This project will address the need to develop technologies to classify targets under complex conditions and identify specific types of targets; especially in a cluttered environment. The historically common practice of disposal of munitions across thousands of underwater coastal and inland sites across the United States and abroad has created a substantial remediation challenge. The potential hazards from unexploded ordnances (UXO) present ongoing needs for detection, localization, movement tracking, and classification. This is particularly important in shallow water environments, where munitions could impact human and wildlife activity. This project aims to build upon previous research in classification between UXO and non-UXO clutter in underwater environments. The project team will focus on characterization and classification of structural acoustic features among UXO through analysis of how geometric scattering and the elastic response of the target change over time in the acoustic return signal. This effort will then seek to develop a framework for interpretability of features as applied in deep neural network (DNN) classification of time varying frequency behavior in target return spatial spectral maps collected in a time volume for each UXO. This work will also evaluate classification decisions to inform what the most influential aspects of the signal in this domain are for characterization and classification.
This work will extend beyond traditional acoustic color domain analysis of frequency versus aspect angle for an entire time series acoustic return by looking at variations, patterns, and correlations that are only visible in the spatial time-frequency domain. The project team proposes the use of statistical data-driven methods to develop a sparse representation of the data as well as application of machine learning interpretability metrics in time sequence classification for refinement of the acoustic color time-frequency feature space. This study will consider acoustic target response data that is representative of the structural acoustics of each object. Data will consist of a mix of legacy data (Applied Research Lab at University of Washington [APL-UW] TIER models, Pondex Experiment 2010, and Target and Reverberation Experiment 2013) and recent collections under SERDP funding (Multi-Sensor Towbody dataset from APL-UW and/or the Sediment Volume Search Sonar data from Penn State with priority placed on recent data, and supplemented with legacy data where appropriate. In support of technical objectives, the primary research activities planned are intended to further the goal of associating physics driven acoustic phenomenon with targets and classification performance. The project team plans to do this by way of application of information measures, identification of mutual information among targets, and evaluation of deep neural network interpretability metrics over time series frequency distribution also known as time varying acoustic color data.
Time-frequency analysis of acoustic returns could offer insights into structural features present in UXO data and could also provide empirical statistical information on differences in geometric and elastic scattering among ordnances. This work will develop methods for extracting latent information from UXO returns, and quantify the most critical acoustic backscatter in frequency and aspect over time. Moreover, this work will contribute to techniques for the characterization of features among and between objects of interest. The expected result of this work is a selection of signal processing methods that proved most effective at improving classification of UXO by optimal selection of time-frequency features. Moreover, the use of interpretable DNN will provide insight into what frequency bands and time steps are more influential in classification to support increased sparsity in UXO acoustic time-frequency representation used for classification.