mars-map-autoregressive

An automated feature extraction model that converts satellite imagery into accurate, attributed, vectorized GIS data for map workflows.

An automated feature extraction model that converts satellite imagery into accurate, attributed, vectorized GIS data for map workflows. Unlike traditional raster-based segmentation methods, MARS generates native vector features for standard feature ontologies, enabling improved generalization across feature classes and delivering high quality, consistent, generation of map features.

MARS uses an autoregressive transformer architecture built on the HuggingFace ecosystem. The model processes satellite or aerial imagery tiles and outputs GeoJSON geometries (polygons for buildings and waterbodies, polylines for roads/railways) with category labels and confidence scores. Waterbodies are detected by a complementary DINOv3-based segmentation model that emits the same attributed vector output, so lakes, rivers, and coastal water are returned as polygons alongside the other categories.

Beyond feature extraction, MARS can also render extracted vectors into a styled, Mapbox-style basemap image through its map rendering endpoint (/map:render). Given a source image and a GeoJSON FeatureCollection — either the georeferenced output of /map:extract or the pixel-space output of /score — it returns a cartographic PNG for quick visualization and map workflows. Rendering supports multiple built-in cartographic themes (default, dark, standard_oil, streets) and optional per-category color overrides, so buildings, roads, railways, and water are drawn as a coherent, styled basemap. See the usage notes for the full /map:render request format and parameters.

Supported Categories

Category	Geometry Type	Model
Building	Polygon	Polygon model
Road	LineString	Polyline model
Railway	LineString	Polyline model
Water	Polygon	Water segmentation model