MULTI-LABEL CLASSIFICATION OF UNDERWATER SOUNDSCAPES USING DEEP CONVOLUTIONAL NEURAL NETWORKS

The detection and classification of passive sonar acoustics is a challenging problem faced by surface, subsurface, and naval air assets. The potential benefit of machine learning systems to assist in this task is appealing. However, little work has been conducted to develop and test machine learning models for this type of data or task. This thesis presents a custom convolutional neural network (CNN) model designed specifically for underwater acoustic classification. This model is compared to several common CNN architectures on two datasets of hydrophone recordings of passing ships. These datasets are some of the largest datasets of ship recordings used for training CNNs to date, composed of over 4,000 hours of recordings and hundreds of unique ships. This thesis's main contribution is in demonstrating multi-label classification on underwater ship acoustics where the proposed model achieved an average micro-F1 score of 0.97. The custom CNN shows marked improvement in performance over standard models in both multi-class and multi-label classification tasks. This work also presents research into the inclusion of synthetic ship sounds and their potential use in training classification models. This thesis demonstrates the capability of machine learning models to enhance human and unmanned systems operating in the undersea domain.Lieutenant, United States NavyApproved for public release. distribution is unlimite