HisMap: A Multi-Temporal Historical Topographic Map Dataset

HisMap is a benchmark dataset of scanned historical German topographic maps annotated for semantic segmentation of land-cover notions. It is built from official topographic map sheets covering two regions of Germany — the Hameln area (Lower Saxony) and Bavaria (Bayern) — each surveyed and published at several points in time between 1898 and 2023. Because the same geographic area is available across many map editions and cartographic styles, HisMap supports research on cross-domain / domain-generalizing segmentation, few-shot and zero-shot learning, retrieval-augmented segmentation, and long-term land-cover change analysis.

At a glance

• Total tiles: 17,676 RGB image tiles with 17,676 matching label masks
• Tile size: 384 × 384 pixels, PNG
• Map editions (region–sheet–year folders): 25 (17 Hameln, 8 Bavaria)
• Approximate size on disk: 2.9 GB

Semantic classes

Pixel-level labels are stored as 8-bit paletted PNGs with the following class-index mapping:

• 0 – background, RGB (0, 0, 0) — all editions
• 1 – forest, RGB (34, 139, 34) — all editions
• 2 – grass, RGB (124, 252, 0) — all editions
• 3 – settlement, RGB (220, 20, 60) — Hameln editions only
• 4 – water, RGB (30, 144, 255) — all editions

Each map-edition folder contains a color_map.json stating exactly which notions are annotated for that edition. The Bavaria sheets do not annotate the settlement class; there the label indices are shifted so classes stay contiguous (forest = 1, grass = 2, water = 3). A nonlabeled category (white, (255, 255, 255)) is reserved for regions without a valid annotation.

Regions and temporal coverage

Folder names follow the pattern region.sheet_year, where sheet is the official topographic map sheet number and year is the survey/publication year of that edition.

Hameln (Lower Saxony) — 17 editions across 4 sheets:

• Sheet 3821: 1974, 2002, 2012, 2023
• Sheet 3822: 1975, 2005, 2023
• Sheet 3921: 1973, 2002, 2017, 2023
• Sheet 3922: 1898, 1974, 1982, 1996, 2005, 2017

Bavaria (Bayern) — 8 editions across 6 sheets:

• Sheet 489: 1910, 1939
• Sheet 490: 1910, 1938
• Sheet 518: 1941
• Sheet 519: 1910
• Sheet 7230: 1959
• Sheet 7330: 1959

Having several years per sheet means the same spatial extent is available in multiple cartographic epochs and drawing styles, enabling multi-temporal change detection and temporal domain-shift studies in addition to single-image segmentation.

Directory structure

Each of the 25 map editions is a folder named <region>.<sheet>_<year>/ containing:

• imgs/ — RGB map tiles, 384×384 PNG, named <x>_<y>.png by pixel offset within the full sheet
• lbls/ — paletted semantic label masks, same filenames as imgs/
• color_map.json — class name → RGB color for this edition
• <edition>.json / <edition>.txt — per-tile class pixel-fraction table and full tile-ID list
• train_samples.txt, test_samples.txt — base train/test split
• train_{10,50,100}shot.txt, val_{10,50,100}shot.txt — few-shot training/validation splits
• cls_selection_{10,50,100}.json — class-balanced sample selections

Intended uses

• Semantic segmentation of land cover in scanned historical maps
• Cross-domain and domain-generalization benchmarking (across regions, sheets, epochs, and styles)
• Few-shot and zero-shot segmentation
• Long-term land-cover and settlement change analysis from cartographic archives

Data provenance and processing

The images are crops of digitized official topographic map sheets of the Hameln region (Lower Saxony) and Bavaria. Each sheet was cut into 384×384 tiles, and land-cover notions (forest, grass, settlement, water) were rasterized into paletted label masks aligned pixel-for-pixel with the image tiles. Each edition is accompanied by per-tile class statistics, predefined few-shot splits, class-balanced selections, and precomputed retrieval indexes and support–query pairings.

Contact: Yunshuang Yuan (yuanyunshuang@gmail.com)