AVISNet: An edge-computing model for acoustic identification of species from the Doñana bird community

The global decline in biodiversity calls for scalable and efficient species monitoring. Birds, as sensitive ecological indicators with rich vocal activity, are particularly well suited for passive acoustic monitoring (PAM), which enables extensive, non-invasive data collection. Yet the large volumes of audio generated by PAM require automated and efficient processing pipelines to extract meaningful biological information in real time. To address this challenge, we integrated automated acoustic species identification directly into embedded devices (i.e., at the edge ), enabling autonomous on-site data processing without reliance on cloud-based computation. Our methodology emphasizes model efficiency and leverages lightweight pre-trained architectures (e.g., MobileNet), alongside advanced learning strategies such as knowledge distillation and few-shot learning to overcome data scarcity for under-represented species. This methodology was applied as a case study to acoustic identification of bird species native to Doñana National Park, Spain. We assessed the performance and efficiency of our models at three levels of acoustic data: focal segments, focal recordings, and soundscape recordings. On focal recordings, our models achieve up to 0.70 F1-score in species identification task. At the soundscape level, we evaluated performance using three real-world soundscape datasets, measuring accuracy metrics at both the vocalization and dataset levels. Our lightweight convolutional neural networks achieve high accuracy at the dataset level (F1-score: 0.56–0.72), solid accuracy at the vocalization level (F1-score: 0.29–0.39; mAP: 0.41-0.69), and fast and energy-efficient inference on resource-limited hardware (2.2–4.7