AI Turned Millions of Wildlife Photos Into Ecological Maps

AI turned millions of wildlife photographs into ecological maps that closely resembled those built from expert identifications. In a study published on May 6, 2026, an automated image-identification workflow processed roughly 3.78 million camera-trap images from three landscapes, then supported occupancy models with strong agreement on many species’ relationships with their environment. For conservation teams, that opens a practical possibility: spending less time sorting photographs and reaching useful ecological answers sooner.

A camera trap waits through darkness and long stretches of nothing, then catches a lynx crossing a trail or a jaguar visiting a water hole. At landscape scale, those moments accumulate into millions of files. The important advance here is a tested route from those files to estimates of where wildlife occurs.

Research references: [1]

Ecological patterns emerged from millions of images

Researchers from Washington State University, Google Research and conservation partners compared an automated workflow with expert-labelled datasets from Washington, Glacier National Park in Montana and Guatemala’s Maya Biosphere Reserve. The surveys were conducted during 2016–2021, across temperate coniferous and subalpine forests and seasonally dry tropical forest.

The image totals were approximately 1.8 million from 638 Washington cameras, 1.7 million from 263 Montana cameras and 280,000 from 326 Guatemala cameras. Together, that is 1,227 cameras. The study focused on primarily terrestrial wild mammals weighing more than one kilogram as adults.

MegaDetector located animals; SpeciesNet classified the cropped images. SpeciesNet is a general classifier whose labels include species and broader taxonomic groups. The researchers combined it with local species lists, sequence processing and confidence thresholds before fitting multi-species occupancy models. The automated image-labelling arm had no human review of its labels; the scientific design and ecological processing choices remained human-made.

Research references: [1] [3]

The models often reached similar conclusions

Across six pipeline variants, the direction of species–environment effects agreed with expert-derived models in 83–94% of the compared cases. Agreement on variable importance was 88–99%, where importance meant that a 95% Bayesian credible interval excluded zero. These comparisons concerned species present in both modelled datasets, and measure agreement between workflows rather than image-classification accuracy.

For the spatial comparison, the researchers used the 50%-confidence variant with cleaned image sequences. For more than three-quarters of species, differences between AI- and expert-derived occupancy maps stayed between −0.1 and +0.1 on the probability scale across the study region: within ten percentage points. These are estimated probabilities of occurrence, rather than animal counts.

That is a meaningful result for conservation science. Ecologists often need to understand how a species is distributed and how that distribution relates to its surroundings. The study shows that imperfect image labels can still support those answers for many mammals, including species of substantial management interest.

Research references: [1]

Faster answers could widen conservation capacity

Washington State University’s account attributes the traditional six-to-seven-month processing period, sometimes reaching a year, to lead author Daniel Thornton’s experience. It reports that automated processing can take a few days, or roughly a week. These are the team’s practical workflow estimates, not a controlled timing benchmark.

Shortening that backlog could help conservation teams investigate habitat use and changes in wildlife distribution while the information is still useful. It could also make larger monitoring projects more manageable for groups with limited staff. Those are promising applications; the study evaluated model agreement rather than conservation interventions.

Research references: [2]

Focus expert attention where it matters

Performance varied by species and location. Some visually similar animals were difficult to separate, and all 94 mountain-goat detections in the Montana dataset were classified as domestic goats. The main analysis also excluded extremely sparse species records. Rare-species inventories and abundance estimates need their own assessment rather than inheriting the occupancy results.

The useful division of labour is becoming clearer: automate repetitive identification where the ecological question and evidence support it, while concentrating expert review on consequential exceptions. A camera network can gather more observations than a field team could watch directly. This work shows how existing AI tools can help turn that extraordinary reach into ecological knowledge.

Disclosure: One study co-author works for Google, which developed SpeciesNet.

Research references: [1]