Zwitscherkasten - Bird Classification through sound and image in Germany

The Zwitscherkasten models are intended for non-invasive, on-device bird species monitoring in Central European (primarily German) ecosystems. They support continuous deployment on resource-constrained edge hardware such as Raspberry Pi, Rubik Pi, and mobile devices, targeting citizen science and biodiversity monitoring applications. Audio classification covers 256 common German bird species. MobileNetV3-Small enables low-power inference with 85.6% top-1 accuracy, while EfficientNet-B0 and EfficientNet-B3 achieve 91.7% and 92.9%, respectively. The highest accuracy is obtained with a PaSST spectrogram transformer (94.4% top-1 accuracy). An upstream bird activity detector reduces unnecessary inference and energy consumption. Visual classification is performed using either a multiclass object detection model (YOLO26, up to 0.82 mAP@0.5) or a two-stage detection–classification pipeline. The latter achieves up to 82.1% top-1 accuracy with EfficientNet backbones, while MobileNetV3-Small reaches 76.1% for lightweight deployment. The models are intended for research and monitoring purposes.