giswqs/WHU-Building-Dataset

WHU Building Dataset

The WHU Building Dataset is a widely-used benchmark for building extraction from high-resolution aerial imagery. It contains aerial images at 0.3m resolution with pixel-level binary building masks.

Dataset Description

Property	Value
Resolution	0.3m ground sampling distance
Tile Size	512 x 512 pixels
Channels	3 (RGB)
Classes	2 (Background=0, Building=255)
Format	PNG

Splits

Split	Images	Masks
Train	5,732	5,732
Val	1,228	1,228
Test	1,228	1,228
Total	8,188	8,188

Directory Structure

├── train/
│   ├── Image/      # 5,732 RGB PNG images (512x512)
│   └── Mask/       # 5,732 binary mask PNG images (0=background, 255=building)
├── val/
│   ├── Image/      # 1,228 RGB PNG images
│   └── Mask/       # 1,228 binary mask PNG images
└── test/
    ├── Image/      # 1,228 RGB PNG images
    └── Mask/       # 1,228 binary mask PNG images

Usage

With GeoAI

import geoai

# Download and prepare the dataset (converts PNG to GeoTIFF, remaps labels 255→1)
# See: https://github.com/opengeos/geoai/blob/main/scripts/train_whu_building.py

Direct Download

from huggingface_hub import snapshot_download

path = snapshot_download(repo_id="giswqs/WHU-Building-Dataset", repo_type="dataset")

Pre-trained Model

A pre-trained EfficientNet-B4 + UNet++ model achieving 0.9054 IoU on the test split is available at: giswqs/whu-building-unetplusplus-efficientnet-b4

Label Format

• Background: pixel value 0
• Building: pixel value 255

Note: When training with CrossEntropyLoss, remap labels from 0/255 to 0/1.

Citation

@article{ji2019fully,
  title={Fully convolutional networks for multisource building identification},
  author={Ji, Shunping and Wei, Shiqing and Lu, Meng},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  volume={57},
  number={1},
  pages={108--120},
  year={2019},
  publisher={IEEE}
}

References

• Original dataset: WHU Building Dataset
• GeoAI package: https://github.com/opengeos/geoai