Ship azimuth prediction using supervised machine learning

Abstract

Machine learning methods are applied to noise from the R/V Endeavor across several days during the SCEx17 experiment to predict the ship's azimuth. Sample Covariance Matrices (SCMs) are formed from received pressures on two vertical line arrays (VLA1, VLA2) and one horizontal line array (SWAMI). Previously, support vector machine (SVM), feed-forward neural network (FNN), and random forests (RF) machine learning methods have accurately estimated the range of a source towed in a linear geometry in shallow water (Niu et al., JASA 142, 1176–1188 (2017); Niu et al., JASA 142, EL455–460 (2017)) where the training and test tracks were close together in time. In this study, we investigate the robustness of SVM and FNN for circular track prediction when the training and test tracks are taken from different days with different sound speed profiles.