Skip to content

Training assets: hand-digitized labels, NAIP chips, U-Net weights

Latest

Choose a tag to compare

@rbhughes rbhughes released this 03 Sep 14:37
· 2 commits to main since this release

The full training bundle behind the README's results, released so the hand-digitized work is reusable:

  • datacenters.geojson — 190 hand-digitized hyperscale data center footprint polygons (QGIS, named facilities). Also committed in-repo under data/.
  • naip_chips_256_rgbn.tar.gz (157 MB) — 757 NAIP aerial chips, 256×256 px, 4-band RGBN, 0.6 m resolution, covering the labeled facilities and surrounding negatives.
  • training_masks.tar.gz — binary segmentation masks generated from the polygons (bin/create_masks.py).
  • validation_previews.tar.gz — validation visualizations.
  • unet_last.ckpt (89 MB) — trained U-Net weights (2.5M params; Dice 76.3% / IoU 61.7% — see README).

NAIP imagery is USDA public domain. Labels, masks, and weights are released under CC BY 4.0 — cite this repo if you use them. The Clay v1.5 checkpoint is not redistributed here; get it from the Clay Foundation.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QZ4CJoQwA6Nx2XP48pjiUj