Understanding uncertain acoustic propagation in deep ocean environments for feature engineering in corresponding machine learning tasks

Transmission loss (TL) predictions obtained using models of deep ocean environments are often uncertain due to imperfect knowledge of environmental properties such as sound speed, bathymetry, and seabed properties. These environmental uncertainties can be transferred to TL-prediction uncertainty by Monte Carlo (MC) sampling over environmental parameters and performing TL-field calculations to obtain an MC probability density function (PDF) of TL. Unfortunately, thousands of TL-field calculations are often required to quantify the TL uncertainty making this approach ill-suited to real-time applications. In an alternative, supervised learning approach, neural networks can be trained to quickly estimate the MC PDF of TL at a point of interest by analyzing the variability in the values of a baseline TL-field prediction within a region surrounding that point. The size, shape, and number of local TL region(s) used as inputs to the neural network can be reengineered by better understanding the means by which the uncertainties in environmental properties affect the TL uncertainty. This process and the resulting improvements in predictive performance are demonstrated for acoustic frequencies of 50 to 600 Hz, and source-receiver ranges up to 100 km. [Work sponsored by an NDSEG Fellowship.]