Skip to content
← Directory

software · Global

Microsoft MARS (Map Autoregressive Transformer)

MARS (Map Autoregressive Transformer): Microsoft feature-extraction model converting satellite/aerial imagery into attributed GeoJSON vectors: building polygons (256×256) and road/railway polylines (512×512). Accessible via the GeoAI SDK on Microsoft Foundry; built on HuggingFace.

Visit source ↗

Claim this listing

Represent Microsoft MARS (Map Autoregressive Transformer)? Send a claim for review so we can discuss listing updates. Claims do not automatically grant ownership or verification.

Explore paid sponsorship →

What's in it

Data offered docs-verified
building polygons (GeoJSON) · road polylines (GeoJSON) · railway polylines (GeoJSON) · category label and confidence score per feature
Formats docs-verified
GeoJSON (output) · GeoTIFF (input) · JPEG (input)
Access docs-verified
Via Microsoft Foundry and the GeoAI SDK; licence: see provider; pricing not stated

Useful for

Use cases llm-inferred
Building footprint extraction from aerial imagery for flood exposure screening · Road and railway centreline extraction for linear infrastructure baseline mapping · Converting aerial or satellite tiles into GIS-ready vector layers for site constraint maps
Audience llm-inferred
GIS analysts · flood risk consultants · planners · BIM/civils engineers

Limitations

Constraints llm-inferred
Experimental Labs project, not a production product · Building polygon model scores modestly (0.637 IoU, 0.492 mAP@50) · Separate checkpoints for polygons (256×256) and polylines (512×512) · Geographic and temporal coverage not stated · Requires Microsoft Foundry access via the GeoAI SDK
Licence
see provider
Authentication
Unknown — review source docs
GitHub stars
Unknown
Last commit
Unknown
Last successful page check
14/10/2026
Page health
healthy
Package
Not recorded
Verification
Not independently verified

Provenance & evidence

Verified 2026-10-15 via curl (HTTP 200).

Metadata observed 14/10/2026 · research sweep: manual-2026-10-15 (Chris sources (Microsoft Foundry Labs))