High sea turtle density in a Mediterranean foraging ground revealed by UAV surveys and machine learning

Monitoring at-sea density and distribution of sea turtles is essential for understanding their seasonal dynamics, identifying hotspots, and mitigating anthropogenic threats. Traditional aerial and boat-based surveys are limited by logistical constraints, being costly and time-demanding. Drones have emerged as cost-effective tools for fine-scale aerial surveys, addressing these challenges, but manual video analysis remains labour-intensive. This study presents and validates an integrated workflow combining simultaneous drone surveys with a deep learning object detection model (YOLOv8) to automate turtle detection and density estimation. The Gulf of Manfredonia (south-western Adriatic Sea) — a key foraging habitat for loggerhead turtles Caretta caretta —where fine-scale distribution, density, and seasonal occurrence remain poorly known, served as a case study. Surveys were conducted at 5 coastal sites between 2022 and 2024, using 2 drones flying standardised double-transect paths. Turtle sightings were both manually annotated and automatically detected by YOLOv8. Automated detections were compared to human-only counts, showing similar accuracy (recall: 91.5 %), with a 3.4 % higher error rate but a 90 % reduction in analysis time. Trade-offs between resolution, survey extent, and inference speed are discussed, and recommendations are provided for applying this workflow to other regions and taxa. Turtles were present year-round, with some of the highest densities (0.722–0.727 turtles km–2) recorded in the Mediterranean, suggesting that lower bycatch rates in summer are likely to reflect offshore shifts in fishing effort rather than turtle movements. Integrating drones and deep learning enhances marine megafauna monitoring, offering an efficient, reliable, and scalable tool for conservation.