AI uncovers nearly 10,000 hidden field islands across Estonia
University of Tartu researchers used artificial intelligence to scan Estonia's farmland and found roughly 10,000 more tree-and-shrub 'field islands' than are listed in the official register. These tiny habitats are important nesting and foraging sites for birds and beneficial insects.
Scientists at the University of Tartu have built an artificial intelligence tool capable of automatically detecting field islands — small clusters of trees and shrubs scattered across farmland. Applied across Estonia, the system located roughly 38,000 such islands, about 34 percent more than appear in the official registry kept by the Agricultural Registers and Information Board (ARIB).
Why it matters
Despite covering tiny areas — typically between 0.01 and 0.5 hectares — field islands serve as nesting grounds for farmland birds and shelter for wild bees and other beneficial insects that pollinate crops and keep pests in check. Because of their ecological value, farmers can receive European Union Common Agricultural Policy support for preserving them.
How the technology works
Until now, mapping field islands relied on farmers manually drawing them in ARIB's digital portal when applying for EU support, a process that often produced incomplete or outdated maps. The Tartu team trained a deep-learning model on more than 15,000 labelled examples drawn from high-resolution aerial photographs and existing farmer-drawn maps. Once trained, the model automatically analyzed nearly 2,000 aerial photographs covering the entire country.
The largest numbers of overlooked field islands turned up in central Estonia, where large farm fields are common. The only region where the AI found fewer islands than the official register was northeastern Estonia, where many field islands consist mainly of rocks and shrubs rather than trees. Notably, the model performed better with slightly lower-resolution images, since viewing a wider area helped it distinguish isolated field islands from hedgerows or forest edges.
The researchers stress that the findings do not indicate the official maps are wrong, but rather show how difficult it is to keep track of thousands of small landscape features through manual mapping alone. The dataset and code have been made publicly available so the method can be applied elsewhere. The study was published in the journal Geocarto International.
