Effects of Spatial Distribution on Change Detection in Animated Choropleth Maps

Abstract

To address unsolved issues of change detection in animated choropleth maps, we proposed the concept of 'gross change detection' and performed an experiment that empirically verifies the incidence of change blindness stems from the 'magnitude of change (MOC)', spatial distribution in animated choropleth maps. We generated experimental materials using the change-characterization arrays and the global Moran's I. Participants had 108 cases of changing maps with time duration (1 to 3 sec) and had questions. The results showed that MOC and duration affect gross change detection, but the most interesting result from our experiment was that different spatial distributions between two adjacent choropleth maps may lead the map reader to under- or over-estimate the level of gross change in the map. It implies that we should consider spatial distribution of change when we design animated choropleth maps.

1. Introduction

Change blindness is a phenomenon that occurs when a viewer failed to detect large changes in a visual display. This indicates that limited human cognitive capacity affects how well they detect changes between scenes in dynamic visual environment (e.g. dynamic displays, or real world). In cartography, these cognitive issues for dynamic visual stimuli are applicable to animated maps, because portray change over time and space (DiBiase et al., 1992; Harrower, 2007).

Perceptual-cognitive approaches to cartography have improved the effectiveness of map designs over the past years (MacEachren, 1995; Harrower, 2007). However, the design of dynamic maps, interactive maps, and threedimensional displays still need more guidelines and empirical evidence to prove its effectiveness (Slocum et al., 2004; Andrienko et al., 2010). In response to this necessity, recent research projects on animated maps have conducted various experiments using survey-based experiments and eye tracking experiments, with the adoption of methods and theories from psychology and vision studies (Montello, 2002; Andrienko et al., 2010). These empirical studies have addressed a wide range of research topics, including comparisons of static small-multiple maps and animated maps (Kossoulakou and Kraak, 1992; Slocum et al., 2004; Griffin et al., 2006), comparisons between visual variables’ effectiveness in animated maps or interactive maps (Cinnamon, 2009; Çöltekin et al., 2009; Garlandini and Fabrikant, 2009; Hegarty et al., 2010), and the impact of interactivity (Mayer, 2001; Mayer and Chandler, 2001) and dynamic variables such as rate of change and abrupt or smooth transitions (Goldsberry and Battersby, 2009; Fish, 2010). Some studies have concluded that dynamic animated maps that are too complex leads to difficulties in conveying information (Bétrancourt and Tversky, 2000; Morrison and Tversky, 2001; Goldsberry and Battersby, 2009), because of human cognitive limits such as split attention, retroactive inhibition, and cognitive overload (Harrower, 2007). Therefore, it is essential to consider how dynamic variables affect the acquisition of map information. The design of animated maps should well capture spatial and temporal traits of change from an original phenomenon, with maximizing map reader’s capability to perceive the animated maps.

With these needs, recent researches on map designs have dealt with perceptual and cognitive issues of dynamic or interactive maps including animated maps. The purpose of this study is to obtain empirical evidence to verify the effect of certain conditions on a map reader’s perception, with adoption of methods and theories from psychology and vision studies. We tried to figure out: perceptive effectiveness at the level of gross change detection, spatial distributions of change in relation to change blindness, and duration in relation to change blindness.

In order to investigate the effect of spatial distribution on gross change detection, we have (1) specified the level of gross change detection in a similar way to Goldsberry and Battersby (2009), (2) designed a cognitive psychological experiment that allowed us to examine effects of ‘spatial distributions of change’, ‘MOC’, and ‘duration’ on gross change detection, (3) produced experimental materials by introducing the change-characterization arrays and global Moran’s I, in order to control MOC and spatial distribution of change in our visual stimuli, and (4) finally conducted an actual experiment.

 

2. Four Major Concepts for Experiment

This paper focuses on unsolved issues of change detection in animated choropleth maps as an extension of Goldsberry and Battersby (2009), and Fish’s (2010) studies. Goldsberry and Battersby (2009) examined limitations of the human visual system for animated map reading and introduced methods to quantify the dynamic variable ‘magnitude of change (MOC)’, the amount of change between adjacent scenes in animated choropleth maps. This means that we can investigate how change detection abilities vary with MOC. To quantify MOC, we used change-characterization arrays that counted the number of enumeration units per transition in animated maps. Fish’s (2010) thesis empirically demonstrated that the type of fill transition used to show change in an enumeration unit between two map scenes affects the amount of change detection. Fish (2010) employed changecharacterization arrays to produce experimental materials and adopted the concept of the level of change detection proposed by Goldsberry and Battersby (2009). In addition, we paid attention to spatial distribution of change and duration, proposed by DiBiase et al. (1992), Peterson (1995), and MacEachren (1995) as well as MOC. While MOC describes how many enumeration units change their fill appearances in animated choropleth maps, it does not measure where those changes are located in the map and finally, change in overall spatial distributions.

2.1 Level of gross change detection

In the animated choropleth maps, ‘gross change’ is as important as change in individual enumeration units. The sum total of values in all enumeration units usually varies with scene transitions and detecting gross change accurately is the key to understanding the overall patterns of dynamic phenomena over time. Therefore, unlike Fish (2010), this study focused on ‘gross change detection’ and defined ‘gross change detection’ as detecting change in the sum of all enumeration unit values. To link ‘gross change detection’ to change blindness, we operationally defined ‘change blindness’ as incidence of incorrect answers to the question of whether gross change occurred. We used two questions about whether he or she was able to detect a change in the map and whether he or she was able to recall the origin state of the highlighted unit prior to the scene change by selecting “high”, “medium”, or “low” from answer choices.

2.2 Magnitude of change

According to Goldsberry and Battersby (2009), changecharacterization arrays can effi ciently summarize changes in each enumeration unit’s fill appearances in animated choropleth maps. More specifi cally, the array for a pair of three-class choropleth maps is conceptualized as Fig. 1. It has nine potential transition behaviors as elements of the array, depending on attributes in the origin and destination state. While three diagonal elements in the array involve persisting transition behaviors, the six off-diagonal elements involve shifting transition behaviors.

The simplest method for quantifying MOC is the basic magnitude of change (BMOC) across the map, which is equal to the sum of all values of off-diagonal elements in the array. In other words, BMOC is the total number of enumeration units that shift during a transition, and normalized BMOC is the proportion of shifting units with respect to the total number of enumeration units, ranging from 0 (no shifting units) to 1 (no persisting units). Another method for quantifying MOC is the magnitude of rank change (MORC), which considers rank distance that is the difference in class rank between the origin state and the destination state. In Fig. 1, while the rank distance between white and grey is 1, the rank distance between white and black is 2. Like normalized BMOC, normalized MORC is the proportion of MORC with respect to the total number of enumeration units. By using a change-characterization array, we controlled both BMOC and gross change to create experimental materials. In Fig. 2, the origin of the 2nd, 3rd, and 4th grids is the fi rst grid. Each grid is based on different change-characterization arrays. While each array has been manipulated to have stable gross change (2nd grid), increased gross change (3rd grid), or decreased gross change (4th grid), all grids have same BMOC of 24. In the same way, with a BMOC of 12, we controlled gross change in our experiment.

Fig. 1.A change-characterization array for a pair of three-class choropleth maps (Goldsberry and Battersby, 2009)

Fig. 2.Relationships between scene transitions and change-characterization arrays

2.3 Spatial distribution of change

We can control MOC and gross change by manipulating a change-characterization array. Unfortunately, starting from one origin map, one change-characterization array can generate large numbers of possible destination maps with many different spatial distributions. The number of possible destination maps varies proportion to the number of enumeration units. In other words, the destination map is likely to have different spatial contiguity from the origin map. By spatial contiguity, we means where spatially adjacent enumeration units with the same class rank are in the choropleth maps.

Spatial distribution of change might affects the perception of animated choropleth maps in several ways: cognitive overload, split attention effect, and foveation distance. Cognitive load theory argues that humans have limited long-term memory and working memory to process and memorize information. This means that too many dynamic changes might result in cognitive overload. Second, the split attention effect (Sweller et al., 1998; Mayer, 2001; Harrower, 2007) is defined as “any impairment in learning that occurs when a learner must mentally integrate disparate sources of information” (Mayer, 2001). According to the definition, in a choropleth map, calculating the sum total of spatially separated enumeration units’ values with the same class rank requires integration of individual units’ value. In other words, in order to calculate the sum total, clustered units require less time and efforts, but dispersed or random units require more time and effort, on account of spatial contiguity effect. Finally, changes in peripheral enumeration units may be less likely to be noticed, due to limited foveation distance (Goldsberry and Battersby, 2009).

Consequently, we created destination map scenes with artificial spatial distributions from the given origin map scene based on consistent change-characterization arrays. We used global Moran’s I (Moran, 1948) for controlling the spatial distribution of choropleth maps. Then, we classified the acquired map scenes into three groups, according to three distribution types, which are ‘clustered (C)’, ‘dispersed (D)’, and ‘random (R)’. The specification of each distribution type is based on the global Moran’s I statistic (Moran, 1948) which measures spatial autocorrelation.

2.4 Scene duration

We considered the dynamic variable ‘duration’ as an independent variable. Duration is an important conditions that affects perceptual and cognitive processes. To be concrete, too short a duration may make it difficult for map readers to detecting changes. We set our duration time as 1, 2, and 3 sec for the experiment. We referred to Griffin et al. (2006)’s duration time, which was 5, 7, 9, and 11 sec for 6 scenes.

 

3. Experiment

3.1 Experimental design

In order to examine change detection in animated choropleth maps, we generated 108 cases, which consisted of pairs of original map scenes called ‘map scene 1’ and destination map scenes called, ‘map scene 2’. In respective cases, we set different experimental conditions. Each condition of experiment was as follows:

(1)State of gross change: No Change, Increased, Decreased (2)BMOC (normalized BMOC): 12 (28.57%), 24 (57.14%) (3)Spatial distribution of change: Map scene 1 (C, D, R), Map scene 2 (C, D, R) (4)Scene duration: 1 sec, 2 sec, 3 sec

The combination of BMOC and state of gross change generated four conditions (BMOC 12_No change, BMOC 24_No change, BMOC 24_Increased, BMOC 24_ Decreased). For spatial distribution of change, there were three conditions in map scene 1 and map scene 2 (Clustered, Dispersed and Random for each scene: CC, CD, CR, DC, DD, DR, RC, RD, RR). Additionally, there were three scene duration conditions (1, 2, 3 sec). One case consists of five slides. The first and second slides showed map scene 1 and 2. The third and fourth slides were question slides. The last slide was an intermediate slide for preparing for the next case. We measured the gross change detection of the first and second level via questions 1 and 2. Test questions were as follows:

Q1. Do you think that the spatial distribution had changed between scene 1 and scene 2?

Q2. If yes, do you think it increased or decreased?

Modified Georgia county map was used for creating maps. We extracted 51 counties from 154 counties and merged some counties with smaller areas to reduce the total number of enumeration unit to 42. There are two reasons to use a modified Georgia county map. First, participants were not familiar with this region. We tried to reduce background knowledge of the study area. Second, we modified the Georgia county map because we needed relatively uniform counties to avoid effects caused by mistake from the size and shape of the region. To create the identical maps, python scripts was used to calculate with two kinds of matrix with BMOC values (12 and 24). Actual values and maps are shown in Fig. 3. We made ‘cluster’, ‘disperse’, ‘random’ maps using Moran’s I index that permutated 1000 times to get a critical value. The critical value was +0.35 for cluster, from 0.01 to -0.01 for random, and -0.35 for disperse (Fig. 3).

Fig. 3.Examples of experimental materials according to experimental conditions

3.2 Experimental procedure and participants

After instruction, participants filled the blank of survey sheet. Participants performed both practice trials and experimental trials. Due to the limits of software and hardware capability, we made presented the stimuli in four groups. Each trial had 27 animations with time duration (1 to 3 sec). A participant looked at animated maps on the monitor. We made buttons on the screen to get an answer. It took about 11 minutes on average for each participant to fi nish the experiment actually.

Eighteen students were participated in this experiment. Participants were undergraduate and graduate school students in geography.

 

4. Results

4.1 Magnitude of change and gross change detection

4.1.1 Magnitude of change conditions: BMOC 12 and 24 in no gross change condition

The overall percentages of correct answers and the averages of time taken to answer are shown in Fig. 4. The percentages and averages are aggregated by two conditions (12, 24) of BMOC. Both conditions had no gross change condition. To confirm their statistical significance, oneway ANOVA analysis was used. In each case of scene duration(1, 2, 3 second), BMOC 12’s percentage of correct answer is higher than BMOC 24’s except in the case of 2 second (1 second: F = 5.615, p = .031 < .05; 2 second: F = .976, p = .338 > .05; 3 second: F = 8.611, p = .01 < .05). In the case of BMOC 12, there is no difference in percentage of correct answer among the scene durations (F = .269, p = .764 > .05). Likewise, in the case of BMOC 24, there is no difference in percentage of correct answer among the scene durations (F = 2.542, p = .08 > .05). Thus, MOC can affect to percentage of correct answer rather than scene duration in this experimental condition. In each case of answering duration per scene duration, there is no difference between BMOC 12 and BMOC 24 (1 second: F = .92, p = .352 > .05; 2 second: F = 1.05, p = .321 > .05; 3 second: F = 1.375, p = .258 > .05). But, in each case of BMOC 12 and 24, there were statistical significances among the scene durations with answering durations (BMOC 12: F = 3.027, p = .05 = .05; BMOC 24: F = 16.064, p = .000 < .05). The time taken to answer is the shortest in 3 second of scene duration. The shorter scene duration, the more time was taken to answer.

Fig. 4.Percentage of correct answer (left y axis, %) and average task duration (right y axis, sec.) of answering with no gross change

4.1.2 Gross change conditions: no change, increase and decrease

The overall percentages of correct answers and the averages of time taken to answer for question 1 and 2 are shown in Fig. 5. In this fi gure, BMOC persisted as 24 and gross change conditions including ‘no change’, ‘increased’ or ‘decreased’ are various in each case. The percentages of correct answers to no change, increased and decreased conditions do not show clear difference between different scene durations (no change: F = .79, p = .465 > .05; increased: F = .049, p = .952 > .05; decreased: F = .058, p = .944 > .05). The time taken to answer is longer to detect gross change in 1 sec duration than 2 or 3 sec, with respect to the question 1 (answering duration of Q1 : F = 3.481, p = .048 < .05). Moreover, the more time subjects consume to answer to the question 1, the less time they do to the question 2 (1 second: F = 5.415, p = .011 < .05; 2 second: F = 5.337, p = .012 < .05; 3 second: F = 7.84, p = .002 < .05). With this result, we presume that while answering the question 1, subjects recall the tendency of gross change simultaneously.

Fig. 5.Percentage of correct answer and the average task duration of Q1 and Q2 with no change, increased and decreased gross change

4.2 Spatial distribution of change and gross change detection

We rearranged distinguishable patterns to explain the relative relationships between spatial distributions of change conditions and gross change detection. We standardized the percent of correct answer and answering duration in each case of scene duration. Scatterplots are shown in Fig. 6. The X axis shows the percentage of correct answers and the Y axis shows the time to answer. The origin of X, Y axis means the average of the percentage of correct answer and answering duration. Thus, there are four cases by quadrant. Each cases are:

Case 1 (quadrant 1): Positive percentage of correct answer and Positive answering duration. This case is inefficient to deliver the map information Case 2 (quadrant 2): Negative percentage of correct answer and Positive answering duration. In this case, people have difficulty to recognize the map information. Case 3 (quadrant 3): Negative percentage of correct answer and Negative answering duration. This case is the most likely to deliver the incorrect map information Case 4 (quadrant 4): Positive percentage of correct answer and Negative answering duration. This case is the most likely to be cognized the map information

Fig. 6.Scatterplot by standardized the percent of correct answer and answering duration

In 3 second scene duration, the BMOC 24 increased case shows the best condition. There are seven spatial distributions (CC, CD, CR, DR, RC, RD, RR) in the case 4. DD and DC are in the case 1 and the case 3 each. Dramatically, there is no case 2. The BMOC 12 no change case, the BOMC 24 decreased case and BMOC 24 no change case shows the best condition in 2 second scene duration. In the BMOC 12 no change case, CC, DR, RC and RD are in the case 4. CR and DD are in the case 3. CD and DC are in the case 2, and RR is in the case 1. The BMOC 24 decreased case, CD, RD and RR are in the case 4. CR, DD and DR are in the case 3. DC and RC are in the case 2, and CC is in the case 1. The BMOC 24 no change case, DC, DR and RD are in the case 4. CC, CD and DD are in the case 3. CR and RR are in the case 2, and RC is in the case 1. In 1 second scene duration, there is no remarkable case. When the gross change condition was ‘increased’, CC, DC, and DC were the worst conditions in each second. Namely, the participants get wrong answers when map scene 2 has clustered distribution. It means that even though the gross change was actually increased, they did not notice. When the gross change was ‘decreased’, CR was the worst case in 1, 2, 3 seconds. Namely, the participants get wrong answers even though the actual gross change was decreased.

 

5. Conclusion

Regarding the magnitude of change in relation to change blindness, we can deduce the following implications from results of each condition of gross change.

The higher MOC, the lower gross change detection. In no gross change condition, the longer duration, the shorter time taken to answer. In gross change condition, scene duration 1to 3 sec. does not greatly affect the effectiveness of animated choropleth maps. In gross change condition, increased and decreased cases are higher percentage of correct answers than no change.

Concerning the spatial distributions of change in relation to change blindness, we can conclude as follows.

The higher MOC, the more people tend to overestimate the sum of enumeration units’ values. When map scene 2 has clustered spatial distribution condition, people are likely to underestimate the sum of enumeration units’ values. We presume that if spatial distribution condition moves from dispersed or random to clustered, people are distracted from gross change detection, due to split attention effect. When map scene 2 has dispersed or random spatial distribution condition, people are likely to overestimate the sum of enumeration units’ values. We presume that if spatial distribution condition moves from clustered to dispersed or random, people are distracted from gross change detection, due to split attention effect. When spatial distributions are dissimilar between adjacent scenes in the animated choropleth maps, map readers are less likely to detect gross change. When original map scene has clustered spatial distribution, more attention is needed in animated map reading.

In conclusion, change blindness is more likely to happen when two adjacent scenes in animated choropleth maps have dissimilar spatial distributions each other, and less likely to happen in the opposite situation. In addition, change blindness is more likely to happen when animated choropleth maps have the higher magnitude of change. Moreover, the impact of duration on the incidence of change blindness varies with spatial distribution conditions. When it comes to level of gross change detection, perceptive effectiveness is higher at the lower level of gross change detection than at the higher level.

In this study, we identified issues: perceptive effectiveness at each level of gross change detection, spatial distributions of change in relation to change blindness, and duration in relation to change blindness. We suggested empirical evidence to prove the effect of each condition of experiment variables on animated map reading. But it is clear that more studies are needed, particularly about scene duration, Magnitude of Change and enumeration unit number, map classifications and attention. We believe that by conducting these series of experiments, we can enhance our understanding of the animated choropleth maps for more effective and efficient design.