Automatic Detection of Acoustic Signals of Beluga Whales and Bottlenose Dolphinsстатья
Информация о цитировании статьи получена из
Scopus
Статья опубликована в журнале из списка Web of Science и/или Scopus
Дата последнего поиска статьи во внешних источниках: 30 января 2024 г.
Аннотация:Passive acoustic monitoring (PAM) is one of the most effective and widely used methods for marine mammal detection and population density estimation. In most cases, it is associated with an inevitably large volume of raw acoustic data, making analysis exclusively by human experts challenging and extremely time-demanding. We present a neural network algorithm for detection of tonal signals produced by toothed cetaceans which allocates time intervals with high likelihood of marine mammal signal presence in underwater sound recordings. The proposed model is based on a convolutional neural network (CNN) of the ResNet152 architecture used as a backbone with additional linear layers. The input data for the model are spectrograms of short segments of underwater sound recordings. Training of the network was performed on a benchmark dataset of bottlenose dolphin whistles labeled by human experts. Analysis of the performance metrics (Precision, Recall, and F1-score) showed the proposed model’s superiority in comparison to widely used acoustic energy-based approaches. In virtue of the model’s versatility, we believe that it can be successfully used for detection of tonal signals of other species characterized by producing tonal acoustic signals. This assumption was confirmed by preliminary results of the application of the developed CNN to a large archive of the White Sea beluga whale recordings.