Wild Boar Detection on Sea Turtle Nesting Beaches Using Camera Trap Images: Automated Image Recognition with YOLO Models and Their Evaluation

Due to its diverse diet and rooting behavior, the non-native wild boar (Sus scrofa) poses significant conservation challenges worldwide. On Tokashiki Island, Okinawa, Japan, introduced wild boars cause various environmental damages, including predation on rare species and increased red soil runoff. In recent years, concerns have grown over their predation on sea turtle eggs, a key tourism resource. To mitigate this threat, effective and efficient capture methods are essential. Camera trap monitoring provides crucial insights into wild boar predation behavior, helping improve control efforts. However, manually identifying and analyzing wild boars in the vast number of images and videos - most of which are false trigger events where no target wildlife species are captured - is highly labor-intensive and time-consuming. Recently, deep learning-based object detection techniques have gained attention as promising tools for wildlife monitoring. This study evaluates the performance of YOLO-based “one-stage object detection” models (GELAN, YOLOv9, and YOLOv10) using image datasets from motion-sensor camera traps set up on sea turtle nesting beaches in Tokashiki Island. The survey recorded a total of 226.6 hours of video, of which 95% (214.4 hours) consisted of empty background footage (without animals). The videos also captured goats (6.5 hours) in addition to wild boars (2.1 hours). Among the seven models tested, GELAN-C showed the highest overall performance (Precision: 0.96, Recall: 0.89, [email protected]: 0.93). For wild boar videos (74 clips), the model correctly identified 92% (68 clips). For empty background footage (without animals) (29 clips), it correctly identified 72% (21 clips). With this empty background detection accuracy, approximately 70% of the empty footage can be pre-filtered, reducing the required video review time by about 155 hours.