[PhD Defence] Gabriel Dubus скачать в хорошем качестве

[PhD Defence] Gabriel Dubus

From crowd-sourcing to supervised deep learning: addressing new challenges in the Detection and Classification of Cetaceans using Passive Acoustic Monitoring Jury : M. Jérôme SUEUR, Rapporteur, Muséum National d'Histoire Naturelle M. Paul WHITE, Rapporteur, University of Southampton Mme. Charlotte CURE, Examinatrice, University Gustave Eiffel M. Nicolas LECOMTE, Examinateur, Moncton University M. Gaëtan RICHARD, Examinateur, SOMME M. Olivier ADAM, Directeur, Sorbonne Université M. Dorian CAZAU, Co-Directeur, ENSTA Bretagne Abstract : Cetaceans have become a significant focus in marine science, not only because of their sophisticated and evolving social structures, but also thanks to their strong contribution to the health of ocean ecosystems. Despite their importance, whales are globally endangered. As cetaceans rely heavily on acoustic signals for communication, navigation, or hunting, passive acoustic monitoring (PAM) emerges as a valuable approach, enabling the simultaneous capture of species-specific and contextual information. The vast amounts of recorded data generated by such monitoring efforts drive the need for automating key processing tasks such as the detection and classification of cetacean sounds. While deep learning algorithms have shown significant potential in this area, their effectiveness is often hindered by the limited availability of annotated datasets, which are crucial for training these models. This data scarcity, combined with the complexity of manual annotation and the large heterogeneity in the quality of datasets, poses a significant challenge which currently limits the performance and generalization capabilities of deep learning approaches in real-world underwater acoustic applications. The work of this PhD thesis is in two main folds: the manual annotation process and the development of automatic detection and classification methods. In a first part, the inter-annotator variability is investigated through multi-annotator annotation campaigns performed on different marine bioacoustics datasets, with a focus on the annotators experiences. This work first suggests that crowdsourced annotations from novice annotators can be a viable alternative to expert annotations, and deeply increase the quantity of annotated data for the training process of detection and classification models. This work also suggests that multi-annotators annotation processes enable the training of detection models using non-expert annotations that achieve performance of models trained with expert annotations. All annotation campaigns realized for this work have been done using APLOSE, a web-based annotation platform developed by the OSMOSE Team for large-scale collaborative annotation campaigns of PAM data. This work underscores the potential of multi-annotators annotation to advance the field of cetacean vocalization detection. By harnessing the collective efforts of novice annotators and optimizing annotation strategies, researchers can increase the quantity of annotated data and, thus, the capacity of generalization of deep learning models for detection. The second part of this thesis investigates the potential interest of recent deep learning methods for underwater PAM applied to low-frequency vocalizations of Antarctic blue whales and fin whales were assessed. The use of multi-label supervised contrastive learning is compared with the most commonly used methods in the community, achieving new state-of-the-art performance. Moreover, the use of temporal context is assessed to enhance detection performance using recurrent models as classifiers on the embedded space. Finally, the use of trained encoders based on large PAM datasets to detect rare species on skewed and limited datasets, instead of classical reference dataset, is assessed. A key point of this chapter is that the main results are based on one of the largest public labeled acoustic dataset and fairly compared with the state-of-the-art thanks to the first benchmark recently published along with this dataset. The results contribute to the current major challenge of automatic classification of underwater acoustic events, with the aim of contributing to ecosystem conservation measures for sustainable oceans.