Value Generalization for Static and Dynamic Choropleth Maps - A new approach considering global autocorrelation and local outliers: Dissertation

For nearly 200 years, choropleth mapping has been used to depict the values of statistical variables for predefined areal units. Although choropleth maps are conceptually less straightforward than laypersons commonly expect, they are prevalent, preferred by users and constructed with technical ease. While static choropleth maps are especially useful to portray the variation of the mapped variable in space, animated choropleth maps have their merits in showing the dynamic change of such spatial patterns. In order to reduce visual complexity and/or emphasize relevant spatial patterns, cartography has established a long tradition of generalizing the mapped values of static and dynamic choropleth maps, either by applying some sort of smoothing procedure or – much more commonly – by reducing the wealth of individual values to a few ordinal classes, with each class depicted by a unique symbol. However, commonly used classification algorithms ignore the spatial contiguity of static- and the spatial and temporal contiguity of dynamic choropleth maps and establish class breaks solely based on the distribution of values along the number line. Together with their strict adherence to crisp class breaks, common classification routines themselves thus introduce “visual noise” in the imposed regionalization, even for data that is highly autocorrelated in space and – in the case of animated time series data – time. This reintroduced fragmentation of the map image due to classification counteracts its purpose of reducing visual complexity.
This thesis tackles this problem and develops a new classification approach that honours the spatial structure of the data. Reasoning that the consideration of space in the classification process only makes sense if space plays a role in the data itself, the influence of the spatial component in the classification adjusts to the degree of spatial autocorrelation found in the data. Applied to positively autocorrelated data, the method leads to a reduction in visual complexity, at the expense of slightly overlapping classes in the value domain. While the latter can be tackled by a specifically developed visualization concept or by referral to somewhat fuzzy but useful “folk” classifications, a major drawback of this initial approach is the contrast reduction or even elimination of local outliers that should be preserved in the final map. Hence, a heuristic approach was developed to identify and preserve local outliers and thus even increase their visual salience against the overall less fragmented map image. Furthermore, the approach was extended to the temporal dimension. It is well known that our cognitive capacity is easily exceeded when watching visually complex map animations of time-series data.
While the first part of this thesis develops the methodology and validates the reduction of map complexity based on well-established spatial and temporal complexity metrics, the second part is dedicated to empirical research. Although the method was initially developed for cartographic classification, it can also be directly applied to unclassed (animated) choropleth maps, smoothing them in space and/or time according to the degree of spatial and/or temporal autocorrelation in the data while preserving local outliers. An extensive empirical study was conducted to determine whether users of unclassed choropleth map animations benefit from spatial, temporal, or spatiotemporal value generalization. While the hypothesized improvement in the detection of general patterns and trends due to value generalization could not be confirmed for the synthetic map stimuli used, results show significant improvements in the detection of local outliers when generalizing in space. Moreover, map users seemingly prefer spatially smoothed map animations. Linking the empirical results to concepts and findings from vision studies allow for plausible interpretation, shedding additional light on our perception of animated choropleth maps.