A convolutional neural network applied to measured time series for source range and ocean seabed classification

Source localization and environmental inference are common problems in ocean acoustics requiring computationally intensive algorithms and knowledge of the search space. Convolutional neural networks (CNNs) learn useful features for making predictions directly from a gridded input signal circumventing the costly practice of selecting features or comparisons to a forward propagation model. To take advantage of these benefits, a CNN was trained and validated on simulated pressure signals generated using four different environments (sandy, muddy, and mixed-sediment layers) for several ranges (2 to 12 km) to make environment class and range predictions. The trained network was then tested on recorded signals from the APL-UW Intensity Vector Autonomous Receiver system during SBC2017. We found that the network could make predictions with 99% accuracy on the simulated validation dataset but struggled to make unbiased predictions on the measured data. The struggles highlight the need to incorporate all significant features in the training data, such as time alignment and addition of noise. The CNN has the power to make predictions on this type and clearly see the differences in the data, but investigations for creating the proper training dataset or building a deeper network are required. [Work supported from Office of Naval Research.]Source localization and environmental inference are common problems in ocean acoustics requiring computationally intensive algorithms and knowledge of the search space. Convolutional neural networks (CNNs) learn useful features for making predictions directly from a gridded input signal circumventing the costly practice of selecting features or comparisons to a forward propagation model. To take advantage of these benefits, a CNN was trained and validated on simulated pressure signals generated using four different environments (sandy, muddy, and mixed-sediment layers) for several ranges (2 to 12 km) to make environment class and range predictions. The trained network was then tested on recorded signals from the APL-UW Intensity Vector Autonomous Receiver system during SBC2017. We found that the network could make predictions with 99% accuracy on the simulated validation dataset but struggled to make unbiased predictions on the measured data. The struggles highlight the need to incorporate all significant...