New research finds there is no substitute for real-world observations, but that — under the right circumstances — AI-generated images can be used to improve the performance of computer models used for species identification and biodiversity monitoring.
In the study, North Carolina State University researchers investigated whether AI-generated, or “synthetic,” images of 20 common North American trees could help train image-recognition models when there aren’t enough real images available.
The researchers combined real images of trees from iNaturalist and the Auto Arborist Dataset with AI-generated images, then fed different combinations into a standard image-recognition model to see whether the synthetic images improved its ability to identify trees.
Synthetic images improved the model’s ability to recognize trees when real images were limited, suggesting that AI-generated images could help supplement real-world data when photographs are scarce.
“The greatest potential for this technology may be in studying species and places we know relatively little about,” said the study’s lead author Thomas Lake, a research scholar in the Center for Geospatial Analytics at NC State.
Computer models require large amounts of data — which, in conservation, means images of plants, animals and other organisms — to accurately identify species for research and monitoring. But collecting enough real-world images can be difficult.
While community scientists have helped address this challenge by contributing photos through platforms such as iNaturalist, gaps remain in the availability of images for some species, particularly those that are less frequently photographed.
“Some species are photographed thousands of times, while others may be rare, occur in remote places or simply receive much less attention. Even for common species, photographs may only come from certain places, seasons or viewpoints,” said the study’s co-author Chris Jones, a senior research scholar in the Center for Geospatial Analytics at NC State.
AI-generated images could be used to expand a small collection of real-world images, helping scientists build initial models for species identification and monitoring in areas where observations are limited, while they continue collecting real images.
Lake said it’s important that the potential of AI-generated images for these kinds of conservation efforts not be overstated. In the study, he and his collaborators found that the synthetic images were less effective than real images overall in training the image-recognition model.
“Most of the AI-generated images looked plausible at first glance,” Lake said. “But really, they didn’t capture the fine details and variations that exist in the real world, telling us that these images can look realistic while still missing subtle information that matters in practice.”
That’s why community science remains essential to data-driven conservation efforts, Lake said. Real images contributed by everyday people provide that missing information. In the case of trees, that includes details such as leaf shape, bark texture and even growth form — all important for identifying species.
“AI-generated imagery does not eliminate the need for field observations or community science; it gives us new ways to use those observations, for recognizing species and noticing how our Earth is changing, potentially sooner,” Lake said.
For Lake, the important part is that it starts with real people and real observations. Community science platforms like iNaturalist are “powerful because thousands of people can collectively observe far more than any individual research team could,” he said.
Anyone can participate in community science and help document the natural world, according to Lake. A single photo of a tree, bird, insect or flower can contribute to a larger record of where species occur and how the environment is changing.
“Tech can help us connect and interpret pieces of data at much larger scales, but if we want to understand a changing natural world, we still need people paying close attention and documenting it,” Lake said.
The paper “Synthetic Imagery Improves Ecological Classification When Real Data Are Scarce, Not Direct Substitutes,” is published in the journal Remote Sensing in Ecology and Conservation. Co-authors include Brennen Farrell and Ross Meentemeyer of NC State.
This work was done with support from the Agriculture and Food Research Initiative (AFRI), Data Science for Food and Agriculture Systems (DSFAS), project award no. 2022-67021-36465, from the U.S. Department of Agriculture’s (USDA) National Institute of Food and Agriculture (NIFA).