Localization of circularly towed sources using machine learning methods

Abstract

Localization of a source towed in a circular geometry is accomplished using machine learning methods, wherein the relationship between received pressure and source range is learned directly from the acoustic data. Feed-forward neural networks (FNN), sup- port vector machines (SVM), and random forests (RF) have accurately estimated the range of a source towed in a linear geometry in shallow water (arXiv: 1701.08431v3). Here, we apply FNN, SVM, and RF to a circular source tow using data from the SCEx17 experiment to examine the effect of varying source-receiver geometries. The Sample Covariance Matrix (SCM) is constructed, vectorized and used as the machine learning input. For FNN and RF, these input vectors are combined to form a d x N matrix and an additional preprocessing step is applied to improve classification results. The input matrix is projected onto a d x k basis formed from the top k eigenvectors of its scatter matrix. This compact representation results in improved computation time and performance for FNN and RF compared with using the SCM inputs.