MARS

MARS — Map Autoregressive Transformer — is a feature-extraction model from Microsoft that converts satellite and aerial imagery into attributed, vectorized GIS data ready for map workflows. Rather than the raster-based segmentation that conventional methods rely on, MARS generates native vector features directly for standard feature ontologies, using a GeoformerModel architecture with autoregressive decoding and a CocoAutoregressiveProcessor for tokenization. The model ingests imagery tiles and emits GeoJSON geometries — polygons for buildings, polylines for roads and railways — each carrying a category label and a confidence score. Two checkpoints are deployed: a polygon model trained on building annotations at 256×256 inference, and a polyline model trained on road and railway lines at 512×512.

Producing vectors natively rather than post-processing raster masks improves generalization across feature classes and yields more consistent, higher-quality map features. On building extraction the polygon model reaches 0.637 IoU and 0.492 mAP@50; the polyline model scores 0.817 TOPO F1 and 0.673 APLS on roads, with railways at 0.732 TOPO F1. Built on the HuggingFace ecosystem and accessible through the GeoAI SDK on Azure AI Foundry, MARS slots vector map generation directly into geospatial pipelines — turning raw GeoTIFF or JPEG tiles into the structured, attributed geometry that downstream GIS and mapping systems consume.