Time-frequency characterization and classification, temporal-spatial, spectral, and source level distributions of fin whale vocalizations in the Norwegian Sea observed with a coherent hydrophone array

Abstract

The vocalization behavior of fin whales in the Norwegian and Barents Seas is monitored and studied using the Passive Ocean Acoustic Waveguide Remote Sensing (POAWRS) technique. The POAWRS technique is employed to provide detection, bearing-time estimation, time-frequency characterization and classification, as well as localization and geographic positioning of the fin whale vocalizations received instantaneously over wide areas greater than 10,000 km^2. The observations were made from 18 February to 08 March 2014 in several regions of the Norwegian and Barents Seas coinciding with the spawning season and grounds for three commercially and ecologically important fish species: the Atlantic herring (Clupea harengus) off the coast of Alesund, the Atlantic cod (Gadus morhua) off the Lofoten archipelago, and the capelin (Mallotus villosus) off the Northern Finnmark coast. For marine mammals that are top predators, such as the fin whale, the concentrated fish migrations and spawnings are a tremendous source of prey. Since the fin whale is currently on the endangered species list, it is unclear how the recovery of fin whales will impact large oceanic fish stocks in terms of future harvesting potential. It is, therefore, of crucial importance to develop methodologies to observe fin whales over wide areas, such as developing a specific level of array processing to effectively detect specific types of fin whale vocalizations, and gather the information required to understand their behavior, such as their interaction with fish species. In addition, millions of acoustic signals can be received by the POAWRS system per day and are classified by visual inspection and using unsupervised clustering algorithms. This method of classification is performed during post-processing of the data and is a barrier to identifying fin whales vocalizations and other individual sound sources in near real-time. A near real-time detection and classification system is essential for organizations that are required to comply with federal laws and regulations that apply to fin whales, such as the Marine Mammal Protection Act. Hence, an automated classification system is necessary to enable sound sources to be recognized for near real-time fin whale classification applications. There are two main goals of this dissertation. The first is to provide a better understanding of fin whale vocalization behavior in the Norwegian and Barents Seas in terms of sound production, temporal and spatial distributions, and detectability using the POAWRS technique. Here, the objectives to achieve this goal are to (i) provide a time-frequency characterization for different call types observed (20 Hz pulses, 130 Hz upsweeps, 30-100 Hz downsweep chirps, and 18-19 Hz backbeats); (ii) compare their relative abundances in three different coastal regions off Alesund, Lofoten and Northern Finnmark; estimate the (iii) temporal and spatial distributions, (iv) source level distributions, and (v) probability of detection (PoD) regions for the more abundant 20 Hz pulse and 130 Hz upsweep call types. The findings presented here will (1) provide interesting fin whale vocalization behavioral and biological insights; (2) provide time-frequency characteristics that form features that can be used in automatic classifiers for near real-time fin whale detection applications; (3) provide fin whale vocalization rates and location-dependent vocalization rate spatial distribution maps, which can be applied in future studies of predator-prey interactions in the Norwegian and Barents Seas; (4) compare source levels as a means of characterizing different vocalization types; and (5) show that a specific level of array processing is required to effectively detect specific types of fin whale vocalizations in the Norwegian and Barents Seas. The final goal is to develop an automatic classifier for near real-time fin whale detection applications using the time-frequency characterization results from the different fin whale call types observed in the Norwegian and Barents Seas. Here, the objectives to achieve this goal are to (i) gather a large training and test data set of fin whale vocalization detections and other acoustic signals; (ii) build multiple fin whale classifiers, including a logistic regression, support vector machine (SVM), decision tree, convolutional neural network (CNN), and long short-term memory (LSTM) network; (iii) evaluate and compare the performance of each classifier using multiple metrics including accuracy, precision, recall and F1-score; and (iv) integrate one of the classifiers into the existing POAWRS software. The findings presented here will (1) provide an automatic classifier for near real-time fin whale detection applications; and (2) lay the foundation in building an automatic classifier for near real-time detection of various biological, geophysical and man-made sound sources in the ocean environment.