Body-part-based individual feral cat identification from camera trap images using deep learning

Feral cats (Felis catus) are a significant threat to Australia's native wildlife, contributing to the decline and extinction of at least 20 native mammal species through predation impacts. To improve the identification and monitoring of populations, individual identification of cats is required. This study proposes a body-part-based computer algorithmic approach that uses deep learning for individual identification from photos that can address a common challenge associated with using camera trapping, where often only a partial or obscured view of the objects of interest is presented. We investigated the discriminatory attributes of the images of four body parts of the cats: flank (‘body’), back leg, front leg, and tail. We use a subset of a dataset of feral cats collected using camera traps deployed across the Glenelg and Otway regions of Victoria, Australia. Due to the skewed and imbalanced nature of images per individual in the dataset, we used a curated subset of 10 individuals, each with a relatively similar number of images, resulting in a total of 1644 images. We trained deep-learning models with a ResNet-50 backbone on these body parts indivdually as well as combinations of multiple body parts through feature concatenation. Results demonstrate that the body was the most discriminatory part for cat identification, with the back leg the next best part. Other parts added to the performance when they were combined. We conclude that individual cats can successfully be identified using partial body images captured using camera traps. While the body was the most distinctive part, the proposed method provides flexibility in cases where the body is obscured. This study shows that deep learning methods can meaningfully contribute to camera trap image analysis, and hence environmental conservation outcomes.