Exploratory Data Analysis and map animation: Using temporal brushing and focusing to facilitate learning about global weather  

Introduction

Our research has two primary goals. The first is to integrate two exploratory data analysis methods (brushing and focusing) with map animation to produce a manipulable dynamic representation that facilitate conceptualization of time as both linear and cyclic. The second goal is to explore the impact of different representations and controls of those representations on the way in which students conceptualize the spatial and temporal aspects of multivariate continuously changing phenomena (specifically weather and climate). To meet these goals, we have built a geovisualization environment (the GlobalWeatherAnalyzer) that facilitates examination of three aspects of global weather (land temperature, ocean temperature, and cloud cover) as they relate to one another in both time and space.

Before describing the GlobalWeatherAnalyzer, a discussion of relevant literature is needed to place our work in the broader context of recent developments in geovisualization generally and spatiotemporal visualization more specifically. This review will touch upon four themes: cartographic representations of change, map animation, integration of exploratory data analysis (EDA) and cartographic methods, and learning from maps.

Cartographic representations of change

The representation of change has been a frequent topic of attention in cartography. Szegö (1987), for example, devoted much of Human Cartography to this topic, integrating a perspective of time geography derived from Hägerstrand (Szegö cites (Hägerstrand 1974)) with cartographic representation methods for static display. Szegö's approach emphasizes a categorization of "cartographic" events that are "acted out" on the "stage" of the map. This metaphor underlies much of the map animation work to date (see below).

In an effort to develop a systematic approach to the representation ofchange on maps, Monmonier (1990) categorizes methods for depicting time on static maps as falling in three types: dance maps (single maps that depict change in an object's location over a time interval), change maps (single maps that depict change in a location's attributes during a time interval), and chess maps (small multiples that depict change in the state of things from one time to the next). Kraak and MacEachren (1994) extend this categorization to include dynamic representations of space-time phenomena and Edsall, et al (1997) propose a categorization of changes in space-time phenomena along a continuum ranging from change in location (but not attributes) to change in attributes (but not location). Given that time can be conceptualized as both linear and cyclic, cartographic representation of time is problematic (see, for example, Szegö 1987; MacEachren 1995; Edsall, Kraak et al. 1997). Using the static "small multiple" technique (see Bertin 1983; Tufte 1983) both conceptualizations can be incorporated in the same visual display, for example, by using an organizations in which column positions of the small multiple matrix depict identical points in the cycle and the sequence across the entire time can be seen by scanning across each row in turn.

Common to all of the above is a goal of understanding how variations in what is being represented might be matched with variations in the form of possible representations. This approach can lead to guidelines for selecting appropriate representation forms for particular data and to hypotheses about the implications of the different representation forms for understanding of spatiotemporal phenomena.

Map animation

Cartographers have been interested in the potential of map animation as a way to depict changing phenomena for approximately four decades (see Thrower, 1959 for one of the first accounts of cartographic animation and Campbell and Egbert, 1990 for a review of early map animation research). Considerable research has also been directed to the application of animation methods to non-temporal data (see, for example: Moellering, 1980; DiBiase, et al, 1991; MacEachren, 1994; and Peterson, 1995). Here we restrict our attention to research on animation directed to temporal data.

Much of the early work in map animation conceptualized time as linear and focused on techniques for generating map movies to represent change over time (e.g. Thrower 1961; Tobler 1970; Gould, DiBiase et al. 1990; Dorling and Openshaw 1992). Within this research stream, attention has been directed to (1) appropriate metaphors for representing change (Gersmehl, (1990), for example discussed the relative merits of flipbook and stage and play metaphors); (2) choice of symbolization methods for depicting changing phenomena with different spatial or attribute characteristics (MacEachren and DiBiase 1991); (3) design of graphic scripts that organize blocks of time into a coherent story (Monmonier 1990); (4) uses of animation in television, such as weather broadcasts (Carter 1996); and (5) issues of temporal interpolation needed to generate smooth animations from temporally sparse data (Acevedo and Masuoka 1997).

The (potential) power of animated over static maps is their ability to prompt a conceptualization of temporal continuity, thus facilitating an understanding of process rather than state. As Monmonier and Gluck (1994) found, however, viewers are often frustrated by complex changing maps that they cannot control. This is echoed by Koussoulakou and Kraak (1992) who found no difference in effectiveness between static maps and animated maps, when the animated maps could not be controlled. The rapid advance in computer technology over the past decade, together with evidence that users want and need to control aspects of the animation has prompted a move away from map movies that a viewer simply watches to manipulable animated displays. Manipulation can take many form, from direction and pace controls (Who?? 1994) to an ability to zoom to a different scale of analysis (Buttenfield and Weber 1994).

Although map animation research (until recently) emphasized a linear perspective on time, more than two decades ago Moellering (1976) demonstrated the power of map animation to facilitate understanding of cyclic space time phenomena with an animation of traffic accidents. As part of his research, Moellering did generate the typical time series animation in which each frame was one hour, with the resulting movie depicting the location and time of accidents across X days in Detroit. However, this typical animation was matched with a complementary representation of "collapsed time." In this movie, accidents happening at each hour were aggregated through time resulting in a map animation depicting a composite day (all accidents that happen at 6am followed by all at 7am, etc.). The result is a clear characterization of the daily cycle of accidents with peaks during rush hour periods and troughs between these times.

Integration of EDA with cartographic representation

An important component of the overall research thrust associated with geovisualization has been the integration of EDA methods into map based analysis environments. Several graphical EDA methods, including linked brushing (Carr, Littlefield et al. 1987; Monmonier 1989), focusing (MacEachren, Howard et al. 1993), and the "grand tour" (Monmonier 1992) have been adapted and extended for application with georeferenced information. The integration of statistical and cartographic approaches to data exploration has resulted in several innovative approaches to visualization of spatial data. For example, the concepts of focusing has been utilized in several ways. One involves application of map animation methods to produce what has been termed "sequencing," where the map display sequentially focuses on subsets of the data range from low to high (or high to low) value (Slocum 1988). Peterson (1995) proposed sequencing, not through data values or classes, but through levels of aggregation, from an unclassed map on which each data value is signified with a different shade of gray to a two-class map in which high values are dark and low values light. Monmonier (1992) combined these ideas into a hierarchical sequence that focuses on increasingly broad subsets of data with each pass -- allowing the analyst to notice a range of features from local details through global scale patterns.

EDA methods have also been extended to deal with temporal data. Monmonier (1989) proposed the concept of a "temporal brush" with which an analyst could highlight a section of a timeline (e.g. a two day sub-period), with the result being display (on linked scatterplots and map) of all data representing the time span selected. Brushing, as implemented in EDA applications, typically allows an analyst to select individual specific entities to highlight, while focusing allows analysts to select a value range within which all included entities will be highlighted. Monmonier's "brushing," thus, corresponds more closely to the EDA method of "focusing" -- since it involves highlighting a value range from a continuous vector (the timeline). In his Atlas Touring implementation, Monmonier combined this method with sequencing to produce a "graphic phrase" through which the observer sees a sequence through time aggregated by year (the "brush" on the timeline is one year wide), followed by one aggregated across five years (a wider brush), to ones aggregated across 10, then 20 years (Monmonier 1992). Focusing directed to attribute values rather than to time periods has also been applied to spatiotemporal data. MacEachren, et al (1997), for example, combine focusing with animation so that an analyst can focus on a particular subset of data (e.g. the top 20% of the data range) then observe changes in the spatial location of that data subset over time by running a time series animation in which the initial focusing threshold is retained for each time. Brushing (in the EDA sense of selecting a potentially non-contiguous set of entities that will be highlighted) has been applied to the temporal component of data through manipulable legends that allow analysts to select specific times to be highlighted--at varied levels of aggregation, such as by hour or day (Edsall and Peuquet 1996).

Learning and problem solving with maps

Learning and problem solving with maps has been researched by cartographers, psychologists, educators, and others. Issues relevant to the current project include: the cognitive prerequisites for learning and problem solving with maps (e.g. Meyer 1973; Downs and Siegel 1981; Winn 1987; Liben and Downs 1989; Liben and Yekel in press); memory for mapped information (e.g., Howard and Kerst 1981; Eastman 1985; Gilhooly, Wood et al. 1988; Mersey 1990); strategies for successful learning and problem solving (e.g. Thorndyke and Statz 1980; Eastman and Castner 1983; Crampton 1992); and the implications of representational choices in the learning and problem solving process (e.g. Shimron 1978; Kulhavy and Schwartz 1980; Yarnal and C ????1982; Peterson 1985).

While there is a substantial body of research on these topics, limited attention has been given thus far to the potential impacts of dynamic maps on learning and problem solving. For learning from maps, MacEachren (1992) demonstrated that a developmentally-based segmentation strategy emphasizing routes makes acquisition of knowledge of city structure easier (in which a map is learned through sequential presentation of a hierarchically organized set of subcomponents). Slocum and Egbert (1993), in contrast, found that sequencing of choropleth maps did not help or hinder map learning but did increase (rather than decrease) response times for tasks based on the learned map. For temporal information, however, Koussoulakou and Kraak (1992) found that an animated display resulted in faster response times to spatiotemporal questions than did a static maps. It appears that maps that change in response to data (but are not user controllable) can result in more complete learning, leading to more accurate or faster response to questions that depend upon information derived from the maps, but that the success of these non-interactive dynamic displays depends on the kind of information to be learned and/or the design of the displays.

Most research to date dealing with interactivity has focused on manipulable static maps. McGuinness (1994), for example, conducted one of the first studies dealing with strategies for problem solving with dynamic systems. She compared novice and expert users of geographic information as they used a GIS to address a pair of typical spatial analysis problems dealing with multivariate data. Clear differences in the strategies employed by experts and novice were found, with the experts being more selective and systematic in their use of map layers--generally combining fewer layers at once but being more likely to consider all combinations of variables. Recently, Evans (1997) examined expert and novice use of an environment that was both interactive and dynamic. The environment was designed to facilitate consideration of both data and data reliability in a problem solving situation focused on use of classified remote sensing images. She found that, when users are given control of which representation form to view and for how long, that both static bivariate maps of data and reliability and dynamic bivariate maps (that flicker between a data and a reliability representation) were effective and that experts spent more time exploring data but were no more accurate in problem solving than novices.

Edsall, et al. (1997) examined the impact of three styles legend types of animated maps of spatiotemporal data. Two versions provided manipulable legends that allowed users to quickly jump to different points in time, one in which time was treated as linear, then other with time treated as cyclic. No significant differences in accuracy of response to simple information retrieval questions was found, but individuals clearly used very different strategies for exploring the information, suggesting temporal maps may benefit by using multiple legends styles. Some made extensive use of the time selection controls while others spent more time viewing repeated animations through the time sequence.

GlobalWeatherAnalyzer (GWA) design and implementation

In accordance with the larger goals of the Visualizing Earth project, our geovisualization system is designed exploit the educational power of remote-sensing and Geographic Information Systems as a way for students to investigate our planet, to learn about the geosciences, and to develop skills of geographic visualization, problem-solving, and investigative learning. More specifically, we are interested in implementing and assessing the utility of temporal brushing and temporal focusing tools in a geovisualization environment, and understanding what differences these tools make in developing an understanding of the global climate system. Although the target group of the Visualizing Earth project is students, these tools have an applicability in systems designed for experts and novices alike.

GlobalWeatherAnalyzer (GWA): Conceptual Level Goals

In accordance with the larger goals of the Visualizing Earth project, our geovisualization system is designed exploit the educational power of remote-sensing and Geographic Information Systems as a way for students to investigate our planet, to learn about the geosciences, and to develop skills of geographic visualization, problem-solving, and investigative learning. More specifically, we are interested in implementing and assessing the utility of temporal brushing and temporal focusing tools in a geovisualization environment, and understanding what differences these tools make in developing an understanding of the global climate system. Although the target group of the Visualizing Earth project is students, these tools have an applicability in systems designed for experts and novices alike.

Temporal Brushing and Temporal Focusing (EDA methods): Operational Level Goals

The GlobalWeatherAnalyzer is designed to be a interactive, exploratory visualization environment. Apart from standard interface tools (i.e. direction and speed controls), users have access to temporal focusing and temporal brushing tools. The former is used to adjust the start and end dates of the animation, while the latter is used to select what times of day are included in the animation. Linked brushing is a powerful tool that has been discussed in other contexts (see Monmonier 1989) and its extension into the temporal realm allows the user to recognize important periods during temporal cycles. For example, if one is interested in changes in daily maximum temperatures over a one week period, the ability to suppress the dominant (and potentially overwhelming) diurnal temperature cycle is extremely useful. Using temporal brushing, the Rossby-wave induced pattern of mid- and high-latitude land temperatures is easily identifiable. Without temporal brushing, these large scale wave-like spatio-temporal structures are lost, or at least, difficult to identify. As visualization environments become increasingly data-rich, the ability to "filter" data using brushing and focusing is increasingly helpful.

Brushing reinforces the notion that that time is cyclic, while focusing reinforces the notion that time is linear. Building on previous research (Edsall et al 1997) we included two types of temporal legends: cyclic (denotes time of day) and linear (denotes day of week). They are complimentary and facilitate the successful use of the brushing and linking tools.

Three data layers are included in the GlobalWeatherAnalyzer (GWA): land temperature, sea-surface temperature (SST), and cloud cover, although other data could be easily incorporated into the system. These data have a temporal resolution of 6 hours, a spatial resolution of approximately 50 kilometres, and show a period of one week (February 10 to 16, 1998). Data are stored as layers in the system which users can turn on and off to create visual overlays. Because each layer is semi-transparent, multiple layers can be visualized at once. This is an important feature of our system because it allows the relationship between the different layers, or climate phenomenon, to be investigated and understood. For example, given that land temperature and cloud cover are related variables, the GlobalWeatherAnalyzer can be used by students to understand their spatial association. Moreover, because this is spatio-temporal data, students could look for a possible lag period in this relationship.

Implementation Strategy

The GlobalWeatherAnalyzer was built using Macromedia's Director 6.0 animation package. Director has become an industry standard in the area of animation and its scripting language, Lingo, provides considerable flexibility in the designing interactive environments. The popularity of both Director and Macromedia's proprietary Shockwave Internet-based technology were other key factors in this decision. Lastly, although vector-based animation tools are starting to be developed (i.e. Macromedia's recently released Flash 3) none offer the power or flexibility that our task demanded.

Creating an intuitive and transparent user interface is always a design challenge. This is compounded by the fact that we had to ensure the system would succeed with novice users working with low-end computer displays. Due to the latter consideration, the stage size was set to 640 x 480 pixels, and all the graphic elements were set to 8-bit color (256 colors). Furthermore, the systems was designed to be Internet savvy. By utilizing Macromedia's Shockwave Internet technology, students can use the GlobalWeatherAnalyzer without special software (it "plays" in their web browser). Furthermore, the GlobalWeatherAnalyzer "streams" which allows it to start playing before it has finished downloading. Although the GlobalWeatherAnalyzer is fairly large digital document (an unfortunate consequence of the raster-based Director) "Streaming" is an elegant solution to the ubiquitous bandwidth issue.

Next Steps

1. Focus Groups

The next step in this research is to elicit feedback from "experts" in an informal focus-group session. Group participants will include specialists in the areas of visualization, G.I.S., and cartography. This feedback help us to refine the GlobalWeatherAnalyzer before commencing with human subjects testing. Focus-groups have proven highly effective at the design stage (Monmonier and Gluck 1994) and pilot tests can help refine the research questions and agenda before costly testing commences (Mersey 1990).

2. Evaluating the impact of EDA methods on exploration strategies and understanding

Later this year, we will test the GlobalWeatherAnalyzer is a formal human subjects experiment. Given our research focus on students and education, subjects will be drawn from the undergraduate population at Penn State University, who will be tested en masse in a large computer teaching room (100+ terminals). We are interested in better understanding how temporal brushing and focusing can help students learn, and how students differ in their development of the underlying skills of spatial, logical and symbolic thinking.

[ABSTRACT]

[PAPER]

[REFERENCES]

[ACKNOWLEDGEMENTS]

[ANIMATIONS]

RETURN MAIN PAGE