Automatic aerial identification model for baleen whales using drones

The use of unmanned aerial vehicles (UAVs) has significantly improved cetacean research by enabling the collection of high-resolution imagery used in several research topics, such as individual identification. This study presents an automated deep learning framework for identifying mainly fin whales (Balaenoptera physalus), but also sei whales (Balaenoptera borealis) and Bryde’s whales (Balaenoptera edeni) from drone imagery. These species exhibit a characteristic dorsal pigmentation right after the blowhole, known as chevron, which is unique to each individual and remains stable throughout its lifetime, making it a reliable natural marker for photo-identification. UAV-based surveys also improve identification efficiency, increasing the success identification rate from 60-65% using conventional methods to 75-80%, while providing a standardized identification approach that facilitates comparisons and data sharing among different research groups. The proposed framework consists of three sequential stages: head localization, species classification and individual identification based on the chevron pigmentation pattern. The first two stages employ YOLO models trained to detect the head and classify the species. For individual recognition, a hybrid architecture combining a ResNet backbone with a recurrent component, and a Siamese neural network was developed to identify and compare pigmentation patterns with a reference catalogue. The current model has been trained primarily using fin whale imagery and includes a database of over 400 individuals. Although the species-recognition step has been already incorporated into the model, ongoing work aims to improve the Sei and Bryde’s identification through expanded datasets and model refinement. Furthermore, the current research is focused on implementing real-time onboard identification during UAV surveys, thereby increasing the efficiency of field research. Ultimately, the proposed model is intended to support long-term global monitoring of baleen whale populations by providing a standardized identification framework that facilitates data sharing and collaboration among organizations working on baleen whale research and conservation.