African continent created by combining images from all months of 2005. Data from MODIS instrument aboard NASA’s Terra satellite. Image Courtesy of NASA’s Earth Observatory.
An article on a newly published research paper caught our attention this week, which used satellite data to estimate that wildebeest numbers in the Serengeti may be less than half what was previously thought.
The paper ‘AI-based satellite survey offers independent assessment of migratory wildebeest numbers in the Serengeti Open Access’ by Duporge & Wu et al was published in Volume 4, Issue 9 of PNAS Nexus in September 2025.
This study was led by researchers from the University of Oxford working with international partners, and they applied Artificial Intelligence (AI) models to satellite images to count the number of wildebeest taking part in the Great Wildebeest Migration across the Serengeti-Mara. This spectacular migration is a vital food source for many of the predators such as lions, crocodiles, and spotted hyaena.
Previously the count has been undertaken by aerial surveys and has estimated the number of wildebeest to be around 1.3 million, and this figure has remained the same for the last half a century. However, these surveys can only cover a small area within the Serengeti and the numbers counted extrapolated to the total. In addition, the aerial surveys themselves disturb the animals and there is a risk of double counting.
The new approach involved applying an AI models to a series of Worldview-2 & 3 high-resolution satellite images covering 4,000 km² of the northern Serengeti National Park in Tanzania and southwestern Kenya, encompassing the Maasai Mara National Reserve. The images were acquired from August 2022 and 2023.
However, the number of wildebeest mean it is unfeasible to count the animals on the images, and instead the team trained two deep-learning models (U-Net and YOLOv8) to identify wildebeest using a dataset of 70,417 manually labelled wildebeest. Both models performed strongly, achieving a high level of accuracy – an F1 score of 0.83, where anything greater than 0.8 is considered excellent for wildlife applications.
The results of the AI models ranged from 502,917 – 533,137 wildebeest for 2023. Surprisingly, this is less than half the previous estimate of 1.3 million animals. Even more concerning is that the researchers suggested the results themselves may be a slight overestimate as the spatial resolution of 30 to 60 cm per pixel may mean the model has not fully been able to distinguish wildebeest from other similar sized animals such as zebras or eland, although the team also acknowledged that some wildebeest may have been obscured by tree cover.
These surprising results on their own don’t necessarily mean the number of wildebeest have drastically reduced, as they may have changed their migration paths or numbers may have been impacted by other factors. However, it is accepted that the changes in habitat within the Serengeti has impacted the available land for prime grazing and this could be impacting numbers. Therefore, getting accurate numbers to understand what is happening to the wildebeest is critical and the use of satellite data and AI models could have a key role to play in the future.
The team have also released the code they used for detecting wildebeest, and this can be found in GitHub repository. The researchers are currently working on a similar approach for African rhinos, and believe it could be used to count other herd mammals. It is exciting to see satellite data being used in innovative and new approaches to challenges such as counting wildebeest.
