Speaker: Gabriel Bidart
Abstract: Amphibian populations worldwide are declining, particularly in biodiversity hotspots such as the Neotropics, posing urgent conservation challenges. Acoustic monitoring offers a non-invasive tool for tracking amphibian presence and activity, but large-scale audio datasets pose bottlenecks. We present a user-friendly workflow leveraging pre-trained convolutional network embeddings from BirdNET combined with a multilayer perceptron classifier to automate multi-label detection of anuran calls in complex, noisy tropical soundscapes. Using the expert-annotated AnuraSet encompassing 42 species across Brazilian Cerrado and Atlantic Forest biomes, we demonstrate substantial accuracy gains over prior ResNet models, especially for rare species (up to 70% improvement in F1 scores). We apply calibration techniques to translate confidence scores into statistically meaningful probabilities, enabling reliable thresholding and risk assessment. Analysis of acoustic activity for three focal species reveals distinct diel and seasonal call patterns consistent with ecological theory on niche partitioning and phenology, though the limited temporal scope cautions against definitive conclusions. The workflow, openly available at https://github.com/ElBatrasio/anura-bat, facilitates replicable, scalable bioacoustic monitoring and offers a foundation to expand amphibian surveillance over broader spatial and temporal scales. The integration of accessible machine learning and robust ecological datasets promises to enhance conservation capability in tropical ecosystems facing increasing pressures.