“Major Migration” Necessitates Far Fewer Wild Taxes Than Expected.

The “great migration” in East Africa is often estimated to consist of around 1.3 million wildebeest. However, a recent AI analysis of satellite images reveals that fewer than 600,000 animals make this yearly journey across the Serengeti Mara landscape.

This significant migration includes wild zebras and antelopes, as they traverse between feeding and breeding areas in both Kenya and Tanzania, while also evading predators such as lions, crocodiles, and hyenas.

Determining the number of migrating animals is a challenging process, traditionally accomplished through aerial surveys with crew members. These surveys typically cover limited areas, necessitating the use of statistical models to estimate animal density across larger regions.

In contrast, satellite surveys offer a solution to these challenges since a single image can encompass extensive areas, minimizing the chances of double-counting and eliminating the need for metabolic calculations. While manually counting wildebeests over such vast expanses is impractical, AI can aggregate the data effectively. “AI automation enhances count consistency and accuracy,” says Isla Duporge from Oxford University.

In a new study, Duporge and her team developed two deep learning models (U-Net and Yolov8) to identify wildebeest using a dataset of 70,417 manually labeled images. These models were then applied to high-resolution satellite images spanning over 4000 square kilometers, with capture dates of August 6, 2022, and August 28, 2023.

The two AI models returned comparable results: counting 324,202 and 337,926 wildebeests in 2022, and 502,917 and 533,137 in 2023. The apparent disparity between the counts from 2022 and 2023 highlights that the surveys were conducted at different times in August. “[What’s encouraging is that deep learning models with differing methodologies have produced consistent findings,” notes Duporge.

Since the 1970s, earlier estimates of 1.3 million were derived from aerial surveys and have remained largely unchanged. “If we can accurately count all individuals with zero errors based on our results, we estimate the true population size to be around 800,000,” Duporge remarked. “We believe the aerial estimates are inflated, and our count likely reflects a slight underestimation. Some animals may be hidden under trees or outside the survey area, but it’s quite surprising that the count doesn’t exceed 533,137.”

A lower count doesn’t necessarily indicate that the wild population is declining; they may have adjusted their migratory routes. Nevertheless, wildebeests face serious threats, such as habitat loss and fragmentation due to agricultural expansion. Accurately estimating their populations is crucial for implementing effective conservation strategies.

The researchers had previously trained AI models to identify elephants using satellite data, marking the first instance of such a method for conducting individual mammal censuses across large, dispersed populations. The team is now working on a similar approach for detecting and counting African rhinoceroses.

“We should shift towards satellite and AI methods for assessing wildlife populations, particularly for species that inhabit large and diverse landscapes,” suggests Duporge.

The researcher’s model code is now accessible at https://github.com/sat-wildlife/wildebeest