Automated species identification for camera trapping in the Iona Skeleton Coast Trans-Frontier Conservation Area

Abstract

The Iona Skeleton Coast Trans-Frontier Conservation Area (TFCA), straddling the border between Angola and Namibia, has suffered through decades of civil war and poaching. While this history has been detrimental to the community of large mammals in the TFCA, data collected on the mammal populations are insufficient to enable effective management. Survey methods such as aerial counts and community-based monitoring have various shortcomings. Therefore camera trapping, which has become important in surveying wildlife worldwide, could become an essential monitoring tool also for the TFCA. However, camera traps tend to capture large numbers of images over short periods of time. The cost and time involved in the manual analysis of such voluminous datasets are the major limiting factors in camera trapping. Deep learning-based computer vision methods proposed to date to address this problem were found unsuitable for application to camera trapping in the TFCA, being computationally too expensive, requiring specialised hardware and large training datasets, focusing on only one species per photograph or relying on static backgrounds between sequential images. On the other hand, the method developed in this study requires only an entry-level computer and relatively few training data while handling multi-species photos with changing backgrounds. It is able to detect and distinguish between humans, vehicles and four large mammal species of importance in the TFCA, namely giraffe, impala, oryx and zebra. Trained on images sourced from the web and applied to 4 000 camera trap photos, the system yielded a recall rate of 85.7% in detecting human-related object classes and 59.1% in detecting the presence of animals in camera trap photos. Its precision in detecting animals was 100% while its precision in distinguishing between the four large mammal species was 96.8%. Furthermore, frequency distributions of photographs inferred by computer roughly correlated to published diel activity levels for each of the four mammal species investigated. The method did not prove useful for the monitoring of rare species, however. Based on the results, the method could be used to filter for photos containing human-related objects as well as animals, and to label or pre-label photos by species. This may make it xiii LIST OF TABLES xiv useful to monitor anthropogenic disturbance, aid in compiling species inventories, document animal migration, map species distributions and pick out images of species for which population densities are to be estimated. Further work would be needed to test the reliability of computer vision inference as an index of activity levels as well as to develop the ability to monitor rare species successfully. Conceptual and technical aspects of using camera traps in combination with the proposed computer vision method are discussed for application in the Iona Skeleton Coast TFCA. However, the utility of the method has the potential to extend far beyond the TFCA and could be applied to a wide range of conservation projects.