A Low-Cost Edge–Cloud Prototype for Continuous Bird-Sound Detection and Ensemble-Based Deep Learning Species Classification

This paper presents OrniSonix, a low-cost distributed edge–cloud acoustic monitoring system for continuous bird-sound detection and species classification. The system employs autonomous Raspberry Pi Zero 2W edge nodes equipped with acoustic sensors and 4G/LTE connectivity. The system continuously acquires environmental audio in fixed-duration segments, preprocesses it, transforms it into log-Mel representations, and evaluates it locally with a lightweight bird/no-bird inference gate. Only observations considered likely to contain bird activity are transmitted through an OpenVPN private network to the OrniSonix cloud infrastructure, where the recording, spectrogram, feature representation, metadata, and subsequent inference results are maintained as a common MongoDB observation document. The cloud architecture additionally integrates asynchronous deep learning inference, distributed-node monitoring, data visualization, and event-driven notification services. We evaluated six lightweight edge models: Isolation Forest, One-Class SVM, Pseudo-LightGBM, SGDClassifier, TinyCNN, and TinyResNet. TinyCNN achieved the highest Edge Efficiency Score, whereas Isolation Forest was selected for deployment because it requires only positive bird training observations and retained 952 of 1000 independently evaluated real bird samples, corresponding to a 95.20% bird pass rate. We evaluated cloud species classification using a Greece-focused Xeno-canto dataset. After filtering severely underrepresented classes, Dataset II contained 31 bird species and 32,289 spectrogram observations. A probability-level ensemble combining YOLO26-S-Cls, EfficientNet-B0, and PANN CNN14 achieved the best overall performance, with Top-1 accuracy of 85.0%, Top-5 accuracy of 97.5%, Macro F1-score of 0.945, ROC-AUC of 0.919, and a Classification score of 0.904. The proposed approach shows that permissive edge filtering, secure distributed data acquisition, document-oriented cloud persistence, and heterogeneous deep learning fusion can be integrated into a practical architecture for continuous, remotely managed bioacoustic monitoring.