Analysing the variety of bird sounds in natural environments could provide valuable ecological information. However, traditional survey methods rely on manual classification and visual monitoring of birds, which can be challenging especially in large or difficult-to-reach areas. This research explores the use of deep learning models to automatically identify bird species based on their vocalizations. We used two datasets of bird sounds recorded in natural habitats in two different areas in Italy to evaluate the effectiveness of BirdNET pre-trained models, namely BirdNET-Analyzer and BirdNET-Lite. In the first data set (recorded in Toscana), BirdNET-Analyzer identified more bird sounds than BirdNET-Lite, though they both missed some bird species. Excluding these missed predictions and cases in which the algorithms misclassified birds outside our target species, both algorithms demonstrated acceptable performance, as shown by ROC and PR curves. For the second dataset (recorded in the Tovanella park in Veneto), we fine-tuned the BirdNET-Analyzer model using data annotated by human experts, which allowed to significantly improve the detection performance compared to the baseline BirdNet-Analyzer model. Overall, our results suggest that deep learning is a mature technology that can be readily used to support monitoring of bird species in natural environments.
FINE-TUNING DEEP LEARNING MODELS FOR AUTOMATED BIRD RECOGNITION: AN EMPIRICAL EVALUATION OF BIRDNET
TEYMOURI, SAMIRA
2023/2024
Abstract
Analysing the variety of bird sounds in natural environments could provide valuable ecological information. However, traditional survey methods rely on manual classification and visual monitoring of birds, which can be challenging especially in large or difficult-to-reach areas. This research explores the use of deep learning models to automatically identify bird species based on their vocalizations. We used two datasets of bird sounds recorded in natural habitats in two different areas in Italy to evaluate the effectiveness of BirdNET pre-trained models, namely BirdNET-Analyzer and BirdNET-Lite. In the first data set (recorded in Toscana), BirdNET-Analyzer identified more bird sounds than BirdNET-Lite, though they both missed some bird species. Excluding these missed predictions and cases in which the algorithms misclassified birds outside our target species, both algorithms demonstrated acceptable performance, as shown by ROC and PR curves. For the second dataset (recorded in the Tovanella park in Veneto), we fine-tuned the BirdNET-Analyzer model using data annotated by human experts, which allowed to significantly improve the detection performance compared to the baseline BirdNet-Analyzer model. Overall, our results suggest that deep learning is a mature technology that can be readily used to support monitoring of bird species in natural environments.File | Dimensione | Formato | |
---|---|---|---|
Teymouri_Samira.pdf
accesso aperto
Dimensione
16.58 MB
Formato
Adobe PDF
|
16.58 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/71036