Le paradigme d'intelligence artificielle 'Data-centric' au service de l'écoacoustique

Conservation biology aims at protecting biodiversity from erosion and extinction through effective monitoring. Ecoacoustics, a field combining ecology, acoustics, and computer science, offers a promising approach using passive acoustic recording devices and advanced analysis techniques, including deep learning, to monitor ecosystems and assess biodiversity over large spatio-temporal scales. In this thesis, deep learning was employed with a data-centric approach to investigate the impact of its representational capabilities applied to ecoacoustics, spanning from species to soundscape scales. The first chapter proposed a novel labelling function based on segmentation, featurization, and clustering to create high quality datasets from weakly labeled bird sounds databases, such as Xeno-Canto. The second chapter investigated the impact of signal-to-noise ratio (SNR) on bird sound classification using a convolutional neural network, and subsequently the application of the model to elucidate the vocal activity of the Tengmalm Owl across years. The last chapter introduced a statistical approach to measure the expressiveness of feature vectors of soundscapes obtained with different deep learning algorithms, across sites, years, months and hours.