Building an Open Dataset of Ubiquitous Map Images for Cartographic Research: Practices and Prospects

Jian Yang, Fenli Jia, Yeqiu Xu, Linfang Ding, Nian-Hua LiuAbstracts 2023 · DOI

Introduction

The landscape of cartography has undergone a radical transformation. We have entered the era of "ubiquitous cartography," where map-making is no longer the exclusive domain of national mapping agencies like the USGS. Today, maps are generated by a vast array of actors, including journalists, scientists, and social media enthusiasts. While this democratization has led to a wealth of spatial information, it has also created a "Wild West" of cartographic design—a diverse ecosystem of images with non-uniform designs, varying themes, and inconsistent quality.

The Cartographic Gap in Computer Vision

Current computational methods, particularly in the realm of computer vision, have become adept at "topic mining." An AI can easily identify that a map depicts a typhoon or a political boundary. However, these tools generally fail to analyze the cartographic quality of the image. There is a distinct lack of automated tools capable of assessing whether a map possesses a balanced layout, clear symbology, or a coherent visual hierarchy. To bridge this gap, researchers are developing a standardized, large-scale dataset specifically designed for cartographic research, moving beyond general-purpose computer vision benchmarks.

A Three-Step Framework for Dataset Construction

The development of the Ubiquitous Map Image Dataset follows a rigorous three-step framework:

  1. Feature Identification: The process begins with a comprehensive review of existing map datasets to identify essential metadata. This includes the map's intended use, its primary theme, its source, and any available auxiliary data.
  2. Hierarchical Tagging: Unlike standard computer vision tasks that might label isolated objects like "trees" or "cars," cartographic tagging requires a specialized semantic system. This system must account for the complex relationships between map elements, such as the placement of legends, the accuracy of scale bars, and the overall visual hierarchy that guides a reader's eye.
  3. Data Enrichment: To truly understand map effectiveness, the dataset goes beyond the static image. Researchers are integrating "auxiliary data," most notably eye-tracking data. By recording where humans actually look when interacting with a map, the dataset provides a ground truth for how information is consumed. This is supplemented by online questionnaires to capture subjective user experiences and interpretations.

The Ubiquitous Map Image Dataset

The dataset draws from a diverse range of sources to capture the full spectrum of modern cartography. These include historical topographic maps from the USGS, thematic maps published in scientific journals, real-time weather prediction maps, and creative designs from social media challenges like the "#30DayMapChallenge." By contrasting these with standard computer vision benchmarks, the authors highlight that map images require a unique set of labels focused on design principles rather than simple object presence.

Key Insights: Maps as Non-Structured Media

A fundamental insight of this research is that maps are not merely "pictures" of the ground. Unlike a satellite image, which is a direct sensor-based representation, a map is a highly structured, symbolic, and often artistic representation of space. Treating a map like a standard photograph ignores its inherent design logic.

Furthermore, the project emphasizes the "Human in the Loop." By integrating behavioral data, the researchers argue that an AI should not just identify what is on a map, but understand how a human interprets that information. This addresses the "semantic gap" between robotics and cartography: while a robot sees a "road" as a physical obstacle, a cartographer sees a "road" as a symbol that must be balanced against text and color for maximum legibility.

Preliminary Results and Future Prospects

Initial experiments using the framework have demonstrated its feasibility. The researchers successfully used the dataset for "balance assessment," employing AI to determine if the visual weight of a map's elements is distributed correctly according to design principles. This paves the way for "map image interpretation," where systems can understand the intent behind a design rather than just the pixels.

Looking forward, the goal is to scale the dataset to encompass even more diverse sources. This will allow researchers to revisit fundamental cartographic problems, such as establishing a mathematical definition for a "good" map layout. Ultimately, this work aims to develop automated tools for cartographic design critique, providing creators with real-time feedback based on established design patterns.

Summary updated

BibTeX
@inproceedings{yang2023building,
  author = {Jian Yang and Fenli Jia and Yeqiu Xu and Linfang Ding and Nian-Hua Liu},
  title = {{Building an Open Dataset of Ubiquitous Map Images for Cartographic Research: Practices and Prospects}},
  booktitle = {Abstracts},
  year = {2023},
  doi = {10.5194/ica-abs-6-281-2023},
}
Figure