FOTBCD

FOTBCD is a large-scale building-change dataset assembled from multi-year French aerial imagery and topographic building data. This release contains 223,675 georeferenced before-and-after image pairs, 447,350 PNG images, and 960,880 COCO-format building instances across 28 departments in mainland France.

Each pair is a 512 × 512 RGB crop at 0.2 metres per pixel. Instance polygons distinguish buildings that are UNCHANGED, DEMOLISHED, or NEW. Spatial metadata uses Lambert-93 (EPSG:2154).

Geographic coverage

Twenty-five departments provide training data. Departments 13, 29, and 92 are held out geographically and supply validation and test samples, preventing local imagery from leaking between training and evaluation.

Dataset summary

Split	Image pairs	PNGs	Instances	Departments
Train	221,712	443,424	952,474	25
Validation	984	1,968	4,261	3
Test	979	1,958	4,145	3
Total	223,675	447,350	960,880	28 unique

The train split includes 99,771 changed patches, 99,770 unchanged patches, and 22,171 empty patches. Validation includes 611 changed, 256 unchanged, and 117 empty patches. Test includes 607 changed, 256 unchanged, and 116 empty patches.

Before images were acquired from 2014 through 2017 and after images from 2021 through 2024.

Repository layout

The dataset is stored as 171 WebDataset TAR shards, totalling approximately 170.006 GiB:

data/
├── train/
│   └── train-000000.tar ... train-000168.tar
├── val/
│   └── val-000000.tar
└── test/
    └── test-000000.tar

Every sample has three members with a shared zero-padded key:

000000123.before.png
000000123.after.png
000000123.json

The JSON member contains the image-pair record, its COCO polygon annotations, and the category definitions. Category IDs are:

ID	Name	Meaning
1	UNCHANGED	Building present at both dates
2	DEMOLISHED	Building present before and absent after
3	NEW	Building absent before and present after

Loading the data

Install WebDataset and Pillow:

pip install webdataset pillow

Stream the training shards without downloading the whole archive first:

import webdataset as wds

url = (
    "pipe:curl -s -L "
    "https://huggingface.co/datasets/abdevd/FOTBCD/resolve/main/"
    "data/train/train-{000000..000168}.tar"
)

dataset = (
    wds.WebDataset(url, shardshuffle=True)
    .shuffle(5_000)
    .decode("pil")
    .to_tuple("before.png", "after.png", "json")
)

before_image, after_image, annotation = next(iter(dataset))

Use the corresponding single-shard URL for validation or test:

val_url = "https://huggingface.co/datasets/abdevd/FOTBCD/resolve/main/data/val/val-000000.tar"
test_url = "https://huggingface.co/datasets/abdevd/FOTBCD/resolve/main/data/test/test-000000.tar"

Dataset creation

FOTBCD combines two source products maintained by the Institut national de l’information géographique et forestière, France’s national mapping agency:

• BD ORTHO: orthorectified RGB aerial imagery from repeated national surveys.
• BD TOPO: authoritative building footprints and attributes.

Successive BD TOPO snapshots were aligned with the corresponding BD ORTHO acquisitions to infer building-level transitions and produce paired crops and polygon annotations. Training examples were generated at scale using rule-based filtering, domain heuristics, learned checks, and aligned imagery. The geographically held-out validation and test examples were manually selected and verified.

Quality control considered acquisition-date consistency, building geometry and metadata, implausible transitions, edge cases, and final human review. Large automatically generated training sets may nevertheless contain annotation noise.

Intended uses

FOTBCD is intended for research and evaluation in:

• building change detection;
• semantic and instance segmentation of temporal aerial imagery;
• construction-versus-demolition classification;
• geographically held-out evaluation and domain-shift studies;
• representation learning on paired remote-sensing imagery.

Limitations

• Coverage is geographically diverse but limited to mainland France; international transfer is not established.
• Training annotations were generated at scale and can contain label noise, imperfect geometry, or missed transitions.
• The class distribution is imbalanced, especially for demolished buildings.
• Temporal imagery and map updates may not be perfectly synchronized.

Licensing

FOTBCD is licensed under CC BY 4.0.

Source imagery and building data are provided by the Institut national de l’information géographique et forestière through BD ORTHO and BD TOPO, under the Etalab Open Licence 2.0.

Citation

@article{moubane2026fotbcd,
  title   = {FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data},
  author  = {Moubane, Abdelrrahman},
  journal = {arXiv preprint arXiv:2601.22596},
  year    = {2026},
  url     = {https://arxiv.org/abs/2601.22596}
}

Paper: https://arxiv.org/abs/2601.22596