Our 3D Greenland terrain system. Clone the repo, follow the install instructions below, and start exploring — all terrain data (heightmaps and textures) is downloaded automatically as you fly around. No bulk downloads or pre-processing needed.

The front end in /webserver folder is running vite. The entry point is main.js → main.terrain.js. The SPA uses a heatmap priority approach to determine what tiles/textures to load or drop, and uses Takram three-geospatial for atmosphere and cloud effects. There is a map view (no clouds) to help with navigation and tile debugging.

The backend in /flaskserver manages terrain heightmaps and textures in a SQLite db (terrain.db). It always assumes the user starts from scratch and will rebuild all missing data as needed. FIRST VISIT TO ANY LOCATION WILL BE SLOW — heightmaps and textures are fetched on demand and cached in the DB. Subsequent visits are fast. All tile demand is driven from the frontend heatmap.

Data Sources

Data	Source	Resolution	Notes
Heightmaps (primary)	ArcticDEM v4.1 mosaic via S3	10m	Cloud Optimized GeoTIFF, free, no auth
Heightmaps (fallback)	Copernicus GLO-30 via S3	30m	Used when ArcticDEM has no coverage
Textures	Dataforsyningen WMS	0.2m–1.6m	SPOT 6/7 + aerial ortho, requires free API token
Labeling reference	Google satellite tiles	~0.26 m/px (z18)	Classification ground truth only — measured, never shipped

Coordinate Systems

System	EPSG	Used For
NSIDC Polar Stereographic North	3413	Internal tile grid, DEM, all spatial queries (meters)
WGS84 Lat/Lon	4326	API input/output, camera positions, user-facing coords
Greenland Transverse Mercator	3184	Dataforsyningen WMS requests only
WGS84 Ellipsoid + ECEF	—	Three.js 3D rendering, camera, lighting, ENU frames

EPSG:3413 is the primary internal system. Everything else converts to/from it.

• coords.py handles to_stereo(lat, lon) (4326→3413) and to_wgs84(x, y) (3413→4326)
• Dataforsyningen textures: 3413 → 3184 for the WMS request, then warped back to 3413 (Lanczos)
• Client-side: lat/lon → Geodetic → ECEF for Three.js scene placement

Texture Pipeline

Real imagery down to depth 12, invented detail below it.

dataforsyningen (faithful, depth ≤ 12) → procedural synthesis (gorgeous + deterministic, depth ≥ 13)

1. Bootstrap: On first run, the server seeds the tile grid as empty skeletons (no heightmaps, no textures). These define the quadtree structure up to the target depth.
2. tex-worker fetches from Dataforsyningen WMS (SPOT 6/7 1.6m, EPSG:3184) on demand. If it fails (rate limit, timeout, no coverage), the tile stays uncached and an ancestor texture is cropped and served as a placeholder until the next request retries.
3. Dataforsyningen WMS requires EPSG:3184 (not 3413). The fetch reprojects 3413→3184 for the request, then warps the result back to 3413 with Lanczos resampling.
4. Dataforsyningen runs out of detail around depth 13 (SPOT is 1.6 m/px). Below that, accuracy vs reality stops mattering: depth 12 already tells us what goes where. A coarse class map at d12 scale (water / grey / dark slopes & shadows / green / white — see flaskserver/classifier/storage.py) is the entire semantic contract.
5. Everything below depth 12 is procedural (work in progress): per-class texture and asset synthesis, seeded from absolute EPSG:3413 coordinates so every visit renders identical detail, color-anchored to the real d12 imagery so the transition doesn't pop. Judged on looks, not fidelity. Google imagery is a labeling/measurement reference only and never ships.

Retired approaches, kept in git history only: SUPIR/ComfyUI enhance (dataforsyningen_enhanced/upscaled sources — never worked right), bathymetry flattening, and learned recoloring from external reference imagery.

The eyeball harness is the tile inspector (pipeline.html?tile=<id>): a tile's progress through heightmap → southness → texture → procgen, with per-stage status and keyboard navigation across tiles and depths.

Classifier output lives in terrain.db's classifier_tiles table as a zlib-compressed, image-oriented uint8 label raster plus a class-schema name, dimensions, optional confidence raster, source/model identifier, and timestamp. A fresh database contains no classifier rows. In the 3D view, press C to toggle classifier presentation: missing tiles keep their satellite imagery but desaturate it, while available coarse_v1 tiles are painted as grey / green / dark / white / water classes. Grayscale elevation is used only while a satellite texture is still loading. Deeper terrain tiles inherit a nearest-neighbor crop from the closest classified ancestor.

Troubleshooting

• Tiles not loading? A hard browser reload (Cmd+Shift+R / Ctrl+Shift+R) may be needed to force the frontend to re-request tiles from the backend, especially after restarting the server or clearing the DB.

Default Camera

The camera starts facing north at Nuuk (64.18°N, 51.72°W). You can override this via URL params, e.g. ?lat=66.5&lon=-53.2. Press R to reset the view if shit gets crazy (e.g. the world suddenly turns into a blue ball).

Map Mode

Press M to toggle a 2D map view (no clouds/atmosphere) for navigation and tile debugging. Right-click on a tile to inspect its metadata or open it in the tile inspector.

Press H to swap the map canvas for the live tile-priority heatmap. Press H again to return to the regular map; both views retain the same heading, pan, and zoom. Regular map mode outlines the terrain meshes currently being rendered as a thin grey grid, with the tile under the pointer outlined in red.

In 3D mode, atmosphere sliders allow adjusting cloud density, coverage, and lighting parameters. (WIP — still working out some bugs.)

Press G to open Google Maps' satellite 3D view at the current camera position, heading, tilt, height above ground, and field of view. This is a debugging reference and opens in a new tab.

Setup

Prerequisites

• Python 3.13+ with pip
• Node.js 18+

Backend (Flask)

cd flaskserver
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Create a .env file in flaskserver/ with your API token:

DATAFORSYNINGEN_TOKEN=<your-token>

To get a free token: create a user account (click "Log ind" → "Opret Profil"), confirm via email, then log in and go to your profile → "Administrer token til webservice og API'er" to generate a token.

Asset Server (TypeScript + SQLite)

The asset server manages vehicles, structures, and their placement via a SQLite database (assets.db). It must be running for vehicles and structures to load in the frontend. Still WIP — schema and endpoints may change.

cd assetserver
npm install

Frontend (Vite)

The frontend depends on a local clone of the Takram three-geospatial monorepo for atmosphere and cloud effects. Vite aliases @takram/* imports to this local source.

cd webserver
git clone https://github.com/takram-design-engineering/three-geospatial.git
cd three-geospatial && git checkout ab3d1cf5 && cd ..
npm install

Running

# Terminal 1 — asset server (assets.db)
./assetserver/runAssetServer

# Terminal 2 — terrain backend (terrain.db)
./flaskserver/runFlaskServer

# Terminal 3 — frontend (snazzy 3d ux)
./webserver/runViteServer

Browser asset calls go directly to http://127.0.0.1:8787 by default. Override in the frontend URL with ?assetServer=http://<host>:<port>.

Scripts log to their respective directories (runAssetServer.log, runFlaskServer.log, runViteServer.log).

Data Attribution

This project uses the following external data sources:

• Satellite orthophotos: Indeholder data fra Klimadatastyrelsen (formerly Styrelsen for Dataforsyning og Infrastruktur). Datasets: "Grønland Satellitfoto" (SPOT 6/7 1.6m regional orthophoto, 0.2m aerial orthophoto). Fetched on demand via Dataforsyningen WMS. Data is free for both commercial and non-commercial use with attribution. Terms of use.
• Heightmaps (primary): ArcticDEM v4.1 10m mosaic, provided by the Polar Geospatial Center under NSF-OPP awards 1043681, 1559691, and 1542736. CC-BY-4.0, free for commercial use with attribution. Fetched on demand via S3. Acknowledgement policy.
• Heightmaps (fallback): Copernicus GLO-30 DEM, provided by the European Space Agency. Free for commercial use with attribution. Fetched on demand via S3.
• Atmosphere & clouds: three-geospatial by Takram.

About

This project exists to promote the development of Greenland. Its incredibly harsh and remote terrain is almost like another planet — perfect for autonomous machinery, robotics, drones, and infrastructure like tunnels. A high-fidelity 3D terrain navigator is a first step toward making that landscape explorable and plannable for all Greenland enthusiasts!