The research investigates the long-term visual evolution of historical maps, recognizing their significance as strategic technologies and cultural artifacts reflecting societal changes. Existing historical cartography research often focuses on individual maps or prominent series, neglecting the vast majority of maps available. The increasing availability of digitized historical maps creates an opportunity for quantitative analysis, akin to corpus linguistics' impact on language studies. The study aims to develop a methodology for cartographic stylometry—the analysis of maps' distinctive visual appearance—to explore the visual language and cultural evolution reflected in large map corpora. The map, as a socially constructed discourse, is analyzed in its multiple roles: representing geographical space, functioning as a power-knowledge practice (influenced by Foucault), serving as a planning device, and reflecting cultural apprehension of the environment through figurative choices. The authors propose to use computational methods for "distant viewing" of cartography, analyzing formal features in a large corpus to bridge the gap between famous map examples and the myriad maps composing the cultural fabric. The study will focus on identifying characteristic formal features within a map corpus, laying the foundation for a quantitative and corpus-based approach to cartographic history.

The study draws upon corpus linguistics' success in understanding language change. It leverages the concept of "pertinence" from semiotics to identify distinctive features within a cartographic system. Contrastive representation learning, a machine learning strategy, is adapted to learn feature representations of low-level graphical units ('mapels') by contrasting higher-level structures (map series). This approach aims to overcome the subjectivity inherent in manually cataloging map symbols and icons, as exemplified by Dainville's work. While the study acknowledges the need for future research on map semiosis, it uses the concept of "pertinence" as a foundational theoretical concept.

The study uses a corpus of 10,046 digitized French and Swiss maps (1600–1950) from 32 institutions, focusing on European France and Switzerland. Maps are geocoded using metadata and natural language processing tools. Eleven coherent map series are identified, encompassing various scales, times, topics, and geographical subjects. The methodology employs 'fragmentation,' inspired by text tokenization in natural language processing, to divide map images into 50x50 pixel tiles ('mapels'). Non-geographic content and low-graphical-load areas are excluded. Mapels are recentered and reoriented to standardize their visual representation. Thirteen candidate visual features are initially considered, covering color (5 features), morphology (1), texture (1), graphical load (3), line thickness (1), and orientation (2). Instead of a black-box neural network approach, the study optimizes the feature space using a genetic algorithm. The algorithm maximizes the median distances between maps from different series and minimizes those between maps from the same series, selecting the six most pertinent features and their weights. This optimization process uses a notion of "radius of free variation" (k) to define the proximity of mapels within a map and between maps in a series. A stylometric distance (D) is introduced to quantify figurative proximity between maps. The resulting feature space is used to group mapels into 'mapotypes' through an iterative space subdivision process. For visualization, mapotypes are projected onto a 2D space using t-SNE and displayed as a 'mapotypic mosaic'. Analysis then involves stratifying the corpus temporally and scalewise, observing the distribution of mapotypes across strata and employing statistical tests (Shapiro-Wilk and Regional Kendall) to identify significant trends.

The optimization process successfully differentiates between map series, with the exception of the nearly indistinguishable Napoleonic cadastral maps, validating the hypotheses. The resulting six-dimensional feature space prioritizes morphology, followed by graphical load, line thickness, and texture; color distribution is not a significant factor. The analysis reveals several key findings: 1. **Abstraction Process (17th-18th Centuries):** A gradual shift from iconographic to symbolic map representation, evidenced by the increasing use of abstract textures and symbols. 2. **Scale-Based Scission (19th Century):** A significant divergence in the visual style of small- and large-scale maps, beginning around the turn of the 19th century. 3. **Line Refinement and Load Increase (19th Century):** A peak in the production of finer lines in the first half of the 19th century, followed by an increase in graphical load and map complexity. 4. **Color Prevalence (Late 19th-20th Centuries):** The widespread adoption of color printing in the later 19th and early 20th centuries, reflecting technological advancements. 5. **Technological and Cultural Influences:** The findings correlate with technological innovations (lithography, mechanical rulings) and cultural shifts (rise of nation-states, tourism, and aviation). Statistical tests such as the Shapiro-Wilk test for normality (measuring the characteristic figuration in time slices) and the Regional Kendall test (for monotonic trends in macroscopic variables) support these observations. Analysis shows a high correlation between mapotype distribution and time periods, revealing distinctive figurative characteristics for each period. The study also reveals a shift in the relationship between scale and figuration over time. The initial figurative similarities between different scales eventually diverge significantly in the 19th century.

The findings address the research question by demonstrating the effectiveness of the fragment-based approach for analyzing the long-term visual evolution of historical maps. The results align with existing knowledge of cartographic history, confirming the approach's validity. The identification of a persistent split between small- and large-scale maps after the 19th century is a particularly significant contribution. The findings support the study's initial hypothesis. The observed changes in map features reflect underlying technological, administrative, and cultural influences. The study's approach opens up new avenues for exploring cartographic history by moving beyond the analysis of individual maps to examining patterns and trends within large map corpora. The study's focus on quantitative methods complements qualitative interpretations, enhancing the understanding of cartographic development.

This study introduces a novel fragment-based methodology for cartographic stylometry, providing insights into the long-term visual evolution of maps. The approach effectively distinguishes map series and reveals key historical shifts and trends. Future research could explore the integration of semantics into the model, extending the methodology to other corpora, and investigating more nuanced cultural mechanisms reflected in map design.

The study's corpus, while extensive, is limited to French and Swiss maps, potentially affecting the generalizability of findings. The selection of candidate features and the optimization process may have influenced the results, and further exploration of other features could provide additional insights. The interpretation of mapotypes relies on visual inspection and lacks a formal framework for establishing direct correlations between visual style and semantic meaning.