APPLICATION OF MACHINE LEARNING TECHNIQUES TO IDENTIFY FORAGING CALLS OF BALEEN WHALES

Abstract

An unsupervised machine learning algorithm has been applied to passive acoustic monitoring datasets to detect and classify foraging calls of blue whales, Balaenoptera musculus, and fin whales, Balaenoptera physalus. This approach involves using a k-means clustering algorithm to cluster data based on common features, which produces a number of specified centroids. The centroids are then compared to machine-selected candidates for classification. Once divided into initial clusters, further clustering is done to fine-tune results. Preliminary testing of the algorithm yielded promising results. The cross-validation method and the DCLDE 2015 scoring tool were used to estimate out-of-sample performance of the detection algorithm. The automated detector/identifier has been applied to data collected during different seasons, and its performance was analyzed for various types of noise present in data, signal-to-noise ratios, and acoustic environment. The advantages of this approach over traditional manual scanning are increased reliable performance, and time and cost efficiency. This approach could potentially be a faster method of sorting and classifying large acoustic data sets.