Will artificial intelligence revolutionize aerial surveys? A first large-scale semi-automated survey of African wildlife using oblique imagery and deep learning

Large African mammal populations are traditionally estimated using the systematic reconnaissance flights (SRF) with rear-seat observers (RSOs). The oblique-camera-count (OCC) approach, utilizing digital cameras on aircraft sides, proved to provide more reliable population estimates but incurs high manual processing costs. Addressing the urgent need for efficiency, the research explores whether a semi-automated deep learning (SADL) model coupled with OCC improves wildlife population estimates compared to the SRF-RSO method. The study area was the Comoé National Park, in Ivory Coast, spanning 11,488 km2 of savannas and open forests. It was surveyed following both SRF-RSO standards and OCC method. Key species included the elephant, western hartebeest, roan antelope, buffalo, kob, waterbuck and warthog. The deep learning model HerdNet, priorly pre-trained on images from Uganda, was incorporated in the SADL pipeline to process the 190,686 images. It involved three human verification steps to ensure quality of detections and to avoid overestimating counts. The entire pipeline aims to balance efficiency and human effort in wildlife population estimation. RSO and SADL-OCC approaches were compared using the Jolly II analysis and a verification of 200 random RSO observations. Jolly II analysis revealed SADL-OCC estimates significantly higher for small-sized species (kob, warthog) and comparable for other key species. Counting differences were mainly attributed to vegetation obstruction, RSO observations not found in the images, and suspected RSO counting errors. Human effort in the SADL-OCC approach totaled 111 h, representing a significant time savings compared to a fully manual interpretation. Introducing the SADL approach for aerial surveys in Comoé National Park enabled us to address the OCC's time-intensive image interpretation. Achieving a significant reduction in human workload, our method provided population estimates comparable to or better than SRF-RSO counts. Vegetation obstruction was a key factor explaining differences, highlighting the OCC method's limitation in vegetated areas. Method comparisons emphasized SADL-OCC's advantages in spotting isolated, small and static animals, reducing count variance between sample units. Despite limitations, the SADL-OCC approach offers transformative potential, suggesting a shift towards DL-assisted aerial surveys for increased efficiency and affordability, especially using microlight aircraft and drones in future wildlife monitoring initiatives.