hughbertlongshanks/us_southeast_building_footprints

Southeast U.S. Building Footprints Segmentation

~50,000 train and ~7,500 test building segmentation imagery chip pairs for the U.S. Southeast (South Florida - urban , North Florida - rural , Eastern Kentucky - rural). Imagery is 512x512 0.3M RGB NAIP Imagery in GeoTIFF in EPSG:5070, segmentation masks are 512x512 with 1 being a building detection and 0 being background. Data is attempted to be time aligned between building footprints and imagery but discrepancies may occur ( Florida and the U.S. Southeast is quickly growing ). Images and ground truth masks are LZW compressed and should be fully readable by standard raster reading software but use caution with standard image readers.

Dataset Sources

OSM and MS Builidng Footpints and USGS NAIP imagery for 2021-2025.

Uses

Building segmentation, each instance has at least 1 building classified pixel but the amount varies significantly between rural and urban examples.

Dataset Structure

Imagery and ground truth masks are stored in seperate directories but share the same basename. E.g. , fl_county_123_01_01.tif in /image_chips corresponds to the same name as a ground truth mask under /gt_masks.

Curation Rationale

Fine-tuning and training of building segmentation models tuned for the U.S. Southeast.