Date created: | Last Updated:

Category: Project

Description: Automated detectors such as camera traps allow the efficient collection of large amounts of data for the monitoring of animal populations, but data processing and classification are a major bottleneck. Deep learning algorithms have gained increasing attention in this context, as they have the potential to dramatically decrease the time and effort required to obtain population density estimates. However, the robustness of such an approach has not yet been evaluated across a wide range of species and study areas. This study evaluated the application of DeepFaune, an open-source deep learning algorithm for the classification of European animal species, and Camera Trap Distance Sampling (CTDS) to a year-round dataset containing 895,019 photos recorded in ten protected areas across Germany. For all wild animal species and higher taxonomic groups on which DeepFaune was trained, the algorithm achieved an overall accuracy of 90%. There were no significant differences between the CTDS estimates based on manual and automated image classification for all species and seasons with a minimum sample size of 20 independent observations per study area except for red deer in one area. Meta-regression revealed an average difference between the classification methods of -0.01 (95% CI: -0.26–0.23) animals/km². As expected, classification success correlated with the divergence of the population density estimates, but variability was introduced by the complex effects of false negative and false positive detections on CTDS parameters. Our results demonstrate that readily available deep learning algorithms can enable a largely unsupervised workflow for estimating population densities of large and visually distinct species. With rapid technological developments, the reliability for a wider range of species can be expected to increase in the near future.