Ground-truth validation of behavioral classifications from accelerometer data using an integrated GPS-accelerometer-camera device: application to an elusive scavenger

Animal-borne cameras integrated with multi-sensor devices offer a powerful tool for hypothesis-driven research. Video recordings enable ground-truth validation of behavioral modes inferred from triaxial accelerometer data in wild animals. However, retrieving video data can be challenging due to battery limitations and the lack of remote download stations. We tested a GPS-accelerometer device integrated with a customized camera and a harness drop-off system on two adult King Vultures ( Sarcoramphus papa ), an elusive Neotropical scavenger. Video recordings from animal-borne cameras were used to establish behavioral reference annotations, which were temporally aligned with accelerometer data and compared against the output of a Self-Organizing Map algorithm to validate behavioral classification performance. Model performance was evaluated using precision, recall, and F1-score metrics. Prototypes were successfully deployed, yielding a total of 764 video files and 2 h 16 min of footage for ground-truth validation. Five to six behavioral modes were identified per individual and matched with accelerometer data. Resting was the most common behavior (59.2–69.9% of time), followed by flight modes such as soaring-gliding (~ 19%) and flapping (~ 3%), while preening accounted for 7.6–17.3% of activity and feeding was infrequent (1.36%, observed in one individual only). Model performance varied across behaviors and individuals. Resting showed high classification performance (F1-score: 0.72–0.89), as did soaring-gliding (F1-score: 0.73–0.94). Flapping achieved moderate performance (F1-score: 0.68–0.76), while preening showed lower classification performance (F1-score: 0.37–0.38). Feeding was poorly classified (F1-score: 0.26), reflecting higher behavioral variability and class imbalance. Animal-borne video provides behavioral reference data for evaluating accelerometer-based behavioral classification models. We evaluated the performance of our prototype, demonstrating its utility for recording behavior in elusive species and offering a solution where direct validation of acceleration data is difficult. We recognize several aspects that should be improved in future designs, including prototype weight, shape, and sensor synchronization. Once these limitations are overcome, integrated GPS-accelerometer-camera systems, combined with reliable drop-off mechanisms, could provide a valuable tool for behavioral validation under realistic field conditions, with applications for both ecological research and conservation.