rbhughes/hyperscale-datacenter-segmentation-naip

Hyperscale Data Center Segmentation (NAIP)

Hand-digitized training data for detecting hyperscale data center footprints in aerial imagery, with a trained baseline model.

Contents

path	what it is
datacenters.geojson	190 hand-digitized data center footprint polygons (QGIS; named facilities, e.g. vantage_0)
chips/	757 NAIP aerial chips, 256×256 px, 4-band RGBN, 0.6 m resolution (GeoTIFF, georeferenced) — labeled facilities plus surrounding negatives, largely around the Texas hyperscale corridors
masks/	binary segmentation masks (datacenter / background) generated from the polygons
validation/	validation visualizations from training
model/unet_last.ckpt	trained U-Net baseline (2.5M params, 4-band input): Dice 76.3% / IoU 61.7% / P 75.9% / R 76.7%

Why this exists

This dataset came out of an experiment fine-tuning the Clay foundation model for data center segmentation. The negative result is part of the point: Clay's frozen encoder — pretrained on earth-observation features — topped out at 37% IoU on these out-of-distribution industrial structures, while a small U-Net trained from scratch on this dataset reached 62% IoU. Full writeup, training code, and reproduction scripts: rbhughes/clay-ai-hyperscale.

If you can make a geo-foundation model beat the from-scratch baseline on this data, the author would genuinely like to hear about it.

Provenance & license

• NAIP imagery is USDA public domain.
• Labels, masks, and model weights: CC BY 4.0 — free to use, please cite this dataset / the GitHub repo.
• Digitized July–December 2025 against then-current NAIP; data centers get built fast, so expect the ground truth to age.

Citation

Hughes, B. (2025). Hyperscale Data Center Segmentation (NAIP).
https://huggingface.co/datasets/rbhughes/hyperscale-datacenter-segmentation-naip