scope.