The interactive engine is real: findings come from Sentinel-2 surface reflectance acquired on two actual dates, not from a simulation. The whole path is keyless and free — ESA Copernicus data via the AWS Open Data program.
1 · STAC
Scene search
Earth Search finds Sentinel-2 L2A scenes covering the point for both dates, ranked by cloud cover.
2 · COG
Pixel streaming
Red, NIR and green windows (2.56 km) are range-read straight from the cloud-optimized GeoTIFFs.
3 · INDICES
Change math
Per-pixel NDVI, NDWI and brightness deltas classify vegetation, water and bare-ground change.
4 · CLUSTER
Polygons
Changed pixels cluster into hectare-measured map polygons.
5 · FUSION
Analyst brief
A local LLM (ollama) writes the brief from the raw detections.
S2 LIVE vs DEMO
Results carry an engine badge. S2 LIVE means the numbers were computed from real reflectance; the scene IDs and their cloud cover are listed under "Data used", and the report hash commits to those scene IDs. DEMO means no usable scene covered the point (open ocean, persistent cloud, archive outage) and a deterministic placeholder engine answered instead — the response says why.
Why not Sentinel-1 radar
The Sentinel-1 GRD archive is requester-pays on AWS and account-gated at the Copernicus Data Space, so it cannot be served keyless. Rather than fake radar findings, the engine only claims the optical data it actually reads.
Historical imagery
The before/after comparison uses the Esri World Imagery Wayback archive: ~195 dated releases of the global basemap back to 2014, matched to the investigation dates and blended with a slider directly on the map.
