Efficient and Accurate Spatio‐Temporal Population Monitoring of Mobile Animals Using Drones and Computer Vision

ABSTRACT Estimating population sizes is critical for fish and wildlife management but often time‐intensive and sometimes unreliable. Aerial imagery coupled with automated computer vision tools has emerged as an efficient way to obtain robust abundance estimates in certain contexts. However, to accurately estimate abundance from survey imagery without double counting, current methods typically rely on human observers or image stitching that requires motionless animals, thus limiting scalability and the range of suitable survey conditions. Here, using simulations and a field case study, we demonstrate two computer vision methods for surveying animal populations from aerial imagery that, when animals are correctly detected, generate unbiased counts without human intervention, even when animals move during surveys. One approach partitions sections of the raw survey imagery and associated detections into a continuous non‐overlapping mosaic of the survey area using calculated 3D landscape information. The second approach combines all counts from all images and weights detections inversely by the number of overlapping images at each survey location. In our case study of spawning salmon in Alaska, both automated approaches yield counts with mean absolute percentage error (MAPE) of < 12%, with the best achieving 5.5% MAPE. This is similar to on‐the‐ground counts made by human observers as part of a long‐term field monitoring effort (MAPE: 9.4%) and far outperforms existing orthomosaic counting methods (MAPE: 44.4%). Our method also generates fine‐scale locations of each individual, providing novel ecological descriptions of the system. Our approach to surveying highly mobile animals is directly applicable to aerial surveys across many populations and species. Its robustness to animal movement and the ability to easily incorporate detections from artifact‐free raw survey imagery versus orthomosaics will allow researchers and resource managers to monitor a wider range of animal populations while collecting more detailed ecological information at higher frequency than current methods allow.