Non-invasive extraction of white shark swimming kinematics from unoccupied aircraft system (UAS) imagery

Context Consumer-grade unoccupied aircraft systems (UAS) are increasingly being used by both scientists and hobbyists in the coastal environment. Marine megafauna are observed via UAS as part of monitoring programs, recreational interests, and scientific research, amassing aerial imagery datasets. Because manual documentation of these datasets is infeasible at scale, efficient approaches leveraging computer vision and deep learning have emerged to detect and classify marine megafauna. Aims This study provides a workflow to quantitatively estimate swimming kinematics tailbeat frequency (TBF) and tailbeat amplitude (TBA) of white sharks (Carcharodon carcharias) from aerial UAS video data. Methods Body pose estimation was performed using computer vision model DeepLabCut to track six key white shark body parts across UAS videos. The relative positions of these body part coordinates were used to compute tail position over time and quantify TBF and TBA across a population of white sharks in Monterey Bay, California. Key results With a training set of just 52 images, the deep residual neural network reaches human-level labeling accuracy of body parts (root mean square error of <1.3 cm). This workflow is applied to 76 focal follows representing 34 individuals to produce TBF (0.43 ± 0.07 Hz) and TBA (0.24 ± 0.10 BL) values similar to those derived from biologging devices previously deployed on individuals in this population. Conclusions The results indicated that body pose estimation via DeepLabCut can allow for the rapid extraction of quantitative kinematics such as TBF and TBA in juvenile white shark populations that aggregate in coastal habitats. Implications This approach provides a non-invasive, scalable method to understanding megafauna kinematics in sensitive species that overcomes the logistical barriers of traditional biologging approaches.