Whale Vision: A tool for identifying sperm whales and other cetaceans by their flank or fluke

Abstract

The Mediterranean sperm whale (Physeter macrocephalus) is classified as endangered, and effective conservation strategies require accurate knowledge of individual whales’ movements and behaviours. However, non-invasive techniques to track individuals require photographic identification of single subjects, a time-consuming process that requires manual curation. This study presents machine learning tools developed to automate the identification of individual sperm whales from photographic data. While fluke images are traditionally used for cetacean identification, this research extends deep learning-based identification to include flank images, a novel approach for sperm whales. Two Residual Neural Network models were trained using a contrastive learning process to distinguish individuals from either fluke or flank images, representing each whale as a point in a 128-dimensional latent space for fast re-identification. Evaluation on the Oceanomare Delphis dataset demonstrated identification accuracies of 81.2% for fluke images and, notably, 76.5% for flank images, highlighting the effectiveness and potential of flank-based identification. A user-friendly interface was developed to facilitate practical application, enabling both the identification of known individuals and the incorporation of new subjects. The methods and tools presented are adaptable to other cetacean species, offering a scalable and non-invasive solution to support conservation efforts.