Date of Award

5-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Natural Sciences

First Advisor

Rycyk, Athena

Area of Concentration

Marine Mammal Science

Abstract

Automated detection of manatee vocalizations across a wide range of habitats, from coastal and estuarine systems to rivers and freshwater springs, is challenged by domain shift, the reduction in model performance when applied to acoustic conditions not represented in training data. We compiled a training dataset of 61,924 annotated clips (19,461 manatee vocalizations, 42,463 background noise) from seven locations spanning West Africa and Florida, including both African (Trichechus senegalensis) and American (Trichechus manatus) manatee vocalizations, and evaluated four Convolutional Neural Network architectures for binary vocalization detection. Leave-location-out cross-validation revealed substantial variation in cross-location generalization, with on-animal tag recordings producing the largest performance decrease (F1 = 0.98 to 0.07). Acoustic augmentation and domain adversarial training did not improve performance, while location-specific noise injection produced substantial gains (DTAG: F1 = 0.067 to 0.626; Sarasota Bay: 0.769 to 0.919). A 3-model ensemble (BirdNET, MobileNet, and PANNs) combining noise injected models achieved the strongest performance (DTAG: F1 = 0.829; Sarasota Bay: F1 = 0.937). These results suggest that when a robust multi-location vocalization dataset is available, domain shift may be driven primarily by background acoustic environment mismatch, and that location-specific noise injection combined with ensemble modeling provides an effective and practical strategy for adapting detectors to novel acoustic environments.

Rights

The author has granted New College of Florida the nonexclusive right to archive, make accessible, and distribute for educational purposes this work in whole or in part in all forms of media, now or hereafter known. The copyright of this work remains with the author.

Share

COinS