Improving Greater Caribbean manatee vocalization detection across habitats using neural networks

Abstract

The detection and classification of Greater Caribbean manatee vocalizations ( Trichechus manatus manatus ) present unique challenges due to the complexities of underwater acoustic environments. This study explores the application of neural networks for improving the identification and classification of Greater Caribbean manatee vocalizations, which can provide valuable insights into their behavior and aid in conservation efforts. Utilizing a large dataset of underwater recordings, we trained a known CNN architecture without domain-relevant pretraining to identify and classify Greater Caribbean manatee calls. Our approach combined advanced signal processing techniques such as filtering and normalization with deep learning algorithms to account for the dynamic and noisy conditions of marine environments, employing data augmentation and feature extraction strategies to focus on relevant and informative sound characteristics. The neural network demonstrated promising results, with an overall F1 score of 95.6% on the Wildtracks test dataset, and an F1 score of 64.4% on the Placencia dataset after fine-tuning on less than 10 seconds of vocalizations. This highlights the ability of the model to generalize to novel datasets collected in different regions with vastly different noise profiles. Although there is room for improvement in terms of generalization, these findings represent an advancement in the automated detection and classification of Greater Caribbean manatee vocalizations. This could potentially lead to more effective monitoring of their populations and contribute to the development of improved conservation strategies.