Bear biometrics: developing an individual recognition technique for sloth bears

Abstract

Identifying individual animals, especially in large mammals, is an important goal for wildlife biologists and managers. Bears, occupying diverse habitats, face and experience significant conflict. Among Asian bears, the sloth bear Melursus ursinus (Shaw, 1791; Vulnerable IUCN Red List) is reported vulnerable due to negative interactions with humans, requiring solutions like identifying bear individuals using morphological features. To do so, we used an image-comparison algorithm to evaluate the uniqueness of chest markings using structural similarity index (SSIM) and trained a deep learning model based on the EfficientNet architecture for predicting an individual bear classification. We collected 1567 images (of 144 bears) to examine individual-level differences in chestmark patterns. The comparison yielded 98% accuracy in differentiating chestmarks as a unique pattern for an individual. Subsequently, we trained a circular classification model based on EfficientNet framework using augmented 5628 images for training (80%; of 115 bears), which was validated over 95% for top one and 99% for five individual predictions on 1407 testing images (20%; of 115 bears). The final step involved passing 58 non-augmented images (of 29 out-of-train bears), and the top five predictions of closely similar patterns suggested by the model were then manually compared for similarities in shapes, which suggested whether the image belonged to a new individual. The high accuracy of comparison and classification models suggests the potential applicability of this technique for helping maintain the ex-situ bear database, identifying the conflict individual and estimating bear populations.