Effect of environmental uncertainty on source localization from mid-frequency tonals using convolutional neural networks

In ocean acoustics, the lack of labeled training data creates the need to train on simulated data. The simulated training data need to represent the environmental variability in the anticipated area where acoustic data are collected. Further, the manner in which environmental uncertainty should be incorporated is a significant question facing all ocean acoustics applications of machine and deep learning. To begin addressing this question, a case study is presented using mid-frequency tonal data recorded on a VLA. The time-varying tonal levels are input to a convolutional neural network (CNN) which finds the source range, depth, and speed and seabed type. The CNN is trained on simulated data and then tested on simulated data with different sound speed profiles and seabed types than used in training. The impact of this mismatch on the CNN predictions highlights the need to carefully account for environmental variability during training in order to provide robust machine and deep learning applications in ocean acoustics. [Work supported by the Office of Naval Research.]