Speaker
Description
Automatic detection and classification of bat echolocation calls is essential for large-scale biodiversity monitoring, yet remains challenging under real-world acoustic conditions and constrained hardware environments. This work presents a lightweight and deployable pipeline for bat call analysis that combines simple signal processing with efficient time-series machine learning. An adaptive amplitude-based segmentation using the Otsu algorithm is applied to extract dominant acoustic events from continuous recordings. These candidate events, primarily representing bat echolocation chirps, are transformed into feature representations using MiniROCKET. The resulting feature vectors are classified by a Multi-Layered-Perceptron trained on a reference dataset of bat taxa common in Germany. To improve robustness, an auxiliary anomaly detection stage is introduced. An Isolation Forest model is trained on the same feature space to identify atypical events. During inference, detected segments are first evaluated for similarity to known bat calls before being passed to the classifier. This cascaded strategy reduces false detections in acoustically challenging environments.The proposed approach provides a practical balance between performance, robustness, and deployability for scalable acoustic wildlife monitoring.