Advanced methods for passive acoustic detection, classification, and localization of marine mammals

Abstract

For effective long-term passive acoustic monitoring of large data sets, automated algorithms provide the ability to detect, classify, and locate (DCL) marine mammal vocalizations. Several DCL algorithms were developed and enhanced, with emphasis on methods robust to non-Gaussian and non-stationary noise sources. (1) A subspace model was developed to separate odontocete clicks from noise sounds. (2) A multi-class support vector machine was improved by resolving confusion among species’ overlapping-frequency clicks, and a beaked whale buzz class developed. (3) For dolphin whistles, shape-related features, extractable automatically, were shown to carry species-specific information. (4) Equipment and site differences were discovered to affect Gaussian mixture model classifiers, and methods were developed to mitigate these differences. (5) A nearest-neighbor approach to detection association and 3D localization across multiple phones with multiple arrivals was developed (and applied to beaked whales) using time-difference-of-arrival (TDOA) hyperbolic methods, retaining TDOAs with fewer than the usual three detections and using associations between a given phone’s detections with nearest neighbors. (6) Minke whale “boing” frequency estimates were improved to differentiate individuals, and a kinematic tracking algorithm was developed. (7) A generalized-power-law detector for humpback whales was improved. (8) A software interface for detection was developed, then tested by sending data from Ishmael to a detection process in MATLAB.