Petitpierre, Remi · di Lenardo, Isabella · Hudson, Polly · et al.
4 files · 1.4 GB · geopackage, zipdeclared
The dataset provides a large-scale vector layer of historical building footprints for the Greater London area at the end of the 19th century. It comprises 1,299,029 individual building footprints, extracted automatically from the Ordnance Survey five-feet-to-the-mile (1:1,056) map. The original maps were surveyed between 1891 and 1895 and published between 1893 and 1896, covering approximately 450 km² of urban and suburban London. The dataset was produced using a deep learning-based semantic segmentation pipeline, achieving approximately 97% precision and 95% recall in building detection. Source The source maps were digitised by the National Library of Scotland at 400 ppi and manually georeferenced in collaboration with the David Rumsey Map Collection. The extraction relies on 753 georeferenced map sheets, provided that six sheets are missing from the original archive. Extraction methodology The dataset was generated through an automated workflow: 262 image patches were annotated manually with three classes: (i) regular buildings; (ii) compound buildings; (iii) building boundaries A Mask2Former [1] semantic segmentation model is used. For training and inference, we followed the approach described by [2]. At inference time, boundary predictions are supplemented by the predictions of a specialist model, trained on cadastral plans [3]. Polygon geometries are extracted, based on predicted building contours, and vectorised. Compound buildings components are merged. Content Two versions of the dataset are provided: london_buildings_1891-96_raw.gpkg Raw output from the automated extraction pipeline (after vectorisation and merging) london_buildings_1891-96_corr_v1.gpkg Minimally corrected version including manual adjustments for major structures (e.g. railway stations, monuments, large buildings) Each geometry has a field buil_class , taking values in ['regular','compound'] , relative to the cartographic representation of building classes and the associated processing approach. In corr_v1, manually added geometries take the value NULL . In addition, we relase the labeled images used for training the segmentation model: annotations.zip ZIP folders containing the training and validation data The archive contains subfolders images , containing the tif image samples and labels , corresponding to the semantic class indexes, in png format. labelmap.txt details the class index codes. Data format Format: Geopackage Coordinate reference system: WGS 84 / Pseudo-Mercator (EPSG:3857) Temporal coverage Survey period: 1891-1895 Publication period: 1893-1896 Data creation: 2024-2026 Descriptive statistics Number of building footprints: 1,300,831 (raw), 1,299,040 (corr_v1) Coverage area: 450 km² Detection performance (raw): Precision: 97% Recall: 95% Use and reuse potential This dataset supports research in: urban history historical GIS urban morphology economic and social history It is particularly suited for studying long-term urban change and fine-grained spatial patterns in industrial-era cities. Related publication This dataset is described in a data paper submitted to the Journal of Open Humanities Data Corresponding author Remi Petitpierre Email: remi.petitpierre@epfl.ch Funding The research was supported by the College of Humanities at EPFL and the European Union Horizon Europe Programme (Grant No. 101233051). License This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. Citation If you use this dataset, please cite: @misc{london_footprints_petitpierre_2026, author = {Petitpierre, R{\'{e}}mi and di Lenardo, Isabella and Hudson, Polly and Herold Hendrik and McDonough, Katherine and Hecht, Robert and Vaienti, Beatrice and Fleet, Christopher}, title = {{A Layer of Late Victorian London: The Building Footprints from the 1:1,056 Ordnance Survey Map (1891-1896)}}, year = {2026}, publisher = {EPFL}, url = {https://doi.org/10.5281/zenodo.19497434}} Limitations Lower accuracy for very small buildings Occasional segmentation errors in dense or complex areas Minor inconsistencies near map sheet boundaries Liability The authors assume no liability for the use of this dataset. References Cheng, B., Misra, I., Schwing, A. G., Kirillov, A., & Girdhar, R. (2022). Masked-attention Mask Transformer for Universal Image Segmentation . arXiv. https://doi.org/10.48550/arXiv.2112.01527 Petitpierre R. (2026) Generalizable Multiscale Segmentation of Heterogeneous Map Collections . arXiv. https://doi.org/10.48550/arXiv.2603.05037 Petitpierre, R., di Lenardo, I., & Rappo, L. (2024). Revealing the Structure of Land Ownership through the Automatic Vectorisation of Swiss Cadastral Plans . Digital History Switzerland. https://doi.org/10.13140/RG.2.2.26632.33281