FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data

We introduce FOTBCD, a large-scale building change detection dataset derived from authoritative French orthophotos and topographic building data provided by IGN France. Unlike existing benchmarks that are geographically constrained to single cities or limited regions, FOTBCD spans 28 departments across mainland France, with 25 used for training and three geographically disjoint departments held out for evaluation. The dataset covers diverse urban, suburban, and rural environments at 0.2m/pixel resolution. We publicly release FOTBCD-Binary, a dataset comprising approximately 28,000 before/after image pairs with pixel-wise binary building change masks, each associated with patch-level spatial metadata. The dataset is designed for large-scale benchmarking and evaluation under geographic domain shift, with validation and test samples drawn from held-out departments and manually verified to ensure label quality. In addition, we publicly release FOTBCD-Instances, a publicly available instance-level annotated subset comprising several thousand image pairs, which illustrates the complete annotation schema used in the full instance-level version of FOTBCD. Using a fixed reference baseline, we benchmark FOTBCD-Binary against LEVIR-CD+ and WHU-CD, providing strong empirical evidence that geographic diversity at the dataset level is associated with improved cross-domain generalization in building change detection.

💡 Research Summary

The paper presents FOTBCD, a new large‑scale benchmark for building change detection that leverages two authoritative French geospatial products: the BD ORTHO aerial orthophoto archive (0.2 m/pixel, radiometrically corrected, nationwide coverage) and the BD TOPO vector database (building footprints, construction/demolition dates, semantic attributes). By aligning temporal snapshots of BD TOPO with corresponding BD ORTHO imagery, the authors automatically infer building‑level changes (new, demolished, unchanged) across 28 French departments. Two public releases are provided:

• FOTBCD‑Binary – ~28 000 before/after image pairs with pixel‑wise binary change masks (CHANGE/NO‑CHANGE). The dataset is split at the department level: 25 departments for training, 3 geographically disjoint departments for validation and testing. Each 256 × 256 patch carries Lambert‑93 (EPSG:2154) georeferencing metadata, enabling seamless GIS integration.
• FOTBCD‑Instances – a smaller, research‑focused subset (≈4 000 pairs) with instance‑level polygon annotations in COCO format, distinguishing NEW, DEMOLISHED, and UNCHANGED buildings. This subset illustrates the full annotation schema and supports instance‑level and multi‑class change studies.

To guarantee label quality, the authors implement a four‑step quality‑control pipeline: (1) temporal alignment of imagery and vector snapshots, (2) topological and semantic validation of building geometries, (3) AI‑based detection of residual inconsistencies, and (4) manual verification of all validation and test samples. The training set remains automatically generated and may contain minor noise, but experiments show this does not materially affect performance.

For benchmarking, the authors adopt HybridSiam‑CD, a Siamese change‑detection architecture that fuses a frozen DINOv3‑sat493M Vision Transformer (semantic features) with a ResNet‑34 backbone (spatial and boundary cues). The model is trained with a fixed protocol across all datasets (50 k steps, batch size 128, AdamW, cosine LR schedule, combined Lovasz‑hinge and boundary‑aware BCE loss, standard augmentations).

Cross‑domain experiments reveal stark asymmetries: models trained on geographically homogeneous datasets (LEVIR‑CD+, WHU‑CD) achieve low IoU scores (≈0.30–0.34) when evaluated on FOTBCD‑Binary, indicating poor transfer to a nationally diverse setting. Conversely, a model trained on FOTBCD‑Binary attains an IoU of 0.70 on WHU‑CD and 0.30 on LEVIR‑CD+, demonstrating that exposure to varied building typologies, urban densities, land‑cover contexts, and imaging conditions yields more robust feature representations. The authors also observe mutual degradation when transferring between LEVIR‑CD+ and WHU‑CD, confirming that geographic domain shift is a general challenge in building change detection.

The discussion emphasizes that geographic diversity at the dataset level is more influential for out‑of‑distribution generalization than sheer dataset size, echoing findings in other computer‑vision domains. Training on 25 French departments exposes the model to multiple climate regimes (Mediterranean, Atlantic, continental, alpine, semi‑arid), architectural styles (dense apartments, detached houses, industrial facilities, agricultural structures), terrain variations, and differing acquisition conditions, encouraging learning of change‑specific cues rather than region‑specific appearance patterns.

Limitations are acknowledged: the benchmark is confined to mainland France, focuses solely on building changes (excluding roads, vegetation, water bodies), and relies on automatically generated labels that may contain residual noise. Future work is suggested to expand to multi‑country or multi‑continent coverage, incorporate additional change categories, and develop models capable of simultaneous multi‑class and instance‑level detection.

Overall, FOTBCD provides the remote‑sensing community with a high‑resolution, geographically diverse, and well‑validated benchmark that enables more realistic assessment of building change‑detection algorithms and promotes research on domain‑generalizable models.