Hydroacoustic signal classification using support vector machines

Abstract. Kernel-based algorithms such as support vector machines (SVMs) are state-of-the-art in machine learning for pattern recognition. This chapter introduces SVMs and describes a specific application to hydroacoustic signal classification. Long-range, passive-acoustic monitor-ing in the oceans is facilitated by propagation properties for underwater sound. In particular, the deep sound (SOFAR, Sound Fixing and Rang-ing) channel can act as a waveguide for underwater signals. In this chap-ter, SVMs are employed for classifying hydroacoustic signals recorded by the sensor network for verification of the Comprehensive Nuclear-Test-Ban Treaty. Constraints in the early signal processing chain and limited data require tailored kernel functions and careful SVM model selection. We demonstrate how problem-specific kernel functions can in-crease classifier performance when combined with efficient gradient-based approaches for optimizing kernel and SVM regularization parameters.