Generalisation and Multiple Representation

Production, Generalization, and Conflation

6704.3 - Towards assessing generalization quality with visual complexity measures

Thursday, July 6
4:50 PM - 5:10 PM
Location: Virginia B

Generalization is perhaps the most persistent challenge in cartography, and with the advent of digital maps, its relevance has only increased. As we zoom in and out, some objects are removed; others are enlarged, aggregated or displaced, and all objects are simplified (Ruas, 2008). Traditionally, generalization has been manually performed and visually assessed (Bard, 2004). Visual assessment of the generalization results by experienced cartographers is still probably the best mechanism for detecting errors. Only a trained eye can detect whether important features are missing, whether shapes retained their characteristics, or if an unwanted fusion of two distinct objects occurred. However, with the automatization and on-the-fly generalization of big spatial databases, visually assessing every map has simply become unrealistic due to the amount of materials to be evaluated. Thus, there is a need for an efficient, computer based, approach for quality assessment. In this study, we assess the suitability of two algorithmic approaches in determining some aspects of generalization quality, specifically to measure the preserved information after generalization. The metrics we feature in this paper are Feature Congestion (FC) and Subband Entropy (SE), proposed by Rozenholtz et al. (2007). Both the FC and SE are based on psycho-physiological principles of visual attention, thus we hypothesized that they would overall correspond with cartographers’ visual assessment of generalization quality.
To test our hypothesis, we first conducted a computational experiment to measure the visual complexity of 12 online maps in 16 zoom levels (5 to 20) from various map providers (Bing Maps, Bing Hybrid, ESRI Topo, HERE WeGo, Google maps, Google Terrain, Google Hybrid, OSM Hike and Bike, OSM Road, OSM Topo, Google satellite and ESRI satellite). Then, we conducted a user experiment, in which we asked cartographers to evaluate the amount of information on subset of the studied maps (zoom levels 6, 10, 14, 16 and 18).
Overall, we found a moderate positive correlation between the cartographers’ ratings and visual complexity metrics (cartographers-FC: rs= 0.33, p < .001; r = 0.33, p < .001; cartographers-SE: rs= 0.38, p < .001; r = 0.41, p < .001). These results suggest that the cartographers’ judgment can be substituted with the FC and SE metrics only to a small degree. Surprisingly, cartographers and algorithms have agreed the most in the assessment of Bing Maps (cartographers-FC: rs= 0.61, p < .001; r = 0.62, p < .001; cartographers-SE: rs= 0.58, p < .001; r = 0.53, p < .001), while the weakest agreement was for Google Road maps, which was also most preferred by cartographers (cartographers-FC: rs= 0.20, p < .001; r = 0.21, p < .001; cartographers-SE: rs= 0.21, p < .001; r = 0.23, p < .001). These results open another question: Do the FC/SE work better for certain maps, or do cartographers overlook some issues on their favorite maps?
Furthermore, we studied the interactions between zoom levels, cartographers’ assessment and algorithmic visual complexity. At zoom levels 14, 16 and 18, cartographers’ assessment corresponds to the algorithmic visual complexity: Stimuli with the lowest FC/SE were rated 1 (too little information), while those with the highest FC/SE were rated 5 (too much information). Importantly, cartographers’ assessments constantly increased as the FC/SE increased.
There are still some unanswered questions about the suitability of FC or SE to assess the quality of cartographic generalization, and more studies are needed. However we believe that our research contributes to an overall understanding of how visual complexity measures can reflect the quality of geographical data generalization and visualization.

Alzbeta Brychtova

UX Designer
University of Zurich / Lufthansa Systems

Alžběta Brychtová is currently an UX designer and cartographic visualization expert in Lufthansa Systems since 2016. Before, she was a postdoctoral researcher with the Geographic Information Visualization and Analysis group of the GIScience Center of the University of Zurich.
She completed her PhD in Geoinformatics and Cartography at the Department of Geoinformatics, Faculty of Science, Palacký University Olomouc in Czech Republic. During her PhD studies she was a visiting researcher at the ETH Zurich, University of Zurich (multiple times), and University of St Andrews (UK). Her primary research interests are in cognitive and usability issues in geovisualizations.


Send Email for Alzbeta Brychtova

Arzu Çöltekin

Group Leader

Dr. Arzu Çöltekin is a Research Group Leader and a Senior Lecturer at the GIScience Center of the University of Zurich and a research affiliate at the Seamless Astronomy group (which specializes in data science and scientific visualization) at Harvard University. Her interdisciplinary work covers topics related to GIScience, visualization, vision (perception and cognition), eye tracking, virtual environments, and human-computer interaction. She is an active member of several international commissions and working groups, specifically with the ICA and ISPRS, and chairs the ISPRS working group Geovisualization and Virtual Reality.


Send Email for Arzu Çöltekin

Nicolas Regnauld

Product Manager

Nicolas Regnauld has many years of experience in researching solutions for automating generalisation. After a PhD on generalisation at IGN (Paris), and a post Doctorate at the University of Edinburgh, he led the research team on automated generalisation at Ordnance Survey for 11 years. During this time, his team successfully developed a fully automated generalisation process that was used for the creation of OS VectorMap District. Nicolas is now Product Manager at 1Spatial, responsible for 1Generalise and 1Publish. These two products form the core of 1Spatial map derivation solutions, for deriving high quality maps from large geospatial databases.


Send Email for Nicolas Regnauld


6704.3 - Towards assessing generalization quality with visual complexity measures


Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Towards assessing generalization quality with visual complexity measures