A satellite spotted 688 elephants. A plane counted 816. The missing animals reveal an AI blind spot

A satellite can photograph an enormous landscape in a single sweep. That sounds like a solution to one of conservation's most expensive recurring problems: how to count wild animals across terrain that is vast, difficult and sometimes dangerous to survey by aircraft. But when researchers compared satellite and aerial elephant counts in Botswana, the most useful discovery was not that AI had made wildlife surveying automatic. It was a measurable gap between the elephants visible from above and those recorded by observers flying closer to the ground.[1][2]

In a study published in Ecological Solutions and Evidence in September, Rebecca Wilks and colleagues developed a framework for counting African savanna elephants using high-resolution satellite imagery and an AI detector. Their key advantage was a synchronous satellite-and-aircraft comparison. Much AI wildlife research tests whether an image classifier can locate animals already visible in curated pictures. A population survey must answer a harder question: how many animals were present in a landscape, including those nobody could see in the picture?[1][2]

The count that would have made two headlines

The aircraft recorded 816 elephants during the matched count; the satellite imagery yielded 688 visible animals. That is 128 fewer, an undercount of about 15.7% relative to the airborne comparison. A selective headline could praise the 688 animals found from orbit. Another could condemn the 128 missed. Both would obscure the scientific finding: detection and abundance are different statistical problems.[1][2]

Some missing elephants may simply have been under vegetation. Others can blend into terrain, huddle closely with herd mates, be small relative to 30-centimetre image pixels or resemble rocks and shadows. Even a person with the best possible classifier cannot identify an animal hidden completely beneath canopy in the moment of imaging. The machine's pixel-level accuracy is bounded by the information that reaches the sensor.[1][2]

The researchers' proof-of-concept AI detector also illustrates why model-validation numbers need their context. In one data configuration, its F2 score—an accuracy-family statistic weighting missed detections more strongly—was 0.89 on validation images but fell to 0.63 on held-out test images. Those figures are scores, not percentages of the total elephant population found. Training and testing across the same familiar landscape is not equivalent to transferring a detector reliably to a new habitat or season.[1][2]

Correcting one bias can create another

To estimate abundance, researchers considered the fraction of animals likely hidden by different vegetation classes. In their illustrative canopy-correction analysis, the raw satellite count became an estimate of 929 elephants. That figure exceeded the 816-animal aircraft count by roughly 13.8%. The correction went from undercounting to overcounting, which tells us why a plausible adjustment is not automatically an accurate one.[1][2]

This is a familiar challenge in survey statistics. When an observation system sees only a proportion of what exists, statisticians divide or otherwise adjust by an estimated detection or availability probability. If that probability is wrong, the correction can amplify error. Canopy varies by habitat; herds move; detection rates change with lighting, season and how tightly animals gather. A population estimate needs confidence intervals and field-grounded assumptions, not only bounding boxes drawn around impressive photographs.[1][2]

Satellites should join the survey team

Aircraft surveys remain expensive and cannot realistically visit every location continuously. Satellites cover large areas without overflight permits and could help identify concentrations, changes and places worth closer investigation. The authors' current recommendation is integration, not instant replacement: combine satellite detections with aerial observations, evaluate bias corrections, and use additional information such as GPS collars where appropriate.[1][2]

The lesson reaches beyond elephants. AI often performs well at identifying visible objects while the decision-maker needs an estimate of a population hidden partly outside the dataset. Counting the animals you can see and inferring all the animals that exist are two different tasks. Botswana's 128 missing elephants put that distinction into numbers a conservation manager can use.[1][2]

Report an error or suggest a correction