AI-driven edge-device bioacoustics classification system for avian calls

Traditional avian monitoring methods require significant effort and are not scalable for continuous observation. This presents the need for automated solutions. Avian calls vary in frequency, duration, and context, making classification challenging. This study explores the use of artificial intelligence (AI) to classify bioacoustic avian calls through a real-time inference system deployed on an edge device. Among the three models evaluated, ResNet50 performed best with an F1-score of 94.7%, accuracy of 97.2%, and MCC of 96.9%, outperforming MobileNetV2 and EfficientNet. Inference averaged 1200 ms per prediction. Future work will focus on expanding species coverage.