SpaceEcoVision/buildout-unet

Built-up segmentation U-Net (Space EcoVision)

Compact 3-level U-Net that segments built-up / impervious surface from multi-band Sentinel-2 tiles. Trained on US spaceport scenes with ESA WorldCover (class 50) as labels; validated on held-out sites (unseen geography).

• Input: B02, B03, B04, B08, B11, B12 (Sentinel-2 L2A), shape (C, H, W)
• Output: per-pixel built-up logit (sigmoid > 0.5 = built-up)
• Trained: 2026-06-03
• Purpose: annual built-up maps to track spaceport buildout over time — WorldCover covers only 2020/2021; this model extends the signal to all Sentinel-2 years.

Code: https://github.com/SpaceEcoVision/terrestrial-impact-engine (cv/)