Elastic Target Modeling for Physics-Based Automatic Classification
The motivating goal of this research was to develop modeling methods and physics insights to improve the state of the art in sonar-based automatic detection classification of elastic targets such as unexploded ordnance (UXO). In particular, in this effort the project team developed methods for relating the scattered field produced by an interrogated object to the structure and internal physics of that object so that they could better understand when particular return components will be observable and develop feature sets and classification algorithms tailored to these components. The project team describes motivations for data collection efforts supporting and directions they have taken toward this goal. This project focused on developing methods for relating the elastic physics of the targets of interest to the scattered returns and leveraging the spectral and temporal structure of those returns to improve automatic classification.
The primary results presented demonstrate two different approaches to relating observables in scattered returns to the underlying internal physics. First, the project team modeled the returns from a target manufactured for this project to break down the observed structures in the return and categorized those structures by the type of associated elastic behavior. They compared the modeled responses to the data collected in the ClutterEx17 experiment, both to validate the modeling and to understand the types of variation to which a feature set must be robust. Second, they analyzed two existing classification systems used in studies involving simulated data and data collected in various underwater environments for the purpose of identifying the features currently being used, testing system robustness, and to identify opportunities for improvement. Finally, the project team considered the implications of the movement to a down-looking sonar in the development of feature sets and classification architectures.
The overall assessment of the project was that the modeling approach can be successfully used to identify observables which can be understood in terms of the underlying physics and used as a basis for targeted feature sets. The ClutterEx17 data that was collected using the science target fabricated under this effort was compared to the models. The results were a good fit in terms of their ability to use the models to relate elastic target behavior components of the observed return. The project team is focused on a change in detection/classification systems to a multi-static download sonar.
The results and conclusions obtained in this effort will encourage follow-on work. Since the origins of this work in the SEED effort, the focus has evolved from a concept based around a monostatic side-look sonar to that of a multistatic down-look sonar. The implications of this change are significant; there are many advantages for this approach but they will require concepts of feature extraction and classification architecture that, while driven by the same underlying physics, are very different from the approach that has been taken in the side-look scenario. The remainder of the work in this effort will be adapted to classification strategies for this new system concept.