Construction of Sound Velocity Field Based on Deep Learning Combined With WOA18 Temperature and Salinity Model

The change in sound velocity is an important factor affecting the precise positioning of underwater. Due to the existing sound velocity profile acquisition methods, the current sound velocity representation error seriously affects the underwater positioning accuracy. In view of the difficulty in realizing continuous temporal and spacial observation in a certain sea area in reality, this paper uses the temperature and salinity data of the Array for Real-time Geostrophic Oceanography (Argo) as the true value and a neural network model based on Long Short Term Memory (LSTM) with added attention mechanisms as well as the World Ocean Atlas 2018 (WOA18) historical thermohaline data to construct local ocean sound velocity field. The results show that this method can be used to retrieve relatively accurate sound velocity profiles in the depth range of 500-1 500 m in local waters of the Pacific Ocean, and the root-mean-square error of sound velocity inversion by LSTM neural network model with added attention mechanism is 0.34 m/s. The root-mean-square error of sound velocity in the local waters of the Atlantic Ocean is 0.78 m/s. Compared with the traditional Back Propagation Neural Network, the accuracy of BPNN and Back Propagation Neural Network Genetic Algorithm (BPNN+GA) neural networks with added genetic factors has been improved.