ABSTRACT Detection and analysis of body scars using photo‐identification data of cetaceans can provide insights into life history, behavior, and exposure to threats, but such analyses typically require manual examination of large image datasets. To explore whether machine learning can assist this process, we developed a pipeline for automatically detecting and counting conspecific scarring events on killer whales. The approach combines an nnU‐Net segmentation model, which generates binary masks containing individual scar lines, with hand‐crafted feature representations and HDBSCAN clustering to group scars into events and estimate the number of events per image. Across a dataset in which images contained between 1 and 10 annotated scarring events (median = 2), the best‐performing configuration produced a conservative estimator with a mean absolute error (MAE) of 0.82 events per image and exactly recovered the annotated number of events in just over 50% of cases. These results demonstrate that machine learning‐assisted analysis can extract biologically meaningful information on scarring burden from standard photo‐identification images of killer whales. While accurate recovery of absolute scarring burden from single images is achievable with useful reliability, detecting small inter‐annual changes remains more challenging under heterogeneous field conditions. Rather than replacing expert annotation, the present system is best viewed as a semiautomated framework that can extend existing photo‐identification workflows toward longitudinal scarring analysis and broader population‐level assessment. It therefore provides a practical foundation on which more comprehensive detectors of cetacean scarring events can be built.
Barnhill et al. (Mon,) studied this question.