A Tutorial on GeoAI: Designing Footprint Extraction from NAIP Imagery Using U-Net, Grounding DINO, SAM, and Mask R-CNN

There's a particular kind of tedium in geospatial work that most people never see: someone, somewhere, is tracing building outlines by hand on satellite imagery, tile after tile, county after county. A new tutorial from MarkTechPost, built around the open-source GeoAI Python package, is aimed squarely at killing that job, and it does it by walking through nearly every approach currently in vogue for automated footprint extraction.

The workflow starts with NAIP imagery, the USDA's high-resolution aerial dataset that covers the continental US at roughly one-meter resolution, paired with building polygons pulled from a Hugging Face-hosted dataset. From there the tutorial slices the raster into 512-pixel georeferenced chips with 256-pixel stride, generates matching label masks, and trains a U-Net with a ResNet-34 encoder pretrained on ImageNet. Nothing exotic — 12 epochs, batch size 8, a learning rate of 1e-3 — but the point isn't novelty, it's that this exact recipe is what most practitioners actually run in production.

What makes the piece worth a look is the second half, where it stops being a segmentation tutorial and turns into a comparison test. After training the U-Net, the author runs sliding-window inference on an unseen scene, then pushes the raw pixel masks through polygon cleanup: region-grouping to strip noise, vectorization, and a regularization step that squares off building corners so footprints look like buildings instead of blobs. Then, separately, they try Grounding DINO paired with Segment Anything for zero-shot detection — no training data at all, just text prompts like