Advancing Whale Research: Using Deep Learning for Efficient Identification

_{Photo: Aqqa Rosing-Asvid}

Our world is changing, and the tools we use to research it must change too. In cetacean (whale) research, photo identification is a crucial technique for tracking individuals and studying population size and dynamics. It provides insights into their lives, helping us understand their whereabouts and protect them more effectively. However, the task of manually matching photographs can be quite demanding, especially with extensive image databases.

Traditional methods involve capturing images of whales and comparing them based on unique markings, which can be time-consuming and labor-intensive. In our research project FINWAIID, we explore an automated approach to streamline this process by developing a deep learning pipeline that can identify individual Southern Hemisphere fin whales from vertical aerial drone footage. Our method focuses on the unique pigmentation patterns found on the backs of these whales, specifically the Central Chevron Pattern and Blaze.

By employing a semi-supervised workflow, we aim to train our system with a smaller set of labeled data, reducing the need for extensive manual labeling. In the next step, we will use an auto-labeling loop with human validation to explore the standardization and automation. This includes the detection of a whale within videos and extracting multiple image crops of the same whale from videos potentially containing many different fin whales. To ensure accurate individual identification, we utilize a deep convolutional neural network (CNN) architecture, similar to those used in human facial recognition technology. Our approach is promising as an alternative identification procedure. We believe that it could significantly improve the efficiency of studying whale populations. This allows for better tracking of individual movements, more accurate estimation of population sizes, and more effective monitoring of the Southern Hemisphere fin whale populations. Additionally, our method could be extended to other datasets containing animals with distinct markings, ranging from seals to terrestrial animals.

Here you can find a video introduction for the project: