Wild Cetacea Identification using Image Metadata

Abstract

Identification of individuals in marine species, especially in Cetacea, is a critical task in several biological and ecological endeavours. Most of the times this is performed through human-assisted matching within a set of pictures taken in different campaigns during several years and spread around wide geographical regions. This requires that the scientists perform laborious tasks in searching through archives of images, demanding a significant cognitive burden which may be prone to intra- and interobserver operational errors. On the other hand, additional available information, in particular the metadata associated to every image, is not fully taken advantage of. The present work presents the result of applying machine learning techniques over the metadata of archives of images as an aid in the process of manual identification. The method was tested on a database containing several pictures of 223 different Commerson’s dolphins (Cephalorhynchus commersoni) taken over a span of seven years. A supervised classifier trained with identifications made by the researchers was able to identify correctly above 90% of the individuals on the test set using only the metadata present in the image files. This reduces significantly the number of images to be manually compared, and therefore the time and errors associated with the assisted identification process.