Detecção e classificação de sinais bioacústicos de cetáceos odontocetos de cetáceos odontocetos na Bacia de Santos

Abstract

Passive Acoustic Monitoring (PAM) has proven to be an effective tool for studying the ecology of marine environments. Biological sounds play a vital role in processes such as survival and reproduction of marine organisms. Among vertebrates, marine mammals are particularly notable for their rich and diverse acoustic repertoires. Advances in technology over recent decades have enabled the deployment of a wide array of equipment and platforms capable of collecting bioacoustic data on broad spatial and temporal scales, supporting the implementation of PAM as a tool for environmental management. This approach offers numerous advantages; however, certain challenges remain, particularly regarding the need for more comprehensive knowledge of species-specific acoustic repertoires and the development of robust classification models. In this study, classifiers were developed for different vocalizations of cetaceans (odontocetes) in the Santos Basin, using a reference dataset collected by the Santos Basin Cetacean Monitoring Project/Petrobras. A set of machine learning-based classifiers were trained and validated using whistle data from seven dolphin species (Odontoceti: Delphinidae), achieving an overall balanced accuracy of 80%. Despite these encouraging results, interspecific variability suggests that behavioral factors may influence acoustic patterns. Using the same methodological framework, a classifier was also developed for echolocation clicks. This classifier exhibited lower performance compared to whistle-based models, suggesting that whistles may encode more species-specific information. Nevertheless, the click classifier remains relevant, as it enables the inclusion of species with limited whistle recordings. It included nine dolphin species in total, among which two were not represented in the whistle dataset and still demonstrated good performance. Combined models integrating results from both whistles and clicks were also tested, achieving a balanced overall accuracy of 76%. This approach demonstrated to be a robust alternative for species identification, even in situations where only clicks or whistles are available, thus broadening the practical applicability of the proposed methodology. A multi-model classification system was also proposed, integrating the weighted probabilities from six base classifiers (whistles, clicks, and combined models) to support the classification of novel acoustic data. Furthermore, automatic click detectors and classifiers were tested for other odontocete species found in the Santos Basin, including members of Physeteridae (sperm whale), Kogiidae (pygmy sperm whale), Ziphiidae (beaked whales), and Pontoporiidae (franciscana). The acoustic features of their clicks were analyzed and compared with existing literature and known geographic distributions to support taxonomic identification at the finest possible resolution. The identification of biological sounds is an important contribution to scientific research, with valuable applications in the monitoring and conservation of marine biodiversity. The acoustic classifier developed in this study represents a significant advancement in this context, offering practical applications for the passive acoustic monitoring of marine mammals.