EarthOS — give it a coordinate, and it reads the land

TL;DR — Type a coordinate. EarthOS pulls the cleanest recent satellite scene, fuses eight observation sources on top of 0.24 m/px aerial imagery, auto-identifies vegetation / water / built-up / surface-temperature / terrain, cross-checks against an international land-cover standard, and then reasons about how much you could actually build from the zoning code. Live demo: https://vidraft-earthos.hf.space/

Why a "world model" and not just a map

Most geospatial tools show you pixels. A world model has to read them — turn imagery into land, land into constraints, and constraints into a decision ("what can be built here?"). EarthOS (AETHER-Earth) is our first step toward that for the Korean peninsula.

The pipeline (5 stages)

1. Acquisition — a coordinate triggers a live STAC query; from the last six months it auto-selects the least-cloudy scene.

Layer	Source	Resolution
Aerial orthophoto	V-World WMTS	0.24 m
Optical	Sentinel-2 L2A, 12-band (ESA)	10 m
Radar	Sentinel-1 RTC VV·VH	10 m
Thermal IR	Landsat C2-L2 ST_B10 (USGS)	30 m
Terrain	ALOS AW3D30 (JAXA)	30 m
Land cover	ESA WorldCover, 11 classes	10 m
Cadastre / zoning	V-World vectors	vector
Climate	Open-Meteo reanalysis	point

2. Preprocessing — everything is reprojected to EPSG:4326 for pixel alignment, resolution is adapted to the request extent (10 m ↔ 26 m), and each channel gets a percentile stretch.

3. Auto-identification

• Vegetation / water / built-up → NDVI · NDWI · NDBI thresholding → area ratios
• Surface temperature (urban heat) → DN×0.00341802 + 149.0 − 273.15 → °C
• Structures / water via radar → 10·log₁₀(γ⁰) dB, VV·VH·polarization ratio — works through clouds and at night
• Change / disturbance → ΔNDVI vs. the scene two years earlier → vegetation-loss pixels
• Terrain → differentiate the elevation array → slope & aspect → flat / south-facing ratios

4. Cross-validation — our own classification is compared against the ESA WorldCover international 11-class standard to check for bias.

5. Regulatory inference — zoning → the Korean National Land Planning Act (§84·§85) building-coverage and floor-area ratios → building footprint, gross floor area, estimated storeys.

The interesting engineering bit

Overseas hosting can't reach Korea's V-World service. So EarthOS splits the work: the browser (inside Korea) fetches cadastre, parcels and aerial tiles directly, while the server only computes. FastAPI + a custom 3D-globe UI, in Docker. At the 250 m zoom level you can resolve individual vehicles and lane markings.

Honesty & license

This is reference material — not a basis for decisions or legal judgment. Cadastre is not a survey; coverage/floor-area ratios are legal ceilings and you must check local ordinances. The research track (FINAL-Bench/AETHER-Earth-KR-v1, an OlmoEarth-v1.2 continued-pretraining experiment) is not used in the demo's judgments — only public standard indices and labels are — because its performance edge is not yet statistically established. The research model inherits the Ai2 OlmoEarth Artifact License: no military, surveillance/policing or resource-extraction use.

Try it: https://vidraft-earthos.hf.space/

Built by VIDRAFT (비드래프트). More at https://vidraft.net/