Deep Learning Techniques for Historical Map Analysis

Deep learning approaches have transformed the automated interpretation of historical cartographic records, enabling researchers to extract, classify and compare geographical features at unprecedented scale and accuracy. Historical maps vary widely in age, scale, symbology and preservation quality, presenting challenges such as faded ink, distortions, non-standardised colour palettes and shifts in cartographic conventions. Convolutional neural networks and related architectures now underpin semantic segmentation, object detection and feature extraction pipelines tailored to these idiosyncrasies. Self-supervised and weakly supervised methods harness spatial co-occurrence and ancillary modern geospatial data to overcome limited annotations, while domain adaptation techniques ensure models generalise across map series from different eras or regions. Fragment-based stylometry further treats map imagery as a visual language, quantifying stylistic evolution over centuries. Collectively, these advances facilitate large-scale studies of urban expansion, landscape transformation and cultural change, enriching our understanding of human–environment interactions and informing disciplines from historical geography to heritage conservation.

Research from Nature Portfolio

A fragment-based methodology has been proposed to analyse the long-term visual evolution of cartographic styles. By partitioning each map into elementary fragments and computing an interpretable feature representation—encompassing texture, morphology and graphical load—researchers distinguish between coherent map series and track stylistic transitions from the 17th to the 19th century. This approach reveals a steady abstraction trend in early modern maps, a divergence between small- and large-scale mappings in the industrial era, and broader cultural fashion cycles in line weight and symbol density. The fully interpretable framework opens new avenues for exploratory research on visual language systems in large historical map corpora.

Deep Learning Techniques for Historical Map Analysis publication trend

The graph below shows the total number of articles in deep learning techniques for historical map analysis across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional neural network: A class of deep learning model composed of convolutional layers that automatically learn spatial hierarchies of features from input images, widely used for image recognition and segmentation tasks.

Semantic segmentation: The process of assigning a class label to every pixel in an image, enabling precise delineation of map elements such as roads, water bodies and urban areas.

Domain adaptation: Techniques that adjust a model trained on one dataset (source) to perform effectively on a different but related dataset (target), addressing variations in style, scale or quality.

Cartographic stylometry: An analytical approach treating map imagery as a visual language system, where statistical and visual-feature analyses reveal patterns of stylistic and cultural evolution across map corpora.

References

1. Domain adaptation in segmenting historical maps: A weakly supervised approach through spatial co-occurrence. ISPRS Journal of Photogrammetry and Remote Sensing (2023).
2. Dataset of building locations in Poland in the 1970s and 1980s. Scientific Data (2024).
3. A fragment-based approach for computing the long-term visual evolution of historical maps. Humanities and Social Sciences Communications (2024).
4. Automatic Road Extraction from Historical Maps Using Deep Learning Techniques: A Regional Case Study of Turkey in a German World War II Map. ISPRS International Journal of Geo-Information (2021).