The stack
Five ways to read one mine.
Two optical passes seven years apart, shortwave infrared, radar and terrain, over the same 250 km² on one 12.5 m grid.
Sentinel-2 · Sentinel-1 · Copernicus DEM
Industry · Mining
Watch a mine from orbit.
A new cut, a dump outside the plan, a stockpile that grew. MineScope reads optical, radar and terrain over a mine and hands a person the few places that need a look.
- MineScope
- TerraScope
- Atlas
17.10° S · 70.63° W · 250 km²
Seven years of one mine, in seven layers.
Terrain
Radar
Shortwave2024
2017
Flag 01 · 1,007 ha
Contains modified Copernicus Sentinel data (2017, 2018, 2024). Copernicus DEM GLO-30, provided under COPERNICUS by the European Union and ESA.
Real
Every layer is public satellite or elevation data, processed by a script in this site's repository. The change map and the flag are computed, not drawn.
Illustrative
The district is a public example, not a client site. The pipeline here is a simplified stand-in for the product engine, and nothing on this page was checked against ground truth.
What runs here.
Each product below works on mining today. Atlas is the engine they share.
MineScope
Change, encroachment outside the permit, plan against actual footprint, and stockpile volumes from radar and terrain.
TerraScope
The land suite. MineScope, AgriScope and ForestScope ship inside it.
OpenAtlas
The engine under every product: fusion, registration, detection, and an agent on the map.
Open
What it has shown.
Numbers from real runs, each with the caveat it needs. Nothing here is a claim the run did not make.
- 185
change polygons on a public coal mine, 671.47 ha in total
Atlas change detection, one run. Largest 483.64 ha, median 0.11 ha.
- 2 → 187
detections on the same scene once tiling went multi-scale
A count from an ablation, not an accuracy figure.
- 5.3%
gap between radar and terrain volume methods on four stockpiles
Two independent methods at a South American mine. No ground truth yet.
How it works.
Open any part for the method, the numbers and where to read more.
Before and after are cut into tiles at three scales and passed through the same frozen DINOv2 backbone. A small head reads the difference.
Pixel difference, structural similarity and edges run beside it and always return a map. Where the learned score is unsure, the weight leans on the classical one, so the polygons stay clean on a site the model has never seen.
A CLIP pass proposes a name for each polygon from a prompt registry, zero-shot.
Backbone
DINOv2 ViT-B/14, frozenLabels
CLIP ViT-B/32, zero-shotMean confidence
30.2%, the honest mid band
On a public mine, one run gave a net land-use change, a comparison of the planned quarry footprint against what was actually disturbed, encroachment flags outside the boundary, and cut and fill volumes from two radar epochs.
These are readings, not validated results. No ground truth was available for that site.
Plan against actual
541 planned, 227 disturbed acresEncroachment flags
167Radar volume
+4.78 Mm³ cut, −2.91 Mm³ fill
At a South American mining operation, a hybrid of radar and elevation estimated about 276,000 m³ across four stockpiles. The radar method alone read 270,000 m³ and the terrain method 285,000 m³.
Survey ground truth has not been compared yet, so the number is a cross-check, not an accuracy.
Mining
On screen.
Product screens from the cleared bundle. Names and accounts are painted out.
Why it matters.
Public figures on the size of the problem.
About 1.4 trillion litres of oil sands tailings were in storage as of 2023, and every planned release will need monitoring that can be verified.
Source: industry research (Alberta Energy Regulator)
Mining · Next step
See it on your own ground.
Send an area and a question. I will say what the sensors can and cannot answer before anyone builds anything.