SAM, but for satellite imagery: point at a rooftop, get a shapefile

What it does SamGeo takes the Segment Anything Model — plus variants like HQ-SAM, Fast SAM, and SAM 2/3 — and points it at GeoTIFFs and satellite imagery. It can pull map tiles from TMS servers, segment objects with interactive markers, box prompts, or plain text prompts (via Grounding DINO), and export results as GeoPackage, Shapefile, or GeoJSON. There’s also a REST API option and a QGIS plugin for the desktop-GIS crowd.

The interesting bit The glue is the value: SAM knows nothing about coordinate reference systems, and SamGeo quietly handles the boring georeferencing so masks come out as actual vectors you can drop into a GIS. Text prompts like “swimming pools” over a city image are where it feels almost unfairly easy.

Key highlights

• Multiple model backends: SAM, HQ-SAM, Fast SAM, SAM 2 and 3, with granular extras so you only install what you use
• Text-prompt segmentation via Grounding DINO, including batch mode
• Interactive foreground/background markers, or markers loaded from existing vector data
• Results export to GeoPackage, Shapefile, GeoJSON; prompts saved as GeoJSON
• QGIS plugin, ArcGIS Pro integration, and a FastAPI-based serving option

Caveats

• GPU with at least 8 GB VRAM is recommended for large datasets — this is not a laptop-CPU toy
• SAM 3 on Windows is described by the README itself as “a bit tricky” to install
• The tile-downloading feature comes with a legal notice: get permission from basemap providers before bulk downloads

Verdict Remote sensing folks and GIS analysts who want deep-learning segmentation without writing their own model plumbing should absolutely look. If you don’t know what a GeoTIFF is, this isn’t your starting point.