Shallow Water Sound Speed Estimation with Neural Networks-Based Nonlinear Regression of Space-Time Variability

Traditional way to estimate propagation losses in a region with no actual measures is to use oceanographic climatologies, built from archived data. Usually such statistical models have 0.25°-1° resolution. While it is enough for large-scale ocean acoustic simulation, higher-resolution climatology reflects regional ocean state better. With increasing of resolution, size of the models also increases, which makes it difficult to use them in small autonomous underwater systems, such as underwater sensor networks nodes, where space and power resources are limited. To minimize the size of a model the artificial neural network regression is proposed. To check applicability of method, shallow water area near Jeju island (East China Sea) was choosen. Set of neural networks was trained on data from World Ocean Database 2013. To estimate the error of sound speed profile reconstruction data from SAVEX15 shallow water acoustic experiment was used. Although the RMS error of prediction was high, vertical gradients of sound speed profile was reconstructed with good accuracy, which was shown using propagation loss calculations.