Efficient wildlife monitoring: Deep learning-based detection and counting of green turtles in coastal areas

• Published in: Ecological informatics Vol. 86; p. 103009
• Main Authors: Noguchi, Naoya, Nishizawa, Hideaki, Shimizu, Taro, Okuyama, Junichi, Kobayashi, Shohei, Tokuda, Kazuyuki, Tanaka, Hideyuki, Kondo, Satomi
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.05.2025; Elsevier
• Subjects: Abundance estimation; algorithms; automatic detection; Chelonia mydas; Coastal area; corals; Deep learning; Sea turtle; seagrasses; Unmanned aerial vehicle; wildlife; YOLOv7; Deep learning; CNN; Unmanned aerial vehicle; Coastal area; YOLOv7; png; Abundance estimation; Sea turtle; mAP; IoU; AP
• ISSN: 1574-9541
• DOI: 10.1016/j.ecoinf.2025.103009

Drones have recently been used to assess wildlife populations and their abundance. The automatic detection of target animals in drone footage enables efficient abundance estimation. However, accurately detecting animals remains challenging, especially in complex field environments. Moreover, automating the tracking of individuals across consecutive images and counting them along transect lines is necessary to apply drones to line-transect surveys. In this study, deep-learning-based You Look Only Once, Version 7 (YOLOv7) models were developed to automatically detect green turtles (Chelonia mydas) in Japanese coastal areas featuring coral reefs and seagrass beds. Drone footage yielded 103,296 annotated images of green turtles. The model was trained and validated using 78 % and 22 % of the images. The best model performances were 0.848, 0.853, and 0.922 for precision, recall, and mean average precision at the threshold of the intersection over union = 0.5, respectively. Then, the BoT-SORT object-tracking algorithm was implemented to track green turtles detected using the YOLOv7 model, and the counting of individuals was automated. When this automatic counting model was tested using eight drone footage clips, green turtles at the sea surface were successfully tracked and counted (n = 3/3); however, the performance in counting underwater green turtles was relatively poor (n = 27/59). The reduced performance might be attributable to accumulated errors in detecting green turtles while processing numerous images in the footage (approximately 60 fps). Nonetheless, relatively high precision was achieved by reducing false positives in complex coastal areas. The methods in this study should enhance the efficiency of long-term wildlife monitoring programs. •YOLOv7 models were developed to automatically detect green turtles in coastal areas.•Green turtles at the sea surface were successfully tracked and counted.•However, the performance in counting underwater green turtles was relatively poor.•High precision was achieved by reducing false positives in complex coastal areas.•The models would contribute to efficient long-term wildlife monitoring programs.