Semantic Segmentation Benchmark Dataset for Homogeneous Historical Map Corpora

This dataset comprises 125 test samples derived from a homogeneous corpus of historical urban maps. It is designed to serve as a benchmark for machine learning and computer vision approaches in the context of historical cartography and land-cover classification. The source material consists of 39 historical maps of Berlin, Germany, provided by the State Library of Berlin (Staatsbibliothek zu Berlin). Each map has a scale of 1:4000 and shares a consistent cartographic style, ensuring a high degree of visual and thematic uniformity across the dataset. The dataset distinguishes five land-cover classes: Buildings – including public, private, military, and religious structures Infrastructure – such as bridges, tunnels, streets, and railway tracks Recreational surfaces – including parks, gardens, and cemeteries Sealed surfaces – such as arterial roads, squares, and courtyards Water bodies – comprising rivers, ponds, and fountains The homogeneity in scale and cartographic representation makes this dataset particularly suitable for evaluating semantic segmentation algorithms that rely on consistent visual semantics. For further information, please visit:https://kartographie.staatsbibliothek-berlin.de/https://stabikat.de/Record/1786888610