Abstract by Mason Acree
Physics and Astronomy
Exploration of Convolution Neural Networks for Source Localization and Seabed Classification using tones on a VLA
Source localization in a shallow ocean environment is complicated due to variations in the ocean environment such as different types of sediment found in the seabed. Incorrect seabed properties can change the source location estimates. Some researchers are currently applying deep learning techniques for either source localization or seabed classification separately, we are predicting both simultaneously. We use a convolution neural network (CNN) applied to five towed tones, which are recorded on a vertical line array to predict source range, depth, speed and concurrently determining the seabed type. Simulated spectrograms are generated for mid-frequency tonals over a long time period and are used to train a CNN. Our model is then tested on simulated samples that are designed to model environment mismatch like a real-world scenario. Our approach is enabled by how the time-varying source location and seabed both affect sound propagation in a shallow ocean environment. This dependence allows our CNN to focus on extracting features that are relevant for both tasks in parallel and show great potential for accurately predicting an acoustic source and seabed type.