Organizations: Institute of Cartography and Geoinformation, ETH Zurich, Switzerland · CRH, École des hautes études en sciences sociales (EHESS), France · Photogrammetry and Remote Sensing, ETH Zurich, Switzerland · LASTIG, Univ Gustave Eiffel, IGN, Géodata Paris, France
Prior to modern Earth observation technologies, historical maps provide a unique record of long-term urban transformation and offer a lens on the evolving identity of cities. However, extracting consistent and fine-grained change information from historical map series remains challenging due to spatial misalignment, cartographic variation, and degrading document quality, limiting most analyses to small-scale or qualitative approaches. We propose a deep-learning-based framework for fine-grained urban change analysis from large collections of historical maps, built on a modular design that integrates spatial map alignment, multi-temporal object extraction, and change estimation. This framework shifts the analysis of historical maps from ad hoc visual comparison toward systematic, quantitative characterization of urban change. Experiments demonstrate the robust performance of the proposed alignment and object extraction methods. Applied to Paris between 1868 and 1937, the framework reveals the spatial and temporal heterogeneity in urban transformation, suggesting its potential relevance for research in the social sciences and humanities. The modular design of our framework provides flexibility for future adaptation to different cartographic contexts and downstream applications.
Figures & tables
Figure 1: Historical maps violate assumptions underlying standard change-detection and image-alignment methods. (a) Two satellite images [ 19 ] with a time difference of 5-10 years exhibit stable object locations across years, allowing simple overlays to reveal changes. (b) Historical maps [ 91 , 92 ] from editions a decade apart display obvious object misalignment: object locations shift across sheets (yellow boxes at mosaic boundaries), making direct comparison unreliable. In addition, cartographers frequently adjust text labels for layout or legibility (yellow arrows), introducing displacements unrelated to real-world change. (c) Video frames [ 36 ] contain smooth, coherent motions where dense-matching algorithms are typically used to align frames, track objects, and estimate per-pixel motion trajectories. (d) Conversely, historical maps exhibit substantial content changes (yellow boxes) [ 92 , 93 ] . These drastic changes, together with deliberate cartographic layout adjustments, undermine the assumptions of dense-matching algorithms, i.e., stable object correspondences for accurate image alignment.
Figure 2: Method overview. (a) Workflow for historical map alignment and change analysis. Maps covering the same location are first rectified to a common image plane. For map pairs from different editions (years), we estimate a dense displacement field for pixel-level alignment and extract urban instances for each edition. The displacement field is used to align and match instances across years, enabling fine-grained instance-level change analysis. (b) Dense displacement estimation without ground-truth correspondences. A dense-matching network is trained using self-supervision from synthetic transformations with known displacements, and refined with a triplet cycle-consistency constraint linking real and synthetic views. (c) Instance extraction with temporal knowledge. A shared encoder–decoder integrates information from multiple years via temporal feature fusion to predict instance segmentation for the target year. The patch-wise predictions are stitched to produce a complete instance map. (d) Instance-level changes are aggregated into urban blocks, a fundamental spatial unit in urban morphology and planning, to derive block-level change profiles. These profiles enable planning-relevant interpretation of long-term urban morphological change.
Image
Object
Disp.
Temp. Cons.
Method
SSIM ↑
CD(10%) ↓
CD(20%) ↓
CD(40%) ↓
mV ↓
L1 ↓
Training Area
Unaligned
0.72
5.50
3.96
1.89
–
–
GLU-Net
0.78
2.57
1.73
0.82
0.66
7.16
WarpC
0.83
1.47
0.95
0.44
0.29
3.87
Ours
0.83
1.44
0.93
0.43
0.23
2.83
Table 1: Model performance for displacement estimation in both training areas (15 districts) and unseen testing areas (5 districts). We measure the CD by removing the 10%, 20%, and 40% largest distances. Smoothness is evaluated by mV, and temporal consistency by L1 within triplets from consecutive years. Arrows indicate the direction of better performance. The best values are marked in bold.
Figure 3: A qualitative comparison of image alignment across different methods. In (a-c), all methods effectively align images despite large displacements between the originals. Our method shows robustness to text displacements in (d) and (e), as well as to object changes in (f-h), maintaining stability even in extreme cases, as seen in (e).
Figure 4: Average precision (AP) scores for objects categorized by area: small (area ≤500 m 2 ), medium ( 500 m 2< area ≤2000 m 2 ), large (area >2000 m 2 ), and non-small (area >500 m 2 ). We compare single-map segmentation (Mask2Former as the baseline) with our proposed multi-map fusion strategy. For each box plot, n=12 test areas were used to derive the displayed statistics, with each test area contributing one AP value. The test areas span two years, and each contains 400–3000 object instances. Box plots show the median as the center line, with the lower and upper bounds of the boxes corresponding to the 25th and 75th percentiles, respectively. Whiskers extend to the smallest and largest observations within 1.5 times the interquartile range (IQR) from the lower and upper quartiles. Observations beyond the whiskers are shown individually as outliers. Our model improves overall accuracy and reduces the IQR and number of outliers.
Figure 5: Qualitative results of our instance extraction method. Individual instances are shown in distinct colors. Our model performs comparably or slightly better to the baseline Mask2Former in (a) building blocks and sidewalks, (b) intricate building structures, and (c) railway neighborhoods, while demonstrating clearly improved performance in detecting faded features such as buildings and gardens (d,e).
Figure 6: Temporal change patterns of Paris from 1868–1937. Each map covers 20 arrondissements (see (g)), subdivided into four quartiers each, for a total of 80 quartiers. Each map displays the average Intersection over Union (IoU) values per building block, calculated from the overlaps of all aligned instances within that block across two consecutive years, serving as a geometric proxy for changes—lower values indicate larger changes potentially. (h) highlights areas in the outskirts associated with rail freight (pink) and waterborne freight (blue), as well as areas that remained relatively stable (yellow). (i) highlights changed areas in central Paris, including hospitals (yellow), parks (green), railway stations (pink), world exposition sites (orange), and other public buildings (blue).
Figure 7: Major types of changed areas and additional applications of building block representations. Maps are sourced from [ 90 , 91 , 93 , 94 ] . Three main types of change are identified: (a) restructuring of building blocks driven by street openings and widenings; (b) infrastructure-related areas exhibiting internal modifications; and (c) apparent changes caused by false detections, typically due to construction lines indicating planned roads (blue arrow). Beyond change characterization, building block representations enable further analyses, such as street network evolution. We illustrate this by visualizing the betweenness centrality of intersections, where nodes in red correspond to the top 10% most globally connected nodes.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Supplementary Fig. 1 : Detailed illustration of triplet cycle consistency and self-supervised data synthesis used for displacement estimation. Images are synthesized with known displacements, with random text and object injections to train the network to ignore these disturbances.
Supplementary Fig. 2 : For displacement estimation, we use the GLU-Net architecture [ 83 ] . The paired images I , J are first down-sampled. Both original and down-sampled images are fed into feature extractors, resulting in multi-level features. Local and global correlations between the source and target features are computed at multiple levels l∈1,2,3,4 . Using these correlations, multi-level displacement fields are iteratively estimated by refining the displacement from the previous level with the current correlation information. The estimated displacement field from the final layer is up-sampled to produce the final estimation.
Supplementary Fig. 3 : Our proposed architecture for instance segmentation. We extend the Mask2former model by introducing multi-scale feature fusion blocks to integrate features from maps at different years, based on the cross-attention mechanism. Note that we apply the estimated displacement fields to spatially align the map editions across years beforehand.
Supplementary Fig. 4 : A qualitative comparison of the estimated displacement field tested on maps of the same area from the years 1905 and 1937. The results, shown from top to bottom, include GluNet, WarpC, and our method. Our model outperforms the others by generating a smoother displacement field and effectively revealing the upward-right shift and the paper-folding artifacts.
Image
Object
Disp.
Temp. Cons.
Method
SSIM ↑
CD(10%) ↓
CD(20%) ↓
CD(40%) ↓
mV ↓
L1 ↓
Unaligned
0.50
1.54
0.97
0.45
–
–
GLU-Net
0.58
1.91
1.10
0.49
0.25
4.15
WarpC
0.60
1.64
0.88
0.32
0.19
1.92
Ours
0.61
1.65
0.88
0.31
0.17
1.79
Appendix
Supplementary Tab. 1 : Model performance for displacement estimation on Siegfried maps . Smoothness is evaluated by mV, and temporal consistency by L1 within triplets from consecutive years. Arrows indicate the direction of better performance
Method
AP_avg
AP_s
AP_m
AP_l
Mask2Former (single-map)
0.87
0.27
0.92
0.87
Ours (multi-map fusion)
0.86
0.30
0.87
0.85
Appendix
Supplementary Tab. 2 : Average precision score for different objects categorized by their areas: small (area ≤ 80 m 2 ), medium (80 m 2 <area<300 m 2 ), large (area>300 m 2 ). We compare using single maps (Mast2former as the baseline) and using our proposed multi-map fusion strategy.
Supplementary Fig. 10 : Misalignment (in pixels) between maps over time. For each region, we compute the median of estimated per-pixel displacement, and the boxplot summarizes their distribution across all regions. This illustrates the potential of displacement estimation for analyzing map distortions and the evolution of cartographic techniques.
Supplementary Fig. 11 : Example of propagated labels from 1925 to 1905, shown in edge maps [ 95 ] . Top: the original edge map from 1925, which exhibits clear misalignment with the map from 1905. Bottom: the edge map propagated using the estimated displacement field from 1925 to 1905. The propagated labels align closely with the target map, illustrating the ability of the proposed method to transfer annotations accurately across decades despite substantial urban change and cartographic variation.
Supplementary Fig. 12 : Temporal sensitivity analysis. We evaluate alignment performance on 12 map-edition pairs with different temporal gaps, grouped into smaller-gap (<15 years) and larger-gap ( ≥ 15 years) subsets with 7 and 5 pairs, respectively.
Historical map collections are highly diverse in style, scale, and geographic focus, often consisting of many single-sheet documents. Yet most work in map recognition focuses on specialist models tailored to homogeneous map series. In contrast, this article aims to develop generalizable semantic segmentation models and ontology. First, we introduce Semap, a new open benchmark dataset comprising 1,439 manually annotated patches designed to reflect the variety of historical map documents. Second, we present a segmentation framework that combines procedural data synthesis with multiscale integration to improve robustness and transferability. This framework achieves state-of-the-art performance on both the HCMSSD and Semap datasets, showing that a diversity-driven approach to map recognition is not only viable but also beneficial. The results indicate that segmentation performance remains largely stable across map collections, scales, geographic regions, and publication contexts. By proposing benchmark datasets and methods for the generic segmentation of historical maps, this work opens the way to integrating the long tail of cartographic archives to historical geographic studies.
Remi Petitpierre
EPFL, Swiss Federal Institute of Technology in Lausanne. Digital Humanities Laboratory (DHLAB), School of Computer and Communication Sciences (IC). CM 1 468, Station 10, 1015 Lausanne, Switzerland. · EPFL. Laboratory of Urban Sociology (LASUR), School of Architecture, Civil and Environmental Engineering (ENAC).
Change detection and scene recognition techniques have been widely applied to Street View Imagery (SVI) to understand changes in scenes across the years. However, metadata alone is often insufficient to reliably find visually aligned image pairs. This study introduces the PairWise image finder, a tool that integrates feature detection and matching, supported by semantic segmentation masks to quantify the visual alignment of two images of varying time periods. The tool outputs the share of matched key features, the matched feature distance and coverage, and the alignment of semantic masks, which enables the user to filter image pairs depending on the alignment quality and use case. The visually aligned pairs derived from the tool can be used to accurately study explicit longitudinal change and help reduce manual effort for perception studies. The usability of the tool is demonstrated through a comparison of longitudinal changes, highlighting the importance of perspective when quantifying changes. The proposed method provides a scalable and open tool for researchers and stakeholders to find high-quality image pairs for urban analysis, perception and related applications.
Jussi Torkko
Digital Geography Lab, Department of Geosciences and Geography, University of Helsinki, PO Box 64 (Gustaf Hällströmin katu 2), 00014, Helsinki, Finland
Geospatial representation learning from satellite imagery is a fundamental problem for large-scale urban analysis and real-world applications. Despite recent advances, current methods struggle with cross-region generalization and semantic interpretability due to their reliance on region-specific auxiliary data and the neglect of semantic alignment within multi-temporal urban imagery. Therefore, we present CoST, a novel Contrastive-based Spatial-Temporal framework that aligns spatial context with multi-temporal semantics to extract universal geographic regularities shared across regions. Specifically, CoST explicitly models spatial correlations to capture transferable geographic structures and exploits multi-year urban change semantics to align learned representations with high-level geo-semantics. Extensive experiments demonstrate that CoST consistently achieves superior performance across various downstream tasks and in unseen scenario, yielding an average relative gain of 8.7% over the strongest competing methods across eight city-indicator settings. The code is available in this repo.
Yutian Jiang, Jiabo Liu, Xixuan Hao +1
The Hong Kong University of Science and Technology (Guangzhou)