This work was concerned with the development of robust methods for detection and classification of military munitions in shallow underwater environments using data collected from synthetic aperture sonar systems. The first objective was to address the problem of detecting the presence of underwater munitions. The goal of the second part of this work was to develop a robust target classification method that could be applied to the detected contacts to discriminate munitions from non-hostile man made objects and competing clutter.
This project first addressed the problem of detecting the presence of underwater munitions using the adaptive multichannel coherence analysis framework. The detection hypothesis was that the presence of munitions in the sonar returns collected from a hydrophone array would lead to higher levels of coherence compared to the returns from the seafloor alone.
The framework for the second part of the project was developed based upon the matched subspace classifier (MSC) using multivariate Acoustic Color data extracted from the raw sonar returns. Scattering models developed by the Applied Physics Laboratory, University of Washington were acquired to generate the required training dataset for various unexploded ordnance (UXO) and non-UXO objects. This was done owing to the fact that actual sonar data from a wide range of UXO and non-UXO objects is scarce in realistic situations. Although, it may be somewhat ambitious to expect model data capture all the essential features of these objects for target characterization, it would provide clues on how to augment the training datasets using perhaps limited training samples from sonar returns of actual objects to improve the robustness in different environmental conditions. The classification hypothesis was that spectral content of the sonar backscatter display unique acoustic signatures providing excellent discrimination between different classes of detected objects.
The first objective found to produce excellent detection performance on a wide variety of sonar data sets. Detection results are presented on two sonar data sets. The first dataset was simulated using a combination of the target-in-environment-response model and Personal Computer-Shallow Water Acoustic Toolset.
The second part of the project yielded classification results of the MSC with incremental dictionary learning, kernelized MSC, and collaborative multi-aspect classification are provided using two real sonar data sets. The first data set, TREX13, was also collected in the Gulf of Mexico near Panama City Beach, FL, using a rail system. The second data set, BAYEX14, was collected in St. Andrew's Bay (Panama City, FL), also with a rail system. Both of these data sets provide realistic challenges, where factors such as schools of fish, water turbulence, seafloor roughness, and target range were prominent.
The detection and classiﬁcation algorithms developed in this project allow for near real-time assessment of large underwater areas using data collected from low frequency sonar systems. These algorithms not only give the user the ability to assess the degree to which the site is contaminated but also provide capabilities to localize and characterize individual detected objects in varying conditions. The developed methods could be useful in a multitude of sonar applications used to search or survey underwater areas including environmental and oceanographic studies, undersea exploration, the search for wreckage on the sea ﬂoor, and mine-hunting. Additionally, the application of the methods can be extended to other multi-sensory systems for remote sensing and surveillance.