Remote sensing of ocean sound speed profiles by a perceptron neural network

Abstract

A method is developed to estimate ocean sound speed profiles through synthesis of remotely measured environmental data and historical statistics of sound speed obtained at a remotely sensed location. Sound speed profiles are represented by an expansion of empirical orthogonal functions (EOF) of the historical sound speed variation, while the remotely sensed environmental data provide real-time information to determine the expansion coefficients. Environmental inputs are limited to sea surface temperature available from satellite infrared sensors, acoustic time-of-flight and ocean bottom temperature measurable from bottom mounted acoustic and thermal transducers. A multilayer perceptron neural network is implemented to learn the functional transformation from the measured environmental input to the desired EOF coefficient output on a set of representative sound speed profiles. Sea surface temperature, time-of-year, and time-of-flight from the acoustic multipath that maximally samples the vertical sound speed are found to be the dominant inputs. The trained network is computationally efficient and produces estimates for untrained environmental inputs with a mean error of 1.1-4.4 m/s.