Characterization of very low frequency ambient noise by principal component analysis

Abstract

Long-term acoustic monitoring datasets compose a dynamic mixture of many source types, creating a unique noise field in each measurement location and for each recording time. Principal Component Analysis (PCA), and the correlation matrix from which the Principal Components are derived, offer an effective means of estimating the fundamental acoustic contributors, and their portion of the noise field. An improved characterization of ambient noise data sets is discussed, by representing the sound field in a smaller parameter space of principal components, each component reflecting the spectral characteristics of a specific sound source (i.e. fin whale vocalizations). This procedure leads to a definitive parsing of the contributing source types in the measured spectra. The datasets used are recordings from the Comprehensive Nuclear-Test Ban Treaty Organization’s (CTBTO) hydro-acoustic monitoring system.