A deep learning approach to source localization and seabed classification using pressure time-series from a vertical array

Abstract

Source localization in a shallow ocean environment has historically been done using optimization techniques such as matched-field processing. However, such optimizations depend on the parameterization of the ocean environment. Due to the complexity of this physical system, some researchers are currently applying machine and deep learning techniques to source localization problems. We propose a convolution neural network (CNN) to better predict the source localization and seabed classification simultaneously using pressure time series waveforms from a vertical line array. Building on research using a CNN to classify the source locale and seabed type using waveforms from only one hydrophone, the method has been extended to a 16-element vertical line array. The additional hydrophones add more physical information from the system as well as simply more features for a CNN to learn source range, depth, and seabed type. The synthetic data were generated using a range independent normal-mode model for multiple ocean environments. Modifications to the CNN are made to exploit the multi-channel waveforms. Ocean acoustic applications require this accurate classification of source locale and seabed environment. In future work we will extend our technique and CNN model to work with real world data.