Automated waterbird detection, classification, and mapping from UAV imagery using a two-stage deep learning pipeline

Abstract

Waterbirds function as sensitive indicators of ecological change in freshwater ecosystems. Conventional monitoring is labour-intensive, spatially limited, prone to observer bias and disturbance. Unmanned Aerial Vehicle (UAV) imagery combined with deep learning (DL) offers a promising alternative for efficient and scalable monitoring.

This study evaluates a fully automated UAV-based workflow to detect, classify, and map waterbirds along a 12 km canal in Belgium. A two-stage pipeline combining YOLO-based detection and convolutional neural network (CNN) classification was developed and evaluated across varying flight altitudes and ground sampling distances (GSD). Model performance was further assessed using an independent seasonal dataset and analysed in relation to object size and habitat complexity. The detection model achieved a high precision of 0.91 and a recall of 0.81, while the classifier reached a weighted F1-score of 0.90. Detection performance was strongly dependent on ground sampling distance (GSD), with optimal results at ≤6.83 mm px⁻¹ (≤25 m altitude) and marked declines in detection performance at higher GSD values. Additional analyses showed that smaller birds were more likely to be missed, while high-entropy environments such as reed beds may further reduce detection reliability. Comparisons between automated and manual counts revealed discrepancies, influenced by environmental conditions and survey timing. Georeferenced detections projected onto orthomosaics enabled spatially explicit mapping of waterbird distribution.

Overall, UAV-based deep learning provides a promising and scalable approach for waterbird monitoring, while highlighting the importance of spatial resolution and habitat complexity in determining detection performance.