What is Russia hiding under fake forests?
At first glance, nothing seems off about the image of the forest outside Kronstadt, Russia's historic port city. It shows some infrastructure, lakes, grass and forest. The satellite image comes from Yandex, Russia's most popular map service. But while Yandex shows the location as seen above, other services depict it very differently: where Yandex shows a quiet patch of trees, other providers show a military missile battery.
In August 2026, the Finnish broadcaster Yle revealed why the base was nowhere to be found on the map: Yandex had copied forest from another part of the same satellite image and pasted it over the missile site. Hiding sensitive locations is not unusual in itself – on map services they are usually blurred. That approach hides details, but anyone looking closely can still tell that something is there. Since Western intelligence services use their own satellite data, it remains unclear whom the base was being hidden from on Yandex. In this case, the fake forest did the opposite of protecting it.
Hidden traces turn into clues
Concealment on maps has led investigators to missing pieces of information before. In 2020, journalists Alison Killing, Christo Buschek and Megha Rajagopalan investigated China's mass detention of Uyghurs in Xinjiang. They noticed that Baidu, China's leading map service, laid blank grey tiles over certain areas – and that these masks clustered around known camps. Buschek turned the map anomaly into a tool to collect the locations of potential mass detention centres. The team analysed tens of thousands of masked tiles and checked each one against satellite imagery from other providers. This method led them to more than 260 structures they believed to be detention camps and forced-labour facilities. The investigation won a Pulitzer Prize in 2021.
Yandex's fake forest is a similar map anomaly, just less visible to the human eye. So we wanted to find out whether a machine would catch it.
What our analysis tools found – and what they missed
Together with our partner Decision Labs, we built a tool to run an automated analysis on the Yandex imagery. In the first run, the detectors built to spot AI manipulation did not flag the edit. Those tools are trained on deepfakes and other generative-AI forgeries. A hand-copied patch is an older, simpler kind of fake – to these detectors, the cloned forest looks ordinary. But two other methods, applied to selected imagery, did work:
1. Tie-point matching. The software picks out thousands of small, distinctive spots in the image – such as the edge of a tree crown or a gap between two trees – and looks for spots that appear twice. In a natural forest, some lookalikes will always be found, but they are scattered at random, in all directions and at all distances. A copied patch behaves differently: every matching spot in the copy sits at exactly the same distance and in exactly the same direction from its twin, as if someone had traced the forest onto paper and slid it across the map. That shared shift is the fingerprint of a copy-and-paste. On the Yandex image, the software found three such cases: three pairs of identical areas. Still, forest looks alike everywhere, so two similar patches prove little on their own.
2. Embeddings. This method uses GeoEmbeddings, an AI model that turns each small square of the image into a numerical summary of what it looks like. Squares that look alike get similar summaries, and we mark them in the same colour on the map. Around Kronstadt, the copied areas light up in matching colours, which supports the finding. On its own, this method proves less, because real forest is repetitive too and many untouched squares also look alike. It is most useful as a second opinion alongside tie-point matching.
In the demo, the labels A and B only show the direction of a match, not which patch is the original. Any finding still has to be triangulated with other satellite images or sources, as open-source investigator Benjamin Strick explains in his latest video.
In partnership with Vertical52, Decision Labs built an interactive demo that runs in the browser, so you can follow the matches yourself. Try the tool and get in touch. The detectors come from the VeriSat project. We will keep building and investigating.
Are you working with satellite imagery and suspect a location has been tampered with? We at Vertical52 are happy to help: reach out, and we can run the analysis on your images and support your investigation.
