Three interrelated representational issues are relevant to the project described here: the application of animation to depiction of temporal data, the role of interactive controls for data exploration by experts, and methods for symbolizing data uncertainty or reliability.
Animation
Much of the recent literature on both scientific and cartographic visualization promotes animation as a valuable tool for both the "visual thinking" and "visual communication" ends of the research sequence (DiBiase, 1990; Peterson, 1995). As noted by several authors, animation has the advantage of representing the dynamic elements of a process, as well as more closely approximating the real world process itself (DiBiase, et al., 1991; Dorling and Openshaw, 1992; Kraak and MacEachren, 1994). Another contribution of animation is its ability to show the correlations and relationships of many phenomena (Karl, 1992; DiBiase, et al., 1994). This ability is grounded in application of one or more "dynamic" variables. Three variables were initially identified by DiBiase, et al. (1992), duration, rate of change, and order. MacEachren (1995) subsequently added three additional dynamic variables to this set, display date ("date" in animation time, typically specified as the frame number), frequency (the logical dual of duration), and synchronization (applicable only when two or more features are being represented).
Atmospheric phenomena in particular can be represented effectively using the dynamic variables. The complexity and fluidity of meteorological and climatological processes along with their continuity across space and time lend themselves to the increased flexibility and power of animation (Koussoulakou, 1994). In the context of climate models, DiBiase, et al. (1992) demonstrated that use of the dynamic variable order can be a particularly powerful tool in understanding model uncertainty as it relates to agriculture. They generated an animation in which precipitation predictions of five climate models were depicted along with an indication of prediction uncertainty (as assessed by dispersion around the mean prediction). When the animation was "played" in order of increasing uncertainty (rather than in temporal order), the earth system scientists for whom the animation was created noticed (for the first time) that model uncertainty was greatest in the spring planting season precisely the time at which certain predictions are most important.
Interactivity
Like animation, the ability to interactively manipulate map parameters (whether static or animated) and to interactively query the database behind the map offers the potential for fundamental changes in how maps are used in scientific thinking (MacEachren and Ganter, 1990). A considerable body of research directed to interactive data exploration has been conducted in the field of exploratory data analysis (EDA) (see, for example, Cleveland and McGill, 1988; Buja, et al., 1991). Several EDA methods have been successfully extended to exploratory spatial data analysis. For example, Monmonier (1989) has extended the concept of scatterplot "brushing" into the spatial domain with his "geographic brushing" and MacDougall (1992) has combined the techniques of "brushing" and "linking" to develop a dynamic variant on cluster analysis. A related EDA technique, "focusing" (a method in which subsets of data are highlighted) has been implemented for exploring data depicting environmental change and metadata representing the certainty of spatial and temporal interpolation of these data (MacEachren, et al., 1993).
Some effort has also been directed to combining animation and interactive control in the exploration of spatial (and spatio-temporal) data. For non-temporal spatial data, the addition of interactive controls to an animated "flyby" facilitates the use of that animation as a tool for exploring the character of terrain (Moellering, 1980) as well as the relationships such as those between terrain and vegetation (Kraak, 1994). Dorling (1992) has applied similar methods (that he terms animating space) to exploring maps of enumerated social and economic data. Egbert and Slocum (1992) have implemented related interactive controls in a tool designed to allow exploration of choropleth maps. In their case, animation was used not to shift the viewer's focus from one place to another, but from one category of attribute to another, through a sequencing of information by data category. For georeferenced temporal data, Monmonier (1990) added the concept of temporal brushing to his earlier geographic brushing. With hypermedia environments (e.g., MacroMedia Director or animation tools for the World Wide Web, such as Sparkle), at least one aspect of temporal brushing has become common place the ability to change the pace of an animation and to move forward and backward, frame-by-frame or at any speed the user desires.
One of the more innovative combinations of interactivity and animation is that associated with Buttenfield and Weber's (1994) concept of "proactive graphics for exploratory visualization." They define proactive graphics as those in which "..., users initiate queries and steer data presentation in a manner consistent with the associative power of the human intellect" (Buttenfield and Weber, 1994, p. 8). To illustrate their concept of proactive graphics, they developed a prototype that allows an analyst to interactively control an animated representation of tree growth over time. The prototype actually includes many individual animations (of maps and graphs at different scales of analysis) that play out simultaneously (in the background) with the user being able to change scale, location, or representation form of the visible display (from map to graph, to diagram) as time progresses.
Uncertainty
With the proliferation of digital data, metadata (or data about data) has become increasingly important. Uncertainty is a key facet of that metadata. To represent uncertainty it must be defined, both generally and specifically. A step toward such a definition is a categorization of components. These include data quality and variability in relation to spatial, attribute, and temporal aspects of the data (MacEachren, 1992; MacEachren, 1994). As part of the U. S. Spatial Data Transfer Standard (SDTS) data quality is further subdivided into positional accuracy, attribute accuracy, logical consistency, completeness, and lineage (Fegeas, et al., 1992). Variability can arise from spatial, attribute, and/or temporal aggregation in which a single characterization is used to represent a group of phenomena that may not be homogeneous (MacEachren, 1992). Variability introduced through aggregation is, of course, directly related to the resolution of that aggregation (Buttenfield, 1993). With the general components of uncertainty identified, the next step in the production of a representation of uncertainty is to determine which aspects are relevant to a specific instance. Once this decision is made, estimates of uncertainty must be matched to appropriate visualization tools and methods.
Visual representation of uncertainty is receiving increasing attention (e.g., Beard and Buttenfield, 1991; MacEachren, 1992; Fisher, 1994; Buttenfield and Beard, 1994; van der Wel, 1994), but remains an area of georeferenced visualization with many unanswered questions. Two major issues can be identified. One is that of symbolization. MacEachren (1992) and others (e.g., McGranaghan, 1993) have examined Bertin's graphic variables for use in representing uncertainty and have added new variables, notably saturation (i.e., purity) of color and clarity. The latter can be further broken down into crispness, resolution, and transparency MacEachren (1995). The second major issue in representing uncertainty is that of display types or user interfaces. These include single bivariate maps, map pairs or multiple maps, sequential presentation, and interactive displays (MacEachren, 1992). Each of these has their own advantages and disadvantages. Some are obviously not suited for animated maps, others are impossible to implement with static maps.
Little attention has been directed to representation of uncertainty over a time series (see MacEachren, et al., 1993 and Mitasova, 1993 for initial attempts). With an animated map, showing phenomena that change over time and uncertainty that also varies with time the choice of display type is limited. Map pairs are often impractical, due to the limited screen size of computers. Sequential presentation and toggling between maps of data and maps of uncertainty would probably be confusing for time series data. Bivariate and multivariate maps seem most suitable for this type of representation. These methods have the advantage of having everything on one map thus allowing the largest possible display given screen size. In addition, the map can be shown as a smooth animation in which viewers can observe the change in uncertainty at the same time they are studying the data. Multivariate maps can, however, easily become complex and confusing (McGranaghan, 1993) and, as Buttenfield (1993) has found, can require viewer training. One goal of our prototype is to determine whether animated bivariate data/metadata maps are practical.