FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data
Abdelrrahman Moubane Retgen AI [email protected] · 2026
Read FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data by Abdelrrahman Moubane Retgen AI [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-22596.txt.
Introduction
Abdelrrahman Moubane built FOTBCD, a large bench for spotting new and gone buildings in aerial photos. Their 2026 preprint draws on French orthophotos and official map data from IGN France. Past benches hug one city or a small zone; this one spans twenty-eight departments.
Twenty-five departments train the models; three held-out ones test them. Wide ground truth makes scores mean more for real map-keeping. Planners and map firms need fresh building maps at national scale. The data set gives that task a fair, hard test. It is a resource paper, built to be used. Anyone can use it to grade a new method on the same ground.
Catalog record
| CREATOR | Abdelrrahman Moubane Retgen AI [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-22596
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data Abdelrrahman Moubane Retgen AI [email protected] Abstract 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 0.2 m/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. 1 Introduction Building change detection from remote sensing imagery is essential for urban planning, cadastral updates, disaster response, and environmental monitoring. Deep learning has driven significant progress, but model development and evaluation remain constrained by the limitations of existing datasets. Current change detection benchmarks suffer from two critical limitations: Geographic Homogeneity. LEVIR-CD+ [ 1 ] covers 20 regions in Texas, USA, while WHU-CD [ 4 ] consists of a single image pair from Christchurch, New Zealand. Models trained on such geographically constrained data learn region-specific features—particular building styles, urban layouts, vegetation patterns, and imaging conditions—that fail to generalize across geographic domains. Limited Scale and Diversity. Existing datasets contain hundreds to thousands of image pairs, insufficient to capture the full variability of real-world change detection scenarios, including dense urban cores, suburban sprawl, rural villages, industrial zones, and varying terrain. We address these limitations with FOTBCD, a dataset family derived from France’s authoritative geographic databases. In this work, we focus on the public release of FOTBCD-Binary , a large-scale benchmark for binary building change detection: • National Coverage : 28 departments across mainland France (25 for training and 3 held out for evaluation), spanning diverse geographic and urban contexts from Mediterranean coastal regions to rural Atlantic areas. • High Resolution : 0.2 m/pixel aerial orthophotos from IGN’s BD ORTHO database. • Binary Change Annotations : Pixel-wise CHANGE / NO-CHANGE masks suitable for large-scale benchmarking. • Spatially Referenced : All patches are associated with patch-level spatial metadata, including projected-coordinate bounding boxes, enabling GIS integration. • Scale : Approximately 28,000 image pairs yielding over 100,000 training 256x256 patches. In addition, we release FOTBCD-Instances (Research Dataset) , a smaller subset with instance-level polygon annotations distinguishing NEW, DEMOLISHED, and UNCHANGED buildings, illustrating the full annotation schema and enabling instance-level experimentation. Our experiments demonstrate that geographic diversity is a key factor for cross-domain generalization. Models trained on geographically limited datasets (LEVIR-CD+ and WHU-CD) exhibit substantial performance degradation when evaluated on FOTBCD-Binary. In contrast, a model trained on FOTBCD-Binary transfers effectively to WHU-CD and comparably to LEVIR-CD+ in relative terms, despite differences in imagery source and regional characteristics. 2 Related Work 2.1 Change Detection Datasets Table 1 summarizes widely used building change detection datasets. LEVIR-CD [ 1 ] contains 637 pairs of 1024 \times 1024 Google Earth images at 0.5 m resolution, covering approximately 20 regions in Texas, USA. Changes are annotated as binary masks focusing on building construction. The dataset is widely used but geographically homogeneous. WHU-CD [ 4 ] provides a single large pair of aerial images (32,507 \times 15,354 pixels at 0.2 m resolution) from Christchurch, New Zealand, captured before and after the 2011 earthquake. While high-resolution, it represents only one city and a single change event. LEVIR-CD+ extends LEVIR-CD with additional image pairs while maintaining the same geographic scope and annotation format. In contrast, FOTBCD-Binary provides large-scale geographic diversity at the national level, enabling evaluation across a wide range of urban, suburban, and rural environments. In addition, we release FOTBCD-Instances , a smaller subset with instance-level polygon annotations distinguishing NEW, DEMOLISHED, and UNCHANGED buildings, illustrating the full annotation schema. Table 1 : Comparison of building change detection datasets. FOTBCD-Binary provides national-scale geographic diversity, while FOTBCD-Instances illustrates instance-level annotations. Dataset Region Coverage Res. Image pairs Patch size Ann. Class LEVIR-CD+ Texas, USA 20 areas 0.5 m 985 1024 \times 1024 Mask 2 WHU-CD Christchurch, NZ 1 city 0.2 m 1 32,507×15,354 Mask 2 FOTBCD-Binary France 28 depts. 0.2 m 27871 512x512 Mask 2 FOTBCD-Instances France 6 depts. 0.2 m 4000 512x512 Polygon 3 2.2 Geographic Generalization in Remote Sensing Domain shift between training and deployment regions is a well-known challenge in remote sensing [ 6 ] . Variations in building architecture, urban density, land cover, terrain, and imaging conditions often cause models to underperform when applied to geographically distinct areas. Prior work has primarily addressed this problem through domain adaptation and normalization techniques. In contrast, we argue that increasing geographic diversity at the dataset level is a more fundamental and scalable approach to improving cross-region generalization. 3 The FOTBCD Dataset 3.1 Data Sources FOTBCD is derived from two authoritative geospatial databases maintained by the Institut National de l’Information Géographique et Forestière (IGN), France’s national mapping agency. BD ORTHO [ 2 ] provides seamless aerial orthophoto coverage of mainland France at a ground sampling distance of 0.2 m. The dataset consists of RGB imagery acquired through regular national aerial survey campaigns, with update cycles ranging from 3–4 years depending on the region. The imagery is radiometrically corrected and orthorectified to ensure geometric consistency across large areas. BD TOPO [ 3 ] is France’s authoritative topographic vector database, providing a structured 3D description of territorial elements and infrastructures with metric precision. The dataset covers the entire national territory and is exploitable at scales ranging from 1:2,000 to 1:50,000. It includes building footprints and associated semantic attributes, along with hydrographic, transport, administrative, land cover, and regulatory features. Temporal differences between successive BD TOPO snapshots are analyzed and aligned with the corresponding BD ORTHO imagery to identify building-level changes, including new construction , demolition , and unchanged structures. This change inference process forms the basis for both the large-scale binary change annotations released in FOTBCD-Binary and the instance-level polygon annotations provided in FOTBCD-Instances . 3.2 Geographic Coverage FOTBCD-Binary spans 28 departments across mainland France , with 25 departments used for training and 3 geographically disjoint departments held out for validation and testing (Figure 1 ). The split is performed at the department level to ensure zero geographic overlap between training and evaluation data, enabling an unbiased assessment of cross-region generalization. The selected departments were chosen to maximize geographic, environmental, and urban diversity. Together, they cover a wide range of French landscapes, including: • Dense urban and peri-urban regions , encompassing large metropolitan areas and their surroundings • Coastal departments along both the Atlantic Ocean and the Mediterranean Sea • Rural and agricultural areas , including plains and low-density regions • Mountainous and