COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL

DATA

Stefan J

anicke

1,2,3

, Christian Heine

, Ralf Stockmann

and Gerik Scheuermann

Image and Signal Processing Group, Institute for Computer Science, University of Leipzig, Leipzig, Germany

ottingen State and University Library, University of G

ottingen, G

ottingen, Germany

Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS, Sankt Augustin, Germany

Keywords:

Visual Data Exploration, Geovisualization, Comparative Visualization.

Abstract:

The amount of online data annotated with geospatial and temporal metadata has grown rapidly in the recent

years. Social networks like Flickr and Twitter are popular providers of masses of such data, but are hard to

browse. Many systems exist that can show the data’s geospatial distribution in a map widget and their temporal

distribution in a time widget, allowing these widgets to become dynamic-query-like ﬁlters for data. We present

a web application that combines several existing approaches and supports comparison of multiple, potentially

large result sets of textual queries in a geospatial and temporal context with extended interaction capabilities.

We validate our approach with several case studies and our tool’s learnability in a ﬁeld experiment.

1 INTRODUCTION

Although the amount and types of data available

through public web resources is seemingly endless,

the main method of ﬁnding information is still largely

performed by text queries. Because of the typically

enormous amount of results for text queries, popular

search engines rank their results using a blend of rele-

vance and popularity. The results often include items

the user is not interested in, requiring him to reﬁne

the query by adding additional or using different key-

words, but repeated reﬁnement can lead to frustration.

While the amount of data is increasing, they also

become more structured. Metadata as well as struc-

tural relationships between results can give rise to

a contextual overview of the data, allowing to ﬁlter

them. Contextual overview can take many forms: top-

ical, geospatial, and temporal being some of the more

popular. Websites such as Flickr and Twitter provide

rich data sources annotated with geospatial and tem-

poral metadata. Users already familiar with searching

in geographic environments like Google Maps, can

ﬁnd results presented directly on a map rather than in

a list. There are also early prototypes for mapping the

events and characters of cultural heritage onto maps

and interactive timelines (e.g. Timeplot

), driven by

http://simile.mit.edu

enthusiasts using the Google Maps API

or the IN-

SPIRE framework

These enthusiasts approached us to implement a

web application that enables the synergetic explo-

ration of multiple topical queries in a geospatial and

temporal context. They want to be able to juxta-

pose and compare spatial distribution and temporal

trends of multiple queries. We searched for other sys-

tems on which we could base our design. Among

our major inﬂuences are a web visualization of The

Guardian (Rogers, 2010), working with many dat-

apoints, but supporting only one dataset and lim-

ited interaction, GeoVISTA CrimeViz (Roth et al.,

2010), which shows three datasets at the highest

zoom but aggregates them at lower zoom levels, and

VisGets (D

ork et al., 2008), which supports single

datasets with a wide variety of interactions. A gen-

eral concern with these systems, which we selected

as representatives of current systems for geospatial

and temporal exploration, is also the clutter that arises

when many data elements are spatially close together.

This becomes especially harmful when the glyphs’

attributes, e.g. which topical query the items arose

from, differ. Also the time widgets in these systems

can be improved as they typically do not support ﬂex-

ible queries like May 2009 to May 2010.

http://code.google.com/intl/de/apis/maps/

http://www.inspire-geoportal.eu

613

Jänicke S., Heine C., Stockmann R. and Scheuermann G..

COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL DATA.

DOI: 10.5220/0003833406130625

In Proceedings of the International Conference on Computer Graphics Theory and Applications (IVAPP-2012), pages 613-625

ISBN: 978-989-8565-02-0

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

In this paper we present the design of a novel web

application that:

• combines showing and comparing multiple result

sets in a geospatial and temporal context,

• improves geospatial presentation of items by

zoom-dependent aggregation of items that avoids

occlusions,

• allows more ﬂexible selections within the map

and time widget, and

• works client-side to allow user-provided data and

improve response time.

2 RELATED WORK

There has been a lot of research on the usage of visual

interfaces to interact with large data resources. Tukey

illustrates the beneﬁt of integrating interactive visual

tools into the data exploration process to expedite the

undirected search for structures and trends (Tukey,

1977). This so called Visual Data Exploration process

follows the Information Seeking Mantra (Shneider-

man, 1996): Overview ﬁrst, zoom and ﬁlter, details-

on-demand. This means, the ﬁrst presentation of the

information is an overview over the whole data, fol-

lowed by the search for and drill down on interesting

patterns, exploiting capabilities of the human visual

system. The search for a speciﬁc information, which

starts with a vague imagination, and in a stepwise

process leads to understanding is called exploratory

search (Marchionini, 2006).

Much research in the ﬁeld of Visual Data Explo-

ration was dedicated to the analysis of data with ge-

ographical and temporal metadata. An overview over

tools, principles, challenges, and the concept of the

analysis of geotemporal data are given by Andrienko

and Andrienko (Andrienko and Andrienko, 2006; An-

drienko and Andrienko, 2005).

Baldonado (Baldonado et al., 2000) proposed

guidelines for the usage of multiple views in visual-

ization to increase the user’s ability to receive deeper

insights if the data is shown under multiple aspects.

Causal relationships can be found easier as well as

unforeseen connections. Because of space restrictions

we cannot address all but a few representative imple-

mentations that exploit synergetic effects from using

linked views for geospatial-temporal data.

VisGets is a web application that displays the

result of one query in four linked widgets (D

ork

et al., 2008), working much like specialized dynamic

queries (Shneiderman, 1994) for spatial and temporal

ﬁltering. A location widget shows the result items as

small glyphs overlaid on a geographic map. A time

widget shows the distribution of results in bins repre-

senting years, months, and days. A tag widget shows

a limited set of frequent words in alphabetic order and

a size proportional to their importance. A results wid-

get shows small textual or image thumbnails of re-

sults arranged as a table. VisGets offers query reﬁne-

ment in space (selecting one glyph), time (selecting

a year, month or day bar), and by topic (selecting a

tag). Each reﬁnement affects the presentation in all

other widgets. VisGets also provides many mecha-

nisms to interact with the results of a single query, but

a comparison of different queries is not possible.

GeoVISTA CrimeViz (Roth et al., 2010) for crime

analysis enables comparison of different crime inci-

dent types. The map widget shows individual cir-

cles on higher zoom levels. However, on lower zoom

levels, the incidents of all types are aggregated into

hexagonal bins, thereby disabling comparison. The

concept of binning items and plotting their total poses

a problem, since the resulting continuous map over-

lay does not preserve the user’s geospatial orientation.

On the CrimeViz timeline, the number of incidents

per time period of different types are stacked, which

shows the sum of incidents and makes obvious dif-

ferences between the incident types and global trends

visible. Wang et al. (Wang et al., 2008) also uses

a speciﬁc stacked graph – the ThemeRiver (Havre

et al., 2000) – for the visual analysis of global ter-

rorism activities. We avoid stacked graphs, because

we do not want to sum up unrelated categorical data.

Furthermore, it is hard to accurately interpret trends

that lie atop other curves (Heer et al., 2010). LISTA-

Viz (Hardisty and Klippel, 2010) uses the concept

of small multiples (Tufte, 2001) for the analysis of

spatio-temporal autocorrelations on maps and time-

bars to allow comparison. The usage of small mul-

tiples makes comparison of our data, which typically

has a highly irregular spatial distribution, difﬁcult.

Maps can also beneﬁt data of historic origin. Tsi-

pidis et al. (Tsipidis et al., 2011) demonstrate the

usefulness of interactive geospatial-temporal visual-

izations for the exploration of archaeological data.

The HESTIA project (Barker et al., 2010) investi-

gates the differences between imagined geographic

distances and real distances in ancient Mediterranean

space by mapping data onto contemporary maps: the

Herodotus Timemap. It is a geo-temporal visualiza-

tion that shows several location types (settlements,

regions, physical features) mentioned in various text

passages. The user is able to browse through time,

space and book chapters. However, there is neither

aggregation in space nor time, causing visual clut-

ter. The system also lacks ﬁltering of locations by

dynamic reﬁnement in space and time.

IVAPP 2012 - International Conference on Information Visualization Theory and Applications

614

3 DESIGN

In this section, we describe the design of our system,

starting from the input and the ﬁltering methods to the

representation inside multiple linked views and ﬁnish

with the interaction techniques for query reﬁnement.

The presentation consists of a map widget showing

the position of query results as small circles suitably

aggregating items for the current zoom, a time widget

showing the distribution of results in a time span, and

a detail widget showing textual contents or thumb-

nails of data items arranged in a table.

The input to our system is the result of one to

four result sets from a classical keyword or term-

based topical search. Each item in the input sets is

required to have a longitude and latitude as well as a

time stamp or time span annotation. In our datasets

the number of items that lack these data is sufﬁciently

low so that we can simply omit them from the map

and time widget.

The system provides spatial ﬁltering based on

containment queries in the form of circles and poly-

gons as well as temporal ﬁltering based on intersec-

tion queries with a given time span. Initially, the sys-

tem starts with a minimally enclosing rectangle over

all items’ positions and a time span that stretches over

all items’ time stamps and time spans.

We choose colors to codify results from different

datasets because of their effective use to discrimiate

categorial aspects of data (Bertin, 1983). We pre-

sume map colors to be mostly dark, cold, or unsat-

urated, and select very light colors for deselected and

very saturated colors for selected circles to ensure that

overlaid circles are not masked by map colors. We use

a ﬁxed number of colors, as the human ability to dis-

tinguish colors is limited. This also limits our system

to four input datasets. From the 12-color palette sug-

gested by Ware (Ware, 2004) for categorial usage, we

chose to use only four colors because we wish to en-

sure that mixtures of these colors are distinguishable

in our time-widget. The four colors used for datasets

1 through 4 are red, blue, green, and yellow. The order

was chosen to reduce the chances of misinterpretation

by color-blind people. We followed the recommen-

dation of Harrower (Harrower and Brewer, 2003) for

diverging colors in thematic maps. The user can also

change the colors from their default values.

3.1 Map Widget

The dominant widget in our visualization in terms of

screen space is the map widget. Following the deﬁni-

tions of Harris (Harris, 1999), the map widget consists

of a geographical map, which is overlaid with bub-

ble glyphs. We use circles of varying size for glyphs,

and each circle represents one or more items of the

dataset. We chose a circle representation on the map

to display data densities instead of, for instance, a heat

map, for multiple reasons. Multiple result sets re-

quire multiple heat maps, which in turn would require

mixing color maps, against which Ware argues (Ware,

2004). Also, humans are more accurate judging areas

than they are judging color tones, making areas a bet-

ter candidate for quantitative values (Cleveland and

McGill, 1984). Using glyphs also allows to group

glyphs and to make every data item individually ac-

cessible through its graphical representant.

In addition, we disallow the glyphs to overlap in

order to avoid visual clutter. To achieve this goal, we

merge circles based on their size, distances, and the

current zoom level in an iterative process. We ag-

gregate by proximity rather than geographic regions

because regions can have a bad area-to-diameter ra-

tio and the glyph resulting from the aggregation may

extend beyond the border, leading to ambiguities.

Aggregation for One Dataset. In the following let

N denote the number of items, k the number of sup-

ported zoom levels, and N

the number of circles at

zoom level l (0 ≤ l < k). r

min

is the radius used for

circles representing one item. We deﬁne r

min

depen-

dent on the average font size of labels of common web

mapping services (Google Maps, Bing Maps,. . . ) so

that small circles have salience equal to labels. The

maximum radius is deﬁned as r

max

= r

min

+ log

min

and A

max

are the resulting circle areas. The area

of an arbitrary circle c

is interpolated linearly de-

pending on the number n

of data elements, that are

assigned to it:



3 −



− 1

N − 1

= πr

= (1 − φ

min

+ φ

max

(1)

The weight in φ

ensures larger area differences for

higher zoom levels. The effect is, that differences be-

tween smaller circles are more salient.

To ensure non-overlapping circles, we perform a

variant of the hierarchical agglomerative clustering

algorithm. Two separate circles c

and c

need to be

merged if their Euclidean distance d

i j

after Mercator

projection is smaller than the sum of their radii r

and

. From this ratio we compute dissimilaries ψ

i j

and

merge circles c

and c

into a new circle c

i j

placed at

the barycenter of all their data elements if

i j

+ r

+ ε

< 1.

The gap value ε denotes the smallest permissible dis-

tance between two separate circle boundaries.

COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL DATA

615

(a) Individual circles. (b) Aggregated circles.

Figure 1: Aggregation of 307 points. The underly-

ing Guardian dataset shows key incidents during the

Afghanistan conﬂict. (a) Each circle with radius r

min

repre-

sents one incident. The high degree of overplotting makes

it hard to access each circle individually, and to determine

the precise conﬂict centers. (b) 307 elements were aggre-

gated to 43 non-overlapping circles. Larger circles clearly

indicate conﬂic centers.

The aggregation of the circles starts with N cir-

cles from which a proximity matrix Ψ is generated.

Initially, each single circle represents all items at a

common point. Iteratively, we merge the two circles

with the lowest value for ψ until there is no value be-

low 1. Each merge step causes two deletions and one

insertion of rows and columns in Ψ

i j

. We end with

k−1

circles for the highest zoom level k − 1 and pro-

ceed similarly to compute the N

l−1

circles at zoom

level l − 1 from the circles at zoom level l for each

0 ≤ l < k. We perform this agglomerative hierarchical

clustering procedure, which takes O(N

l+1

) time for

each zoom level l (with N

= N), until we have found

a non-overlapping circle set for zoom level l = 0. An

example result is given in Figure 1.

Aggregation for Multiple Datasets. In the case of

multiple (m = 2, 3, 4) input sets, we compose multiple

circles c

, . . . , c

into a more complex glyph – a circle

group – whose bounding circle b will be used in the

aggregation. We use a modiﬁed version of the congru-

ent circle packing method of Kravitz (Kravitz, 1967):

after computing the radius of each c

with Equation 1

we place m template circles t

at the same distance

from a hypothetical center z that touch each other at

their boundary. We then place the circles c

inside the

template circles t

and move them as far as possible to-

wards z without them leaving t

. In the case m = 4 the

biggest and second biggest circle are positioned oppo-

site each other. We then construct a minimal bound-

ing circle like in Figure 2.

Although the circle packing method requires a lot

of screen space, it leaves all similarly located ele-

ments very close and quantitative differences that are

reﬂected by varying diameters of the corresponding

circles are easy to spot. Furthermore, all circles can

Figure 2: Example circle packings.

(a) Country level. (b) Region level.

Figure 3: Level of detail of placename tag clouds. The used

Flickr dataset shows photos which are tagged with “Italy”.

(a) most of the places are “Italy” and “Vatican City”, (b)

administrative divisions of Italy, (c) most photos for capital

Rome, (d) Colosseum photos in Rome’s district San Paolo.

be accessed individually. We also chose packed cir-

cles over pie charts, because even though they can

present ratios of the data at one point very good, it

is difﬁcult to compare the size of pie slices between

pie charts of different size.

Placename Tag Cloud. Each of the map’s circles is

optionally associated with a set of placenames includ-

ing their frequency. A details-on-demand tag cloud

for a circle shows the most frequent placenames, their

font scaled proportional to their frequency. This fea-

ture provides a preview of how a glyph arising from

agglomeration would split if zoomed in. If the data

offers different levels of detail for a place, we choose

the label dependent on the current zoom level. We dis-

tinguish four semantic levels: country, region (e.g., a

state or a countryside), city, and borough (e.g, a dis-

trict, speciﬁc place or an address of the given city). If

one of levels is missing, we take the next coarser level

if possible, otherwise the next ﬁner level. An example

of the placename tag cloud is given in Figure 3.

Historical Maps. Because some of our applicants

wish to study datasets with historical context, we al-

low to show political borders for speciﬁc points of

time. In addition to a contemporary map we offer 23

different historical maps from 2000 BC to 1994 AD

provided by Thinkquest

. Because a dataset may span

a time range for which multiple maps are available

http://library.thinkquest.org/C006628/

IVAPP 2012 - International Conference on Information Visualization Theory and Applications

616

and there is no concept of “average political border”

we show the map closest to the median time stamp

occuring in the dataset as default and allow the user

to switch to any map at his convenience.

3.2 Time Widget

Using terminology of Harris (Harris, 1999), we con-

struct the time widget as a segmented area graph, that

has time as unit for the x-axis. The segmented area

graph is a line chart where the area under the line is

ﬁlled. T

, . . . , T

partition the interval T = [t

min

max

]

of the given dataset into intervals of regular duration:

either seconds, minutes, hours, days, weeks, months,

quarters, years, or decades. The resolution is chosen

to maximize the number of intervals without exceed-

ing 400. Short units typically arise from steadily up-

dated data sources and large units arise from data with

historic content. The resolution unit changes automat-

ically when the user zooms inside the time widget.

Whereas the x-axis of the time widget is directly

deﬁned by T

, . . . , T

, the y-direction shows the num-

ber of data items that fall in each interval using bin-

ning. For data items with time stamps the counting

is straight-forward, for data items with time spans

we add a value proportional to the amount of over-

lap with each bin. Although this can lead to an over-

representation of items with long time spans in the

time widget, we found this to be no problem. For the

datasets we considered, either the effect was benign

as the time spans had approximately the same dura-

tion or items with longer time spans were also more

interesting.

Time Binning for Multiple Datasets. In the com-

parative setting we perform the bin counting per

group. In the ﬁnal visualization the bins’ sizes are

shown as overlapping segmented area graphs rather

than bar charts because the former is better suited

to direct comparison of the groups’ time distribution.

We use the same colors for the groups as in the map

widget and shade the area under each line using a

semi-transparent version of that color. This ensures

that each curve is visible and also hints at the area

that would have been present in the bar chart. Despite

the use of blending, the limitation to four colors as

well as the stacked area graphs’ shape helps limit and

resolve ambiguities.

3.3 Detail Widget

For inspection of single data items that match the cur-

rent ﬁltering we present small textual or image thumb-

nails presented as a table in the detail widget. This

widget is the only one which does not include any ag-

gregation, but results are presented on multiple pages

if they exceed a certain ﬁxed number.

3.4 User Interaction

Our application offers multiple ways for the user to

interact with the data. Each widget provides native

navigation that results in updates of the other widgets.

The map and the time widgets provide simple zoom

and pan. Because of the way we aggregate data in

these widgets an animation between zoom levels or

time resolution changes is not performed. When fo-

cusing or selecting items in one view the correspond-

ing items are highlighted in the other widgets too.

The selections are linked. In particular, a selec-

tion within the map widget triggers a proportional se-

lection on the time widget. This is done by overlay-

ing the area graph with a bar chart that shows, how

many elements of a timeslice are associated with the

selected location. A selection on the timeplot triggers

a proportional selection on the map as well. There-

fore, we overlay a deselected circle with a circle with

the highlight color and a radius, that reﬂects the pro-

portion of elements, which are inside the selected time

interval.

Highlight. A glyph or time step that the mouse hov-

ers over is highlighted along in all other views. It is

a preview for selection. Hovering over a circle on the

map frames the circle with its associated placename

tag cloud. For a hovered bin in the time widget, the

number of associated elements and the corresponding

date is shown.

Selection. Each highlight can be turned into a se-

lection by a simple click. Every widget provides its

own means of selection and offers a broad set of fur-

ther user interactions, e.g., any selection can be turned

into a query reﬁnement by a context menu item.

In the map widget the user can select each cir-

cle individually. Thereby, the placename tag cloud

is frozen and the different place labels can be high-

lighted as well as selected.

It is also possible to draw shapes such as circles

or polygons. Selections can be reﬁned by dragging

shapes or control points. As a special selection the

user may also click on a country effecting the selec-

tion of all the items inside its borders. In historical

contexts, using maps of the time period of interest are

of great beneﬁt to this task.

The time widget allows both the clicking on one

bin and the selection of a time range using a mouse

drag gesture. A toolbar is then shown, that offers to

COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL DATA

617

modify the left or right border of the selected time era

and to add a “gray zone” which blends between se-

lected and deselected elements. Finally, the user can

manually move the selected time window. We update

the map widget automatically to facilitate a direct re-

ﬂection how locations change over time. A play but-

ton starts an animation mode and the selected time

span loops smoothly over the entire time range. This

mode directs the user’s attention towards changes in-

side the map.

The boxes in the detail widget reﬂect highlight

and selection by setting the border and background,

respectively, to the dataset’s color. Additionally, the

table can be used for the selection and deselection of

single items, to allow further reﬁnement to the user.

Intra Dataset Comparison. One of the major fea-

tures of the detail widget is the ability to export ele-

ments of a selection of a dataset as a new individual

input set. Hence, the temporal comparison of differ-

ent geographical regions of one dataset as well as the

geographical comparison of different time periods is

possible.

4 IMPLEMENTATION

We implemented our design in a prototype applica-

tion in the context of the European project (Purday,

2009) EuropeanaConnect

. The system’s architecture

is given in Figure 4(a).

The Web-server is mainly used for data process-

ing. Dependent on the requested data source, the

server retrieves the requested data and constructs a

KML-ﬁle with a speciﬁc format. For each item that

ﬁts the given query, we add a Placemark tag to the

KML root. Each Placemark is ﬁlled with a name, lo-

cation in form of latitude and longitude and either a

valid TimeStamp or TimeSpan. Optionally, a place-

mark node can be enriched with a location info (ad-

dress), that will be used for the placename tag clouds.

Slashes are used to separate their levels of detail. A

description entry can be used to commit a CDDATA

section that includes HTML content to offer detailed

information for the placemark. Finally, the KML-ﬁle

will be sent to the client as a response to the XML-

HttpRequest.

The client browser parses the retrieved KML-ﬁle,

creates JavaScript objects and ﬁlls the widgets with

content. We decided to construct a pure JavaScript

client, because it is supported by every browser with-

out any additional package. Another reason for

http://www.europeanaconnect.eu/

(a) Architecture.

(b) Screenshot.

Figure 4: System overview.

choosing JavaScript for our main application is the

rapidly growing browser support in terms of perfor-

mance. All interactions, including data reﬁnement are

performed on the client side. The great advantage of

this system structure is, that each modiﬁcation of a

dataset within the client browser triggers only some

functions on the client side. This beneﬁts the response

time of modiﬁcations, since the browser does not have

to wait for the server. We allow the user to load his

own KML-ﬁles. To the best of our knowledge our tool

is the ﬁrst to support this feature.

For the widgets on the client side we make use

of two OpenSource JavaScript libraries. A modiﬁed

Simile Widgets Timeplot

instance organizes the seg-

mented area graphs and OpenLayers

visualizes the

circles and arranges the different map layers. We of-

fer some popular maps as well as the historic maps,

http://www.simile-widgets.org/timeplot/

http://openlayers.org/

IVAPP 2012 - International Conference on Information Visualization Theory and Applications

618

(a) CrimeViz (low zoom level): Incident types arson, homi-

cide and sex abuse.

(b) Our system (low zoom level): Incident types arson (red),

homicide (blue) and sex abuse (green) in circle groups.

(green) and robbery (yellow).

(e) Co-located incidents at 27th of June. (f) Stolen car and burglary incidents in neigh-

borhood 13 (top) and 39 (bottom).

Figure 5: Crime incidents in Washington D.C. Figure 5(a) reproduced with permission from (Roth et al., 2010)

COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL DATA

619

which are explained in Section 3.1. We use a web

server instance of the GeoServer

, that provides the

tiles of the historic maps. A screenshot of our appli-

cation can be seen in Figure 4(b).

5 RESULTS

To demonstrate the ﬂexibility of our system, we

linked dynamic web sources as well as static data sets

with our system and compared our visualization to ex-

isting ones. Finally, we performed a ﬁeld experiment

(McGrath, 1995) to evaluate our design.

5.1 Crime Incidents

We compared our visualization to the CrimeViz appli-

cation. CrimeViz can only show incidents of one year

at once; we chose 2009 for comparison. In CrimeViz

the total of the frequencies of 3 incident types ar-

son, homicide and sex abuse is displayed in form of

hexagonal bins of different saturated red color hues on

lower zoom levels (Figure 5(a)). Thereby, the over-

lay causes a loss of map context in strongly saturated

areas. In our system places with many incidents are

salient through circle groups with different sizes (Fig-

ure 5(b)). Without the need to zoom or ﬁlter, we can

directly see that there was no homicide and arson inci-

dent in the north western neighborhoods. This cannot

be seen in CrimeViz. Furthermore, the time resolu-

tion in our system is more precise (days) compared

to CrimeViz (weeks). There are also no spatial ﬁlter

mechanisms offered by CrimeViz. Within our system,

we are able to ﬁlter incidents geographically by deﬁn-

ing polygon or circle container shapes, or directly by

selecting an administrative area, e.g., a police district

(Figure 5(c)).

The used crime database

offers a total of eight

different crime incident types. CrimeViz used only

the three types with smallest number of incidents: ar-

son (59), homicide (140), sex abuse (361). Other

types – theft (9,262), stolen car (4,852), robbery

(4,389), burglary (3,670), assault with deadly weapon

(2,623) – are not considered. Since each circle has

one spatial representative in CrimeViz and there is no

spatial aggregation on higher zoom levels, overplot-

ting would be a fatal problem for the larger datasets.

The additional crime types allow us to compare

crime incidents which may be more related to each

other than arson, homicide, and sex abuse. In Fig-

ure 5(d) we see the geospatial and temporal compar-

ison of stolen cars, burglaries, thefts, and robberies.

http://geoserver.org/display/GEOS/Welcome

http://data.octo.dc.gov/

By hovering over the glyphs in the city center, we

ﬁnd out that there are relatively few stolen car (20%)

and burglary incidents (23%) in comparison to thefts

(45%) and robberies (35%). Moreover, we ﬁnd pat-

terns for thefts and robberies near populated places

like metro stations or shopping centers. By exploring

individual days, we discover that a stolen car incident

is often grouped with a theft, burglary or robbery in-

cident (Figure 5(e)). The number of correlations in-

creases further by choosing a two day time range.

With the comparison feature, crime analysts could

improve the detection of connected crime incidents of

different crime types. Furthermore, our tool in combi-

nation with crime data could be used to drive the de-

cision making for apartment search. Interesting ques-

tions could be “Where is a secure region in terms of

burglary?” or “Where should I also rent a garage for

my car?” Figure 5(f) indicates Neighborhood 13 as

substantially safer than Neighborhood 39 with respect

to stolen cars and burglaries.

5.2 Guardian Data

The Guardian datablog

is another open source for

a multitude of datasets with geospatial-temporal data.

On 23rd of October 2010, The Guardian published

the Iraq war logs. The data contains information

about approximately 60,000 incidents during the Iraq

conﬂict from 2004 to 2009. Every incident claimed

at least one casualty; 109,000 total casualties are

mapped. Each incident is annotated with a time

stamp, latitude, longitude, incident type, and number

of casualties for each casualty type. The Guardian

also published a visualization (Rogers, 2010), where

each incident is represented by one circle on a map

(Figure 6(a)). The visualisation is cluttered as the re-

sult of overlapping glyphs. Figure 6(b) shows the geo-

graphical distribution of incidents with our system us-

ing the same color for glyphs. In both visualizations

regions with a low glyph density work well, e.g., a

series of horizontal circles from Baghdad (center) to

Jordan (east) highlights the importance of this con-

necting road. However, the conﬂict centers, Baghdad,

Al Mausi, and Al Basrah, cannot only be discrimated

easily in our visualization.

The temporal distribution shows a peak of inci-

dents between August 2006 and February 2007 (Fig-

ure 7(a)). The decreasing number of incidents after-

wards indicates the success of the Operation Impos-

ing Law, which started in February 14th 2007. The

Guardian map shows information about the incidents’

number of victims only in popups. We split the data

into four sets, depending on the four given casualty

http://www.guardian.co.uk/world/datablog

IVAPP 2012 - International Conference on Information Visualization Theory and Applications

620

(a) Guardian map. (b) Our map widget.

Figure 6: Iraq War Logs: Incidents distribution. Figure 6(a) reproduced with permission from (Rogers, 2010). (a) Three

conﬂict centers can be made out: Baghdad in the center, Al Mausi in the north, and Al Basrah in the south-east. Due to

overplotting, it is difﬁcult to determine which of these three regions suffers the most from incidences. The original visualiza-

tion omits some incidences for unknown reasons. (b) Baghdad in the center is clearly the region with most incidents (23.783

incidents, approx. 5,600 incidents in Al Mausi, 2,400 in Al Basrah, as simple mouse-over reveals). All incidences are shown.

types (civilian, enemy, iraq forces, coalition forces)

for comparison.

A comparison between incidents and civilian vic-

tims reveals two informations: (1) several incidents

without civilian casualties in the north-west of Bagh-

dad (Figure 7(b)), (2) incidents with many civilian

casualties, e.g., an incident which caused 158 civil-

ian casualties in the south of Karbala at April, 14th

2007 (Figure 7(c)). Another of these incidents is an

IED Explosion in the north of the Baghdad’s neigh-

borhood Rasheed that caused 94 civilian casualties at

February, 3rd 2007. The overplotting in the Guardian

Map masks this incident: it is hidden by an event with

one victim (Figure 7(e)). Our spatial aggregation pro-

cedure merges both incidents (Figure 7(d)); the inci-

dents’ details can be found in the detail widget.

The comparison of all casualty types shows that

most of the victims were civilians (65,649), and the

only period with more enemy than civilian casual-

ties was at the beginning of the conﬂict in 2004 (Fig-

ure 7(f)). There was also less ﬁghting at the road to

Jordan in 2004. The roads to Mosul (north) and to

Syria (north-west) were more contested at that time.

We are also able to reject the statement of the Iraq

Military from March 14th 2007, that there were only

265 civilian casualties in Baghdad since the start of

Operation Imposing Law. Figure 7(g) shows the time

range from February 14th to March 13th with a total

of 997 civilian casualties and a peak on March 5th,

caused by explosions around noon in Tigris.

5.3 Dynamic Data Sources

Humanities scholars within the eAQUA project

uchler et al., 2008; Heyer et al., 2011) use our sys-

tem to explore the geographic distribution and propa-

gation of words extracted from texts in ancient Greek.

The TLG research center at the University of Cali-

fornia created the corpus, which is one of the most

important resources that offers around 7200 Greek

works of over 1800 Greek authors. Text mining al-

gorithms extract words of strongly inﬂected ancient

Greek and tag them with temporal and geospatial in-

formation. The metadata spans the eastern Mediter-

ranean with Greece, Italy, Turkey, and North Egypt

and 18 centuries starting from 8 BC. The resulting

data divisions are shown in a suitable historic context

by automatically displaying adequate historic maps.

Figure 8 shows occurences of the words “Plato” and

“Aristotle” on the map of 400 AD.

We integrated our system into the online digital

catalogue of Europeana. It contains over 15 million

entries (images, texts, sounds, videos), which are dig-

itized resources of Europe’s libraries, museums, and

archives. Most of the items have geospatial (e.g., pub-

lication place) and temporal (e.g., publication year)

COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL DATA

621

(a) Peak of incidents 2006/07. (b) Incidents (blue) and civilian (red) casualties.

(f) More enemy (blue) than civilian (red) incidents 2004. (g) 997 civilian casualties (top), explosions on March

5th 2007 (bottom).

Figure 7: Iraq war logs: analysis. Figure 7(e) reproduced with permission from (Rogers, 2010).

IVAPP 2012 - International Conference on Information Visualization Theory and Applications

622

(a) Occurences during middle platonism. (b) Occurences during neoplatonism

Figure 8: Occurences of the words “Plato” (red) and “Aristotle” (blue) in ancient Greek texts: In the ﬁrst period – Middle

Platonism (1st to 3rd century) – we see a widely spread distribution of “Plato” and “Aristotle” (Figure 8(a)) in the Greek-Ionic

region. The second period – Neoplatonism (4th to 6/7th century) – shows a lot of occurences in the territories of Athens and

Constantinople (Figure 8(b)). This indicates a correlated movement of both topics from rural regions to metropolises.

metadata. In the detail widget we use thumbnails,

which link to the Europeana source page. Figure 9

shows the distribution of occurences of different his-

toric watermark types (tokens of letter authenticity).

5.4 Field Experiment

To evaluate our design and spot areas of usabil-

ity improvements we conducted a ﬁeld experiment.

We asked 15 individuals to partake; 7 men and 3

women followed our call. The participant’s occupa-

tion ranged from pupil to pensioner and their age from

13 to 71 years. The participant’s educational back-

ground varied, but none had notable computer sci-

ence background. Each test took place at the partici-

pant’s typical location where he accesses the internet.

The average session duration was approximately two

hours. A session consisted of three parts: a tutorial,

solving tasks with the system, and an interview.

First, each participant was asked to work through a

tutorial, which consists of 10 short ﬂash videos linked

from the application. The tutorial introduced the wid-

gets and features of the system. To understand each

feature, the users were encouraged to solve basic per-

ceptual and manipulation tasks. Parts of the tutorial

could be skipped or repeated, but we asked the partic-

ipants not to ask questions to the present conductor of

the study, who only observed, whether the pace, clar-

ity, and volume of the tutorial was appropriate.

After ﬁnishing the tutorial, each participant was

given a form with 15 prepared questions, which he

had to answer using the system. During that time the

participant could replay tutorials or ask the study con-

ductor for help, but participants only used the oppor-

tunity to verify their understanding of the questions.

The questions had an increasing complexity, and the

solutions required the use of all features mentioned in

the tutorial. Questions could be answered in any or-

der and skipped if considered to be too difﬁcult. The

solutions had different types: single or multiple data

items, a speciﬁc time, place, placename, and even

combinations of them. An example question asked

to verify the Iraq police statement on the success of

Operation Imposing Law (see Section 5.2). For some

questions the users had to estimate solutions or ﬁnd

trends in space and time.

The fraction of correct answers was 75.3%. Even

the oldest subject attempted 8 of 15 questions and all

of his answers were correct. Some answers were in-

correct but “almost” right (10.0%); there was often

just an single overlooked glyph or time step. Only a

small portion of answers were wrong (6.7%) or not

solved (8.0%). By comparing the questionnaires, we

discovered, that the ratio of correct answers was inde-

pendent from the internet usage per day and the reg-

ularly visited platforms. Users with lesser experience

required more time to ﬁnish the tutorial, and they also

took more time in answering the questions, but the

ratio of correct answers was comparable. Even expe-

riences with web mapping services were not relevant

for the correctness of the given answers.

Finally, the subjects were asked openly how they

liked the application, followed by which features they

liked or disliked, and what they would change. The

general reception of the application was very pos-

itive. Although they stated to require the tutorial,

which introducted only elementary interactions with

single widgets, they learned and used the synergetic

effects the more they used the tool. We also noted

that the majority of the subjects did not make use of

the time animation control, but those who took it to

solve the questions, claimed it to be a useful feature.

COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL DATA

623

Suggested improvements were mostly of technical na-

ture, e.g., avoiding overlapping boxes inside the time

widget and reducing the startup time of the system

for very large datasets. Not all of them were using the

browsers we reccomended for fast JavaScript execu-

tion. They also proposed to use pie charts rather than

circle packing, to provide placename tag clouds for

circle groups, selections for continents, and a speed

adjustment for the time animation feature. Some

tasks, for instance the selection of multiple countries,

have been implemented in the mean time.

Figure 9: Watermarks tagged with “Snake” (blue), “Bird”

(red) and “Cloche” (green). Compared to Bird, Cloche

and Snake watermarks were used more often. Whereas the

Cloche watermark is spatially distinct and often used in the

14th and 15th century, the Snake watermark was used be-

tween 1450 and 1700, mostly in northern Italy and central

Germany. Domain experts attribute this to the popularity of

the Snake watermark as the heraldic device of the Milan dy-

nasty of the Visconti, and later on as Swabian papermakers’

hallmark of excellence.

6 CONCLUSIONS AND FUTURE

WORK

We presented a novel approach and web applica-

tion to show, compare, and explore multiple topical

query results in a geographical and temporal context.

We were able to utilize, combine, and improve ap-

proaches from several related works. In comparison

to CrimeViz (Roth et al., 2010), which also offers

comparitive visualization, we found a consistent so-

lution for displaying results of different queries in the

geospatial dimensions. We display them separately

without aggregating them into the same representa-

tives for coarser zoom levels. Furthermore, with our

method we are able to solve the spatial overlap prob-

lem, which is also an issue in the Guardian visualiza-

tion for the Iraq war logs (Rogers, 2010). Compared

to the similar system VisGets (D

ork et al., 2008),

which only works for one set of results, we also made

use of the linked widgets approach (map, timeline,

detail widget) to extend the users exploration abili-

ties. Furthermore, we enriched the ﬁlter capabilities

in both geospatial (e.g., selecting all results inside a

country) and temporal dimension (e.g., selecting dy-

namic time ranges). In contrast to VisGets, where

most user interactions lead to a server request, we

placed nearly all system logic inside the client to ef-

fectively improve the response times.

Our case studies show that visually comparable

datasets extend the exploration and analysis abilities

of the user in an effective way. It helps to detect

equalities and varieties between distinct data contents

that unveil their relationships in space and time. Our

method is limited to four datasets at a time mainly

to ensure that the colors used for discrimination are

properly distinguishable, the splitting of circles does

not waste too much screen space, and the overlapping

segmented area graphs do not occlude each other too

much.

In the ﬁeld experiment, we could establish through

the feedback of the participants, that our design is

easy to learn and suitable to the given task of com-

paring multiple input sets in geospatial and temporal

dimensions. In addition, the aggregation of items of

distinct sets were evaluated as intuitive and the func-

tionalities we offer to interact with the data were easy

to adopt with the help of the provided tutorial – inde-

pendent on the class of age, education level, and the

amount of participants’ internet usage. We also re-

ceived some hints on how to improve our application.

In the future we will direct our attention to the ex-

tension of our method to very large datasets by search-

ing for client-server communication where most of

the data remains on the server but a working set is

transmitted to the client for quick interaction. We aim

to extend our system to extract and show trajectories

of distinct sets containing movement data, potentially

leading to new insights.

The dynamic visualization of recent Twitter feeds

is also one of our future objectives. We hope to see

how topics proliferate. Preparing examples which

automatically update their results by querying social

platforms in equal time intervals could even show the

live change of topics’ relations.

ACKNOWLEDGEMENTS

We like to thank our colleagues Christian Mahnke,

Marco B

uchler, and Muhammed Faisal Cheema for

fruitful discussions and our ﬁeld experiment partici-

IVAPP 2012 - International Conference on Information Visualization Theory and Applications

624

pants for their time. We are also indebted to Cornelius

uller for preparing datasets. We also thank Simon

Rogers and Robert E. Roth for providing screenshots

of their systems. This research was funded by the Eu-

ropean project EuropeanaConnect.

REFERENCES

Andrienko, G. and Andrienko, N. (2006). Visual Data Ex-

ploration: Tools, Principles, and Problems. In Clas-

sics from IJGIS: twenty years of the International

journal of geographical information science and sys-

tems, pages 475–479. CRC Press.

Andrienko, N. and Andrienko, G. (2005). Exploratory

Analysis of Spatial and Temporal Data: A Systematic

Approach. Springer.

Baldonado, M. Q. W., Woodruff, A., Kuchinsky, A., and

Kuchinsky, A. (2000). Guidelines for using multiple

views in information visualization. In Advanced Vi-

sual Interfaces, pages 110–119.

Barker, E., Bouzarovski, S., Pelling, C., and Isaksen, L.

(2010). Mapping an ancient historian in a digital age:

the Herodotus Encoded Space-Text-Image Archive

(HESTIA). Leeds International Classical Studies,

9(1).

Bertin, J. (1983). Semiology of graphics. University of Wis-

consin Press.

uchler, M., Heyer, G., and Gr

under, S. (2008). eAQUA

- Bringing modern Text Mining approaches to two

thousand years old ancient texts. In Proceedings of the

4th International Conference on e-Science (IEEE08).

Cleveland, W. S. and McGill, R. (1984). Graphical Percep-

tion: Theory, Experimentation, and Application to the

Development of Graphical Methods. Journal of the

American Statistical Association, 79(387):531–554.

ork, M., Carpendale, S., Collins, C., and Williamson, C.

(2008). VisGets: Coordinated Visualizations for Web-

based Information Exploration and Discovery. IEEE

Transactions on Visualization and Computer Graph-

ics, 14(6):1205–1212.

Hardisty, F. and Klippel, A. (2010). Analysing Spatio-

Temporal Autocorrelation with LISTA-Viz. Int. J. Ge-

ogr. Inf. Sci., 24:1515–1526.

Harris, R. L. (1999). Information Graphics: A Comprehen-

sive Illustrated Reference. Oxford University Press.

Harrower, M. and Brewer, C. A. (2003). ColorBrewer.org:

An Online Tool for Selecting Colour Schemes for

Maps. The Cartographic Journal, pages 27–37.

Havre, S., Hetzler, B., and Nowell, L. (2000). ThemeRiver:

Visualizing Theme Changes over Time. In Proceed-

ings of the IEEE Symposium on Information Vizual-

ization 2000, INFOVIS ’00, pages 115–, Washington,

DC, USA. IEEE Computer Society.

Heer, J., Bostock, M., and Ogievetsky, V. (2010). A

Tour Through the Visualization Zoo. Commun. ACM,

53(6):59–67.

Heyer, G., B

uchler, M., Boehlke, V., and Schubert, C.

(2011). Aspects of an Infrastructure for eHumanities.

ACM Journal on Computing and Cultural Heritage.

Kravitz, S. (1967). Packing Cylinders into Cylindrical Con-

tainers. Mathematics Magazine, 40(2):65–71.

Marchionini, G. (2006). Exploratory Search: From Finding

to Understanding. Commun. ACM, 49(4):41–46.

McGrath, J. E. (1995). Methodology Matters: Doing Re-

search in the Behavioural and Social Sciences. In

Readings in Human Computer Interaction: Towards

the Year 2000, pages 152–169. Morgan-Kaufmann,

San Francisco, CA, 2nd edition.

Purday, J. (2009). Think Culture: Europeana.eu from Con-

cept to Construction. Electronic Library, 27(6):919–

937.

Rogers, S. (2010). The Guardian - Wikileaks Iraq war logs:

every death mapped. http://www.guardian.co.uk/wor

ld/datablog/interactive/2010/oct/23/wikileaks-iraq-

deaths-map (Retrieved 2011-10-07).

Roth, R. E., Ross, K. S., Finch, B. G., Luo, W., and

MacEachren, A. M. (2010). A User-Centered Ap-

proach for Designing and Developing Spatiotemporal

Crime Analysis Tools. In Weibel, R. P. R., editor, GI-

Science, Zurich, Switzerland.

Shneiderman, B. (1994). Dynamic Queries for Visual In-

formation Seeking. IEEE Software, 11(6):70–77.

Shneiderman, B. (1996). The Eyes Have It: A Task by Data

Type Taxonomy for Information Visualizations. In Vi-

sual Languages, Proceedings, pages 336–343.

Tsipidis, S., Koussoulakou, A., and Kotsakis, K. (2011).

Geovisualization and Archaeology: supporting Exca-

vation Site Research. In Ruas, A., editor, Advances

in Cartography and GIScience. Volume 2, volume 6

of Lecture Notes in Geoinformation and Cartography,

pages 85–107. Springer Berlin Heidelberg.

Tufte, E. R. (2001). The Visual Display of Quantitative In-

formation. Graphics Press, second edition.

Tukey, J. W. (1977). Exploratory Data Analysis. Addison-

Wesley.

Wang, X., Miller, E., Smarick, K., Ribarsky, W., and Chang,

R. (2008). Investigative Visual Analysis of Global

Terrorism. Comput. Graph. Forum, 27(3):919–926.

Ware, C. (2004). Information Visualization: Perception for

Design. Morgan Kaufmann.

COMPARATIVE VISUALIZATION OF GEOSPATIAL-TEMPORAL DATA

625