Manually digitizing building footprints from satellite imagery is still eating weeks of GIS analyst time

Extracting building footprints from high-resolution aerial imagery has historically meant either expensive proprietary software or painfully manual digitizing in QGIS. This tutorial replaces both with a Python-native pipeline that goes from raw raster to cleaned, georeferenced building polygons.

One notebook takes NAIP imagery to regularized building polygons

A Tutorial on GeoAI walks you through configuring a geospatial deep learning environment, downloading NAIP raster imagery and vector labels, then generating image chips and segmentation masks for training. You train a U-Net with a ResNet-34 encoder, run sliding-window inference on unseen scenes, convert predicted masks into regularized building polygons, and benchmark results with IoU and F1 scores against zero-shot Grounding DINO plus SAM and a pretrained Mask R-CNN model. The pipeline also connects to Microsoft Planetary Computer for real-world NAIP imagery and Overture Maps for building labels, so it is not confined to toy datasets.

Geospatial ML teams are the first to put this to work

• GIS analysts who spend days hand-digitizing footprints for urban planning or damage assessment get an automated polygon output they can edit rather than build from scratch.
• Remote sensing researchers who need reproducible segmentation benchmarks get a single notebook comparing four model architectures on the same imagery.
• Infrastructure and insurance data teams who need building counts and footprint geometry at scale get a pipeline that extends to any NAIP-covered region without retraining from scratch.

The specificity here matters: this is not a generic segmentation tutorial dressed up with satellite imagery.

Zero-shot geospatial segmentation has quietly crossed a usability threshold

Grounding DINO and SAM together can now segment buildings in aerial imagery without any labeled training data, a capability that did not exist in a usable form eighteen months ago. As NAIP imagery coverage expands and Overture Maps building labels grow denser, the gap between what a solo analyst can extract and what a full annotation team could produce is narrowing fast.

What you can do with this pipeline

• Train U-Net on custom aerial imagery chips with one script.
• Run zero-shot building detection using Grounding DINO and SAM.
• Convert raw segmentation masks into regularized GeoJSON polygons.
• Benchmark U-Net, SAM, and Mask R-CNN on the same scene.

Pricing not listed — check our directory.

The ResNet-34 encoder is where most beginners hit a wall

This tutorial assumes comfort with Python geospatial libraries including rasterio and geopandas; anyone without that foundation will stall before the training cell runs.

For pure semantic segmentation without the instance separation, TorchGeo offers a cleaner entry point with more pretrained geospatial backbones. If polygon regularization is the only missing piece in an existing pipeline, the standalone buildingregulariser library handles that without the full stack.

Open-source GeoAI is moving faster than most enterprise GIS vendors can respond

We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.