Two Open Neural Networks Found 28 New Construction Sites in Moscow Satellite Imagery

Reviewing satellite imagery quickly turns into work that cannot reasonably be done by eye. You need to find cloud-free scenes, align different dates, compare hundreds of square kilometers and not lose the result in a folder of rasters. In an experiment covering New Moscow, an AI agent assembled this chain while the computation stayed local. The output was not a flashy demo but a genuinely workable queue: 28 sites worth a human check.

The example is straightforward. A rectangle of 455.60 km² near Troitsk, Desna, Krasnaya Pakhra and Shchapovo was taken. Three summer Sentinel-2 images were found for it: one from July 2025 and two from July and August 2026. The system looked for places where the surface had changed relative to 2025 and where the probability of construction had risen at the same time. After filtering, 455.46 km² remained suitable for comparison.

Two Models Instead of One Guess

The first model, FC-Siam-diff, receives a pair of images and marks the pixels that changed. This is a useful signal, but by itself it says nothing about the cause: a field may have turned yellow, soil may have become wet, and a shadow may have shifted. So Dynamic World was run alongside it. This model divides the surface into nine classes, including water, vegetation, cropland, bare ground and built-up areas.

The final filter required two independent signals to agree. FC-Siam-diff had to see a change on both the July and August 2026 images relative to 2025. Dynamic World had to show a steady rise in the "built-up" class score on both new dates. The author chose the thresholds manually: change probability above 50%, built-up score in 2025 below 35%, and in 2026 above 55% with an increase of at least 25 percentage points.

This approach narrows the search considerably. One model flagged 144 areas totaling 108.32 ha, the other 339 areas covering 170.80 ha. Their intersection left 28 polygons covering 12.44 ha. For the six largest candidates, visual inspection did confirm changes to artificial surfaces or construction sites.

Where the Agent Worked

The agent was assigned not image recognition but process assembly. It found the data through the STAC API of the Earth Search catalog, prepared scripts, brought the bands to a common grid, ran the pretrained model weights and uploaded the layers to a mapping service. Satellite bands were resampled to a 10-meter grid, clouds and shadows were excluded using the SCL layer, and categorical data was interpolated using the nearest-neighbor method.

After inference, the raster masks were cleaned of connected areas smaller than nine pixels, that is, 900 m². They were then converted into GeoJSON polygons with area, date and model type. Together with the source RGB images, these polygons were published as separate map layers. A colleague can open the link, switch on the desired date and check a specific site without setting up any geo tools locally.

What Not to Confuse

These 28 polygons cannot be called 28 new buildings. Sentinel-2 resolution is 10 meters for the main bands, and a single outline may include a road, bare ground and several objects. The hard intersection of thresholds also cuts real sites into pieces and misses construction inside already developed territory. Of the 28 candidates, only the six largest were reviewed manually; the method's recall has not been measured at all.

The practical value lies elsewhere. Off-the-shelf open models and an AI agent can turn 455 km² of imagery into a short queue for an analyst without sending the source data to an external AI service. But before any production use, you need a labeled validation sample, an assessment of precision and recall, versioning of thresholds and manual review. This is a good preliminary screening pipeline, not an automatic cadastre and not proof that a building has appeared.