A fine-tuning workflow for CNN-based bat acoustic classification

Convolutional neural networks (CNNs) are widely used in passive acoustic monitoring (PAM) for automated bat call classification, yet models trained on broad reference datasets often fail to generalize across heterogeneous ultrasonic environments. Domain shift driven by variation in noise, species assemblages and recording conditions reduces accuracy and compromises ecological inference. We developed a fine-tuning workflow to adapt a generic full-spectrum CNN bat classifier to new acoustic environments using small volumes of annotated data. Tested across two ecologically contrasting landscapes that represent major sources of error in bat PAM (intense heterogeneous noise versus high activity and species richness), the workflow provides the first systematic assessment of transfer learning under realistic full-spectrum conditions. Fine-tuning with 600–900 annotated recordings increased classification accuracy by up to 31% and improved both recall and precision. Temporal activity patterns derived from fine-tuned predictions closely matched expert annotations (overlap ≥0.88). Our results show that fine-tuning is a practical and data-efficient strategy for improving classifier robustness in full-spectrum ultrasonic bat monitoring. The workflow is computationally efficient, scalable and fully integrated within an open-source platform, providing an accessible tool for long-term and large-scale acoustic monitoring.