An interactive machine learning toolkit for classifying impulsive signals in passive acoustic recordings

Abstract

A typical wide-bandwidth passive acoustic seafloor sensor can record tens of millions of impulsive signals produced by biological, anthropogenic, and physical sources each year. Sources include echolocating toothed whales, snapping shrimp, ship propeller cavitation, echosounders, and weather. The volume and variety of detections make manual classification by human analysts unmanageable without in-depth knowledge of the overall acoustic context of each monitoring location. We describe an interactive machine learning toolkit in development for efficiently detecting and classifying short, highly variable impulsive signals in large passive acoustic datasets. Modules include a configurable event detector, an unsupervised clustering module for identifying dominant signal categories, a deep learning unit for learning and applying event classes, and a graphical user interface for viewing, correcting and evaluating detectors and classifiers. These tools are discussed and illustrated in the context of efforts to extract distinctive features of toothed whale clicks from towed array recordings collected by NOAA SEFSC and to use them to classify detections in unlabeled seafloor sensor recordings. The goal of the toolkit is to facilitate classification of individual signals to species in very large PAM datasets across sensor types and monitoring locations, and to improve quantitative assessment of these sources.A typical wide-bandwidth passive acoustic seafloor sensor can record tens of millions of impulsive signals produced by biological, anthropogenic, and physical sources each year. Sources include echolocating toothed whales, snapping shrimp, ship propeller cavitation, echosounders, and weather. The volume and variety of detections make manual classification by human analysts unmanageable without in-depth knowledge of the overall acoustic context of each monitoring location. We describe an interactive machine learning toolkit in development for efficiently detecting and classifying short, highly variable impulsive signals in large passive acoustic datasets. Modules include a configurable event detector, an unsupervised clustering module for identifying dominant signal categories, a deep learning unit for learning and applying event classes, and a graphical user interface for viewing, correcting and evaluating detectors and classifiers. These tools are discussed and illustrated in the context of efforts to ex...