Presenter: Ankita Shukla (Arizona State University)

Description:
Re-identifying animals in the wild using photo-identification (photo-ID) allows to determine key ecological parameters of wildlife populations (e.g., abundance, carrying capacity, community structure, site-fidelity). A photo-ID database of the St. Lawrence Estuary Beluga population (SLEB) taken from a research boat has been growing for over 31 years. Those shots of belugas’ flanks allowed to identify about 350 individuals among the whole population, which counts around 1000 individuals. The traditional re-identification process involves the examination of each photo by an operator to locate any distinctive marks allowing them to either identify a known individual or add a new one to the database. This is a very tedious task, and in the past 31 years, only 21 years of photo-ID have been partially processed, with more photos taken every year than it is possible to process manually. However, those data are of major importance to characterize the social dynamics of the SLEB population. In this context, the objective of the project is to investigate how machine learning can support the automatic re-identification of known individuals from the beluga population and the discovery of new individuals. The automatization of this process will notably enable the reduction of the biases introduced by the human judgement for the recognition of the individuals with distinctive marks.  We have been successful in developing an initial AI based solution that achieves approximately 75% true acceptance rate in re-identifying the individual when the system is allowed a false acceptance rate of 1%.  Going forward we aim to improve the performance of our systems and develop a hybrid solution that can benefit from human experts as well as more effectively utilize  AI tools.