Developing a simple automatic elephant detection system aimed at achieving individual recognition in a specific forest site using image processing and supervised machine learning

Elephants (Loxodonta africana and Loxodonta cyclotis) living in forested environments are threatened with extinction due to poaching, habitat destruction and human–elephant conflict. However, because of the difficulties associated with monitoring them in bushy and remote zones, their population size, behaviour, ecology and ranging patterns are still poorly understood. Our study, conducted on African elephants ranging in the Sebitoli area in Kibale National Park in Uganda, aimed to detect and then identify individuals recorded on camera-trap clips using a machine learning-based approach for automatic detection and individual recognition. Our objectives include (1) developing an automatic elephant detector from camera-trap clips, (2) implementing a digital image processing method to recognise individual elephants from photographs in these clips, and (3) comparing the reliability and efficacy of our system with human observations. We show that our detector outperforms human operators in speed and efficacy, detecting all elephants in the video footage. Our supervised classifier, based on elephant heads observed in profile, uses eigenfaces features and the K-nearest neighbours algorithm and, under a stricter leave-one-video-out validation framework, reached an overall image-level classification accuracy of 54.6%, with specimen-level performance reaching up to 74.1% for some individuals. Thus, our protocol represents a proof of concept developed for a specific local setting, which may inspire similar low-cost approaches, requiring limited technical and computational resources, in other elephant populations or species, offering valuable insights for site-based conservation and monitoring strategies.