Using neural networks to estimate the probability density function of transmission loss in ocean environments with a database-driven model of uncertainty

Abstract

The prediction of transmission loss (TL) in the ocean depends on the relevant environmental characteristics, such as sound speed field, bathymetry, and bottom properties. These characteristics are often uncertain as they are commonly not all measured in situ, and are instead estimated from available databases. Thus, the resulting predictions of TL are uncertain, and this uncertainty can be quantified via the probability density function (PDF) of TL. Given a model for the environmental uncertainty, Monte Carlo techniques and thousands of realizations of the environment can be used to estimate the PDF of TL, but this method is computationally costly. A variety of alternative uncertainty-estimation techniques have been proposed, including a machine learning approach which uses neural networks trained on large Monte Carlo-produced TL datasets. The resulting neural network inherits the limitations of its training datasets and their underlying model of environmental uncertainty. A new environmental uncertainty model is described here which uses higher-resolution databases. A revised neural network approach to predicting PDFs of TL is developed and tested on ∼4.5 × 106 examples across 300 environments with the new, more realistic model of uncertainty at ranges up to 100 km for source frequencies between 50 and 600 Hz. [Sponsored by an NDSEG Fellowship.]