Detecção e classificação automática de sinais acústicos de baixa frequência de Balaenoptera borealis (baleia sei) utilizando técnicas de visão computacional

Abstract

Studies on the field of marine mammals have traditionally been conducted by observers aboard of aircraft or vessels, which is highly costly and limited by weather and oceanographic conditions. Passive acoustic monitoring (PAM) is an effective alternative for studying cetaceans, even under rough weather conditions. With the increasing amount of recorded acoustic data in recent years, there is a growing need for algorithms capable of automatically detecting and classifying acoustic signals. However, the lack of large ground truth acoustic data from marine mammals calls is a challenge in the bioacustic field. Among the baleen whales, the Balaenoptera borealis (sei whale) is one of the least studied in the scientific community. In this study, a low frequency signal detection system and a binary classifier for sei whales are proposed. Furthermore, data augmentation techniques such as flipping and distortion were employed to enhance the quality and the size of the acoustic data. Image processing techniques were used to detected acoustic signals, along with machine learning (ML) algorithms such as Random Forest (RF), Multilayer Perceptron (MLP), Support Vecctor Machine (SVM), and Gaussian Naive Bayes for binary classification of sei whale calls. Comparisons among the ML algorithms were carried out to determine the best binary classification method. The results show a probability of detection of 98.7% for the sei whale, and more than 94% of precision and f1-score for RF, MLP, and SVM.