‘Great Migration’ involves far fewer wildebeest than we had thought

East Africa’s iconic “Great Migration,” which has long been estimated to involve nearly 1.3 million wildebeest, may actually involve far fewer animals—around 600,000, according to a groundbreaking AI analysis of satellite imagery.

The Great Migration showcases the perilous journey of wildebeest, zebras, and various antelope species as they transition between feeding and breeding grounds in Kenya and Tanzania while striving to evade formidable predators like lions, crocodiles, and hyenas.

Accurately quantifying the number of animals participating in this migration has historically been challenging. Traditional methods relied on aerial surveys, which can only cover limited areas at a time, necessitating statistical models to project population densities over unsurveyed regions. This approach is susceptible to inaccuracies due to the uneven and fluid distribution of herds.

Recent advancements in satellite technology provide a solution to these challenges. A single satellite image can encompass hundreds of thousands of square kilometers, significantly minimizing the likelihood of double counting and eliminating the need for extrapolative estimates. Manual counting remains impractical in such vast images, but artificial intelligence can be employed to conduct these assessments efficiently. Isla Duporge from the University of Oxford states, “Automation through AI does provide greater consistency and accuracy in counting.”

In the recent study led by Duporge and her team, they trained two deep learning models—U-Net and YOLOv8—to identify wildebeest in images containing 70,417 manually labeled animals. The models were then applied to over 4,000 square kilometers of high-resolution satellite imagery captured on August 6, 2022, and August 28, 2023.

Both AI models yielded similar estimates, counting 324,202 and 337,926 wildebeest in 2022, and 502,917 and 533,137 in 2023. The notable variation between the two years stems from the fact that the surveys were conducted at different points in August—a period known for fluctuations in migratory animal numbers. “What’s encouraging is that both deep learning models, which utilize very different techniques, produced consistent results,” Duporge comments. “This consistency enhances our confidence in the reliability of AI-based counts for wildlife population tracking.”

The previous figure of 1.3 million wildebeest has been utilized since the 1970s. Duporge suggests, “Based on our findings, I suspect the true population size is closer to around 800,000 individuals, assuming we could achieve a count with zero errors. I believe that the aerial count is an overestimate while ours could be a slight underestimate.” She underscores the challenge of counting animals hidden under trees or outside the survey area, but it is startling that no more than 533,137 were observed.

Lower counts do not necessarily imply a population collapse. Instead, it is possible that the wildebeest have altered their migratory paths. However, they do face significant threats like habitat loss and fragmentation due to agricultural practices. Thus, precise population estimations are essential for informing effective conservation strategies.

This research team had previously developed an AI model to identify elephants from satellite data; however, this marks the inaugural use of such a method for conducting a census of individual mammals across a large, distributed population. They are also progressing with a comparable technique aimed at detecting and counting African rhinos.

“I advocate for a shift towards satellite and AI-based methods for wildlife population counting, particularly for species dispersed over extensive landscapes,” Duporge asserts.

The coding framework for the team’s AI model is available on GitHub at https://github.com/sat-wildlife/wildebeest.