A Geological Metaphor for Geospatial-temporal Data Analysis

Tom Liebmann, Patrick Oesterling, Stefan J

anicke and Gerik Scheuermann

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

Keywords:

Visual Data Exploration, Abstract Data Visualization, Geo-Visualization, Comparative Visualization.

Abstract:

To provide visual access to geospatial-temporal data, existing systems usually highlight the data’s spatial,

temporal and topical distribution individually in separated, but linked views. Because this design often com-

plicates queries that concern multiple data aspects and also involves more user interaction, in this paper, we

present a geological metaphor that aims to combine relations between orthogonal data aspects. We describe

how our adopted landscape metaphor intuitively depicts global and local relationships based on its surface,

glyph augmentation and inner sediment structure. We validate the geological metaphor with case studies,

compare it with existing systems and describe how it can be integrated into those as an alternative map view.

1 INTRODUCTION

Analyzing data and extracting information from

databases has long been text-based. However, deﬁn-

ing queries with proper keywords can be frustrat-

ing and browsing through textual result sets does not

scale well, depends on language, and excludes human

abilities to distinguish and rank things visually. If

data contains additional meta-information, query re-

sults can be spatialized to rely on preemptive, paral-

lel abilities of the human eye-system. According to

Ware (Ware, 2004), the user can then quickly distin-

guish data aspects based on, e.g., position, color or

shape rather than on words in textual results.

For example, in case of geospatial-temporal data,

i.e. data associated with geospatial position and a

time-stamp, it has proven beneﬁcial to lay out items

on a map to identify interesting subsets and to steer or

reﬁne queries interactively. Several frameworks, such

as GeoTemCo (J

anicke et al., 2013), VisGets (D

ork

et al., 2008), or CrimeViz (Roth et al., 2010) have

been introduced. They typically consist of at least a

map- and a time-widget to provide spatial and tem-

poral context, respectively, and they usually support

linked-selection and linked-brushing. Selections are

often speciﬁed in other (maybe overlayed) widgets

like tag-clouds, histograms, or overview-widgets.

However, because these frameworks illustrate or-

thogonal information in separate widgets, data as-

pects can also be analyzed only individually. As a

result, queries concerning multiple aspects may not

be feasible, or they require more user interaction. For

example, if the data’s spatial distribution over time,

or topical distribution at certain locations over time

is questioned, the user is required to perform single

selections in the time-widget, while watching how

linked selections change in the map-widget. Not only

does this imply many selections, the user also needs

to store a mental map of all changes because linked

selections replace each other. Many approaches also

augment the map with glyphs to indicate data oc-

currence. For large data, this can be problematic if

glyphs overlap or if aggregated glyphs, with size pro-

portional to item count, suggest data in actually void

areas. Furthermore, glyph color is often used to dis-

tinguish queries or data types, and has thus only little

potential to convey other information dimensions.

In this paper, we present a 3-D geological

metaphor for geospatial-temporal data to illustrate re-

lations between orthogonal information dimensions in

one view. To this end, we distort the map so that

data occurrence is evidenced by hills with elevation

proportional to item count in that area. We color a

hill’s surface to display temporal distribution and use

glyphs per hill as a third information channel to re-

ﬂect an item’s class, which could represent its afﬁli-

ation to a particular query, or another data attribute.

Because terrains are familiar and intuitive to humans,

global overview of the data’s main aspects can eas-

ily be extended to local analysis. We allow the user

to dissect hills to watch sediment evolution over time.

Depending on sediment granularity, relations between

orthogonal key information like geospatial position,

time and class can thus be analyzed at the same time.

161

Liebmann T., Oesterling P., Jänicke S. and Scheuermann G..

A Geological Metaphor for Geospatial-temporal Data Analysis.

DOI: 10.5220/0004742901610169

In Proceedings of the 5th International Conference on Information Visualization Theory and Applications (IVAPP-2014), pages 161-169

ISBN: 978-989-758-005-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Landscape metaphors support intuitive data explo-

ration and provide another degree of freedom in 3-D.

Still, using three dimensions rather than two and dis-

torting and coloring the map can limit the metaphor’s

application. Therefore, we view its strengths on less

detailed maps and in scenarios where precise dis-

play of single data aspects is neglected in favor of a

combined illustration of relationships between several

data dimensions. Often these relations disclose fea-

tures that deserve closer investigation regarding one

or two data aspects in a second analysis step. Our

visualization is integrable into other frameworks as a

surrogate map-widget. Therefore, it supports ﬁltering

and highlighting of linked-selections, as well as send-

ing user selections back to other widgets.

2 RELATED WORK

Visualizing data in combination with maps (Tominski

et al., 2005) to provide the user with spatial context

relates to thematic cartography and geo-visualization.

An overview is given by (Dent, 1999) and (Slocum

et al., 2009). Showing information as intuitive land-

scapes has long tradition in information visualization

and was, for example, adopted as ThemeScapes (Wise

et al., 1995) to depict and reorganize information

that is not inherently spatial, or as GraphScapes (Xu

et al., 2007) to reveal multivariate graph clustering

using landscapes as attribute surfaces that are aug-

mented with the underlying graph. An ontology of

the landscape metaphor and an overview how it actu-

ally works is given by (Fabrikant et al., 2010).

Systems to analyze geo-temporal data typically

break down the analysis semantically, by follow-

ing the information-seeking mantra ’Overview First,

Details on Demand’, as proposed by (Shneiderman,

1996), but also technically by providing the user with

linked views on different data aspects. The role of

multiple coordinated views in exploratory visualiza-

tion is described by (Roberts, 2007).

The concepts of visual data exploration and geo-

temporal data analysis, including their principles and

challenges are thoroughly explained by (Andrienko

and Andrienko, 2005; Andrienko and Andrienko,

2006). In our adopted geological metaphor we use

a third dimension to indicate data occurrence and to

reﬂect temporal evolution of inner sediment struc-

ture. Utilizing 3-D for time-geography studies to re-

late event times to locations on a map is often seen

to originate from the Space-Time Cube (Kraak, 1988)

that uses space-time paths and their footprints on the

map to understand movement data. Throughout the

years this model has been improved and adopted, for

example, by GeoTime (Kapler and Wright, 2005) to

visualize the spatial inter-connectedness of informa-

tion over time; including extensions to augment Geo-

Time with a story system that uses narratives, hy-

pertext linked visualizations, annotations, and pat-

tern detection for analytic exploration (Eccles et al.,

2008). In a similar fashion (Tominski et al., 2012)

gain insights into trajectory attribute data by inte-

grating spatial and temporal displays based on color-

coded trajectory bands that are stacked perpendicular

to the map. For multivariate data analysis, the usage

of kernel density estimation to determine probability

values was also employed by (Kim et al., 2013) to

present Bristle Maps, a series of straight lines that

extend from map elements such as roads to create a

multivariate attribute encoding, and by (Maciejewski

et al., 2010) to ﬁnd spatiotemporal hotspots by over-

laying heatmaps and contours for different data as-

pects, linked to time-series plots for user-selected re-

gions of interest.

Most frameworks for geo-temporal data analy-

sis, such as GeoTemCo (J

anicke et al., 2013), Vis-

Gets (D

ork et al., 2008) and CrimeViz (Roth et al.,

2010), share a very similar architecture in that they

employ a map-widget and a time-widget to provide

spatial and temporal context, respectively. They

mainly differ in that not all of them support compar-

ative analysis, that some of them provide additional

tag-clouds and result views, and that temporal selec-

tion in their time-widgets happens at different gran-

ularities — using stacked bar-charts, histograms, or

graphs. Together with the trivial visualization of the

Iraq conﬂict incidents (Rogers, 2010), these systems

use glyph augmentation to indicate data occurrence

on the map, though varying in quality by trying to

avoid glyph overlap and visual clutter using different

binning- and aggregation-solutions.

Our work differs in that we focus on a more ab-

stract, but combined display of several data aspects

in one view. We substitute data item glyphs for hill

elevation to indicate item count - which allows us to

present features with more granularity and higher ac-

curacy on the map. Furthermore, spatial granularity

is decoupled from item count and the map’s zoom-

level. For example, in GeoTemCo large glyphs can

easily cover void areas around local peaks and actu-

ally distant parts could be represented by one glyph

because of high item count at these locations. Then

the user has to zoom-in, which unnecessarily reduces

the number of visible and thus comparable features on

the map.

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

162

3 LANDSCAPE METAPHOR FOR

GLOBAL DATA ANALYSIS

This section describes how our landscape metaphor

is designed to convey orthogonal information dimen-

sions and how the user interacts with the visualization

— both inside the widget and by linked-selections

from other widgets.

3.1 Map Distortion

Conceptually, each data item is assigned a volume so

that data occurrence is ﬁnally evidenced by agglom-

erations with extent and height proportional to item

count. We use a Gaussian-based approach and con-

sider all data positions to be samples of an unknown

random variable whose density distribution we ap-

proximate. The probability density function then de-

scribes the shape and the structure of the map.

Gaussian density estimation is subject to a kernel

ﬁlter radius σ that controls each item’s volumetric ex-

tent. The user can change this parameter to control

the landscape’s level-of-detail and to coarsen or reﬁne

features, e.g. by isolating countries, regions, or single

cities with different ﬁlter radii. For aesthetic land-

scapes, elevation values can be reduced by (linear or

logarithmic) kernel scaling and maximum height val-

ues can be bound to a percentage of the map extent.

Showing data concentration as hills permits fast

localization and comparison, includes hierarchies

(sub-clusters) and illustrates features locally: since

item count translates into elevation, neighbored void

areas are not covered by data representatives; like

large, aggregated glyphs in 2-D would do. Still, per-

spective projection can impede comparing heights of

elevations. Although we consider precise comparison

to be part of local data analysis (cf. Section 4), the

usage of hypsometric tints can counter this issue.

3.2 Landscape Color

In void areas, the map still provides enough spatial

context to relate hills to geographic locations. On the

hills, however, map readability is likely to be reduced

due to distortion. Therefore, map display can be deac-

tivated on hills to utilize surface color to convey time

as a second information dimension.

Because a hill reﬂects the agglomeration of poten-

tially many data items, multiple time-stamps can con-

tribute to each surface point’s color. The idea is to use

the mean time and standard deviation of all involved

time-stamps and to map both values to one color. To

this end, the whole dataset’s time span is mapped to

a color gradient. Then for each set of time-stamps,

(a) (b)

(e)

Figure 1: (a) Landscape with surface colored by time. (b)

Active (green) and inactive line segments to localize pro-

ﬁle generation. (c) Glyph-augmentation showing data item

proportions for class and (d) for time for every component

above the red polygon. (e) Slider to ﬁlter data items by time.

the mean time is ﬁrst mapped onto the gradient and

then the standard deviation deﬁnes that color’s light-

ness. This scheme allows to display the chronological

sequence of the dataset and to distinguish young, old,

and mid-aged regions. Moreover, controlling light-

ness values with standard deviation helps distinguish

mid-aged regions from (lighter) regions where young

and old data items are just mixed (cf. Figure 1a).

3.3 Data Filter

To support handling big datasets and to facilitate

closer investigation of suspicious features, the user

can ﬁlter data by class, time, and hill elevation.

The class ﬁlter determines whether a data item

should contribute to the landscape construction based

on its class afﬁliation. This is helpful if the spatial and

temporal distribution of single classes is questioned,

or if only a few classes should be compared. The ﬁlter

can be set with a simple list or any other third-party

widget that allows to select possible classes.

The time ﬁlter deﬁnes two points in time and ex-

cludes all data items with a time-stamp outside this

period. While a simple two-slider widget, maybe aug-

mented with histograms to indicate item count, is suf-

ﬁcient, more sophisticated widgets, as proposed in

GeoTemCo (J

anicke et al., 2013) or VisGets (D

ork

et al., 2008), can also be used. In any case, we rec-

ommend to augment the time-widget with the color

AGeologicalMetaphorforGeospatial-temporalDataAnalysis

163

gradient used for the landscape surface. This helps

identify thresholds necessary to eliminate particular

hills based on their color (cf. Figure 1e).

The elevation ﬁlter removes or preserves parts of

the landscape based on user-deﬁned minimum and

maximum height values. Hills with an elevation out-

side this range are removed from the landscape. The

ﬁlter is helpful to eliminate small hills in regions with

only little data occurrence, but also to remove promi-

nent hills to concentrate on vague features.

The ﬁlter order is crucial because class and time

ﬁlters affect the height of all remaining hills before

they are eventually ﬁltered by correct elevation.

3.4 Glyph Augmentation

Because map distortion and surface color already

summarize the data’s spatial and temporal distribu-

tion, glyphs can now provide information about a

third data aspect for items below the surface. Note

that in contrast to a single value on the surface,

glyphs also provide another degree of freedom to

present compositions or proportions. The user inter-

actively controls glyph generation per hill by deﬁn-

ing polygons on the map. By default a single poly-

gon comprises the whole map to generate glyphs for

all separated hills. A polygon’s height can then be

changed by moving it parallel to the map, which

changes the number of landscape components that re-

main above. Glyph are ﬁnally generated for every

component within and being crossed by a polygon.

This way, one can effectively control whether sub-

hills should contribute to their parent-hill’s glyph, or

if they should have an own glyph. If placed directly

on the map, a single glyph provides proportions for all

data items within the polygon. Glyph generation thus

depends on hill granularity (cf. Section 3.1), but can

also be adjusted based on sub-hill relationships. Us-

ing polygons at different heights also allows to com-

pare features at different granularities.

We use rectangular bars that always face towards

the user’s viewing direction to indicate the distri-

bution of class or time within the hills. In both

cases, all data items that contribute to a component

above a polygon are analyzed and proportions are

then mapped on the glyph. For the class glyph, pre-

deﬁned colors can be used (cf. Figure 1c) to quickly

identify how classes are distributed over the map and

over time based on the landscape’s color. Glyphs can

also be used to convey proportions of the data’s time

aspect. Because proportions are better perceived if

only a few colors are used (Ware, 2004), we divide the

time span into three and map the proportions of old,

mid-aged, and young data items using the color gradi-

ent that represents time. A semicircle above each bar

additionally indicates the mean time (cf. Figure 1d).

Glyphs and whole polygons support data selec-

tion. Picking them could, e.g., trigger tag-cloud aug-

mentation or linked-selections. Glyph colors can also

be extended to support linked-brushing for items se-

lected in other widgets.

3.5 Deﬁning Sediment Proﬁles

Although proportions on glyphs already provide more

details than surface color, they do not consider

chronological order. Because the data’s time aspect

can be in highly interesting relations with class or

location, we reveal them using the landscape’s inner

sediment structure. To this end, the user interactively

controls for which hills sediment proﬁles should be

exposed. Sediment proﬁles are perpendicular to the

map and thus speciﬁed as line segments to deﬁne lo-

cation and extent. To better relate proﬁles to the map,

line segments are triangular-shaped and colored (cf.

Figure 1b). Furthermore, they are freely movable by

dragging the triangle bar or the vertical plane, resiz-

able by dragging the end-points, and they can be con-

nected to support local analysis from different angles.

It is thus easily possible to juxtapose inner structure

of physically distant parts of the landscape.

4 SEDIMENT PROFILES FOR

LOCAL DATA ANALYSIS

Sediment proﬁles could be exposed by excavation

or by making occluding hills transparent. However,

because this visual approach would still suffer from

problems with orientation and perspective distortion

in 3-D, we display sediment proﬁles side-by-side in a

2-D view. This section describes a proﬁle’s properties

and how multiple proﬁles are managed.

4.1 Sediment Proﬁle

Conceptually, sediments arise during the landscape

construction if data items are handled in groups and

in chronological order. Sediments are thus always

sorted by time and a sediment’s height depends on the

number of group elements and their current weight

on the proﬁle. A data item’s weight is the amount of

proﬁle intersection with its volumetric representation,

i.e., for Gaussians, the density it contributes to the lo-

cation of the line segment in the map-widget.

We use two partitioning schemes. If partitioned

by time, the whole time span is evenly divided into

a user-deﬁned number of groups. Items are then

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

164

(a) (b)

Figure 2: (a) Sediment proﬁle augmented with glyphs. (b)

Proﬁle in split-mode to identify peaks per sediment layer.

grouped based on their time-stamp and each group is

represented by a sediment. Note that a sediment’s ap-

pearance in the proﬁle still depends on the line seg-

ment’s position and the data’s spatial distribution. If

partitioned by number, all data items are divided into

groups of user-deﬁned size. The minimum size is one,

when each data item has its own sediment.

Sediment colors play an important role to display

relations between multiple data aspects. To relate

time to location we color sediments by time. That

is, the mean time of a sediment group’s time-stamps

is ﬁrst interpolated according to their current weights

and then mapped on the color gradient. This reveals

in which order inner structure emerged (cf. Figure 5,

second row). If adjacent sediments have similar col-

ors it can be difﬁcult to identify spatial distribution,

i.e. whether temporally close items are distributed on

different hills. In this case, random colors help dis-

tinguish close-by sediments (cf. Figure 5, ﬁrst row).

Coloring sediments by class allows to relate time to

both location and class. Single-item sediments di-

rectly use the class’ color. For larger groups, colors

are mixed according to the items’ weights. Although

mixing colors can result in invalid class colors, this

technique still provides good results for sufﬁciently

small sediment groups (cf. Figure 5, third row).

A sediment’s shape is inherently affected by its

subjacent layers. Because this can complicate ﬁnding

local maxima for curved layers, we provide a split-

sediments mode. While the chronological order is still

preserved, every sediment is treated as if it was at the

bottom (cf. Figure 2). This permits precise compari-

son of local data increase between two points in time

or at two different locations in the proﬁle.

The sediments allow to display single data items.

Although we could exactly fragment each layer ac-

cording to the data items’ volumetric representations,

for aesthetic reasons, we only place glyphs per data

item. A glyph’s horizontal and vertical position is de-

termined by its item’s location and by the shape of

its sediment layer, respectively. Glyph size is pro-

portional to the item’s weight, i.e. it is maximal if

the proﬁle is placed directly on top of it in the land-

scape view. Glyphs are colored by class and addi-

(a) (b)

(c)

Figure 3: (a) Gaussian sample mask of radius σ (blue) with

additional sample points (black) around data item (red). (b)

Grid to quickly identify other items in σ-distance. (c) Quad-

strips to assemble sample positions into sediment layers.

tional meta-information can be presented at mouse

hover events. Although glyphs can be of arbitrary

shape, we use clam-shaped glyphs to be visually in

line with the sediment metaphor.

A proﬁle provides several means to select data

items either glyph- or sediment-based. Glyph-based

selection includes direct picking or using arbitrarily

shaped polygons. Sediment-based selection means to

select data as arbitrary groups of sediments. Linked-

brushing, i.e. highlighting items that are selected in

other widgets, is achieved by adjusting glyph colors.

4.2 Management of Multiple Proﬁles

To display individual proﬁles, line segments can be

manually (de-)activated in the landscape view. Pair-

wise proﬁle comparison is then achieved by chang-

ing their order in the 2-D view. To better relate a

proﬁle to the map, the corresponding line segement’s

triangular-shaped color bar is placed below the pro-

ﬁle. Furthermore, we allow the fusion of sediment

proﬁles. In this case, an extended proﬁle is generated

based on two line segments and they are combined

visually by merging their item groups. This is help-

ful for line segments that are connected at their end-

points or if physically distant parts of the landscape

should be treated as if they were actually next to each

other (as will be demonstrated in Section 6.1).

5 IMPLEMENTATION ISSUES

The implementation of the landscape metaphor is

straightforward and can be realized as a simple mod-

ule chain. If a module’s parameter changes, repetitive

landscape construction then efﬁciently starts at this

module and only includes subsequent modules. At

ﬁrst, the data is ﬁltered by class and time before den-

AGeologicalMetaphorforGeospatial-temporalDataAnalysis

165

sities are evaluated for remaining data items. After

ﬁltering those items by density, the landscape geome-

try is generated and vertices are provided with correct

elevation and color values. The last modules create

glyphs per hill and sediment proﬁles, as triggered by

user interaction. The remainder of this section pro-

vides details about major implementation issues.

5.1 Landscape Metaphor

Implementing the landscape as a regular high-

resolution grid quickly becomes unmanageable for

accurate display of small hills on a big map. There-

fore, we apply discrete sampling masks at (ﬁltered)

data item positions to obtain a sufﬁciently precise ap-

proximation (cf. Figure 3a). Then a 2-D Delaunay

triangulation of all sample points serves as the ﬁnal

landscape. To color the surface without interpola-

tion artifacts, required information, like mean time

and standard deviation of contributing data items, is

attached to each vertex and processed by a fragment

shader using color gradient textures (cf. Figure 1a).

Many calculations can also be accelerated and paral-

lelized if data items are stored in a grid with a res-

olution of the Gaussian ﬁlter radius (cf. Figure 3b).

To generate glyphs per hill, the Delaunay triangula-

tion can easily be separated into connected compo-

nents above a certain height by starting a depth-ﬁrst

search from every vertex to mark component associ-

ation. This process executes in linear time as every

vertex is processed only once. Data information as-

signed to each vertex is then used to generate glyphs.

5.2 Sediment Proﬁles

Sediment layers are generated by equidistant sam-

pling of the density function on the line segments de-

ﬁned in the landscape view. Simple quad-strips are

used to assemble sediment layers and the whole pro-

ﬁle (cf. Figure 3c). To identify data items that con-

tribute to the proﬁle, data can be stored in a kd-tree.

Then a range query with a size that comprises the pro-

ﬁle on the map quickly returns the desired items.

6 CASE STUDIES

In this section, we demonstrate the key features of our

geological metaphor. At ﬁrst, we perform an exem-

plary investigation on a literature dataset, followed by

a comparison of the landscape to alternative overview

visualizations, based on the Iraq war logs (Rogers,

2010). We use a machine with two 2.4 GHz Quad-

core processors and 8 GB RAM. The whole construc-

(a)

(b)

(c)

Figure 4: (a) Glyphs showing the distribution of Goethe

(red), Lessing (yellow), Schiller (blue), and Shakespeare

(green) at major locations in Germany. (b) Line segments

to localize sediment proﬁle generation. (c) Fused proﬁles in

split-mode, colored by time (left) and writer (right).

tion and user interaction is ﬂuent and respond times

after parameter changes are negligible.

6.1 Exemplary Analysis Process

The dataset in our ﬁrst scenario consists of documents

published by or dealing with four important writers,

namely Goethe, Schiller, Lessing, and Shakespeare.

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

166

The data was extracted from an online public ac-

cess catalog (OPAC) and contains 4.436 records that

spread over some hundred years. Each record is an-

notated with a writer, longitude- and latitude values,

and the publication year. Because the data is primarily

distributed over Germany, we limit our analysis to this

area. To get a ﬁrst impression about the data’s spatial

distribution, we choose a Gaussian ﬁlter radius that

is able to split features at city-level. Then we draw a

polygon around all features located in Germany and

we change its height so that prominent features are

assigned with a glyph to summarize the writer (class)

distribution. As shown in Figure 4a, the data clearly

concentrates in a few cities, and the most remarkable

insight extracted from the glyphs is the dominating

presence of publications by Goethe (red). The poten-

tial occlusion of hills and glyphs can be problematic

in 3-D. However, occlusions can easily be reduced by

rotating the whole scene, and, on the other hand, this

drawback is compensated by better local feature ac-

centuation if item count translates into hill elevation;

instead of large, aggregating glyphs on a 2-D map.

Because Goethe and Schiller represent the major-

ity of records in Germany, we use the class ﬁlter to re-

strict the landscape construction to both writers. Fur-

thermore, we use the elevation ﬁlter to remove small

features. To provide more details about the spatial,

temporal and class evolution of published documents,

we place a line segment at each major hill in order to

analyze their inner sediment structure. The scenario is

illustrated in Figure 4b. Note how the landscape color

already indicates where data items are old, young or

mid-aged. To compare all individual proﬁles at the

same time, we fuse them into one and use a sediment

partition by time with 20 sediments. Figure 4c shows

two versions of the fused proﬁle in split-mode. The

left one is colored by time and the right one is col-

ored by class. Colored triangles below both proﬁles

correspond to the line segments to indicate the cities.

The ﬁrst proﬁle clearly reveals two insights: In terms

of time, the dataset contains publications from the au-

thor’s lifetime until today (yellow), while most of the

documents appeared after their death (dotted lines).

In terms of location vs. time, data occurrence for both

writers follows their major places of activity during

their lifetime (Leipzig, Stuttgart), but spreads to other

cities after they died, most likely because many peo-

ple in capital cities dealt with their heritage (Berlin,

Munich, Frankfurt). The highest peak is Weimar,

where both writers lived and died. The second proﬁle

relates the writer to both location and time. It uncov-

ers that the publication places of Goethe and Schiller

primarily varied in the early years, but mixed later on.

To address the visual effect of using the presented

Figure 5: Sediment proﬁles for different color schemes and

with varying sediment granularity.

sediment color schemes at different granularities, we

concentrate on those documents by Schiller and Less-

ing that are located around Frankfurt and Stuttgart.

We use a larger Gaussian ﬁlter radius to abstract the

whole region instead of single cities. In all proﬁles

shown in Figure 5, the time aspect is encoded in the

order of the sediments. That is, even if sediments are

colored by class, we can still observe that the smaller

hill appeared later than the other one. Of course, the

relation between time and location is best reﬂected

if sediments are colored by time. But at ﬁner gran-

ularities, even this obvious coloring scheme can hide

spatial variance. This can be countered with a random

coloring. In our scenario, this reveals that especially

old data items occur at up to ﬁve locations, before

the two main hills eventually dominate. For coarser

granularities, colors for aggregated class and time are

interpolated. While this could result in invalid class

depiction, the indication of interesting relationships -

that might deserve further investigation using a ﬁner

sediment granularity - still works sufﬁciently well un-

til a sediment group contains very many items.

6.2 Overview Capabilities

The major advantage of the geological metaphor is

the manifold and detailed information conveyed by

inner sediment structure. However, we also want to

compare overview capabilities, i.e. how much static

information is provided without further user interac-

tion. We choose the Iraq war logs for this scenario.

The data was published by Wikileaks and consists of

around 60.000 records that describe incidents with at

least one casualty from 2004 to 2009.

The ﬁrst visualization, as presented by The

Guardian (Rogers, 2010), is shown in Figure 6a. Ev-

ery data item is represented by a small circle-glyph at

the incident’s location on the map. While it is easy to

AGeologicalMetaphorforGeospatial-temporalDataAnalysis

167

(a) (b)

(c)

Figure 6: Iraq war logs: (a) Single glyphs per data item,

courtesy of ’The Guardian’ (Rogers, 2010). (b) Aggregated

glyphs, courtesy of ’GeoTemCo’ (J

anicke et al., 2013). (c)

Our landscape metaphor with surface colored by time.

extract the data’s spatial extent, its spatial distribution

can not be determined reliably. Even though glyphs

occasionally appear as visually dense areas, reading

concrete item count is impeded by glyph overlap.

Conﬂict centers can thus only be guessed and glyph

color is furthermore wasted to indicate a single data

class. Figure 6b shows the visualization provided by

GeoTemCo (J

anicke et al., 2013). The authors use an

iterative process to combine overlapping glyphs un-

til each glyph is overlap-free. Final glyph size then

reﬂects the number of aggregated glyphs. Although

the visualization looks much clearer, there are still

some problems: If many data items occur at one sin-

gle location this results in a very big glyph; that grows

even more due to increasing overlap with neighbor-

ing glyphs. Therefore, to avoid having a Baghdad-

glyph (center of the map) covering the whole screen,

glyph size is logarithmized. This, however, impedes

the comparison of single items to glyphs that poten-

tially aggregate dozens of items; which happens near

Al Bukamal (middle left side) and is visible as a dense

area in Figure 6a, but hidden in Figure 6b. Glyph

color is again wasted for one single class.

Figure 6c shows the landscape metaphor for the

same dataset. Regions of higher item count are eas-

ily identiﬁed as outstanding hills. Because we do not

use 2-D glyph size, but 3-D hill elevation, feature

granularity is much ﬁner because item count space-

efﬁciently translates into height values; thus high-

lighting more distinct features on the same map. Fur-

thermore, we only need to decrease the Gaussian

ﬁlter radius to split hills at this zoom-level of the

map; while GeoTemCo has to zoom-in to split ag-

gregated glyphs and can not compare distant features

anymore. Most importantly, because we do not use

single-colored glyph-augmentation, we can utilize the

2-D landscape surface to convey time as a second data

aspect. The color distribution indicates that the major-

ity of all incidents happened at the data’s mean time in

2007 (green), while some old (blue) and young (yel-

low) regions stand out. At some places (lower right

side), hills are even white due to high standard devia-

tion. This indicates that incidents in this area occurred

in the beginning (2004) and at the end (2009) of the

conﬂict and suggests further analysis.

7 CONCLUSIONS

We presented a geological metaphor to illustrate,

compare and further investigate multiple queries of

different type in a geospatial-temporal context. Com-

pared to existing approaches, we focus on the depic-

tion of relations between multiple data aspects. Al-

though individual aspects could be illustrated with

higher accuracy in separated widgets, we opted for

a more abstract, but combined visualization of their

relationships because we think that, though at lower

granularity, these relationships are more important to

reveal interesting features that deserve further analy-

sis in a next step.

The geological metaphor provides means to con-

vey different information at varying granularity.

While the most summarized information of a data as-

pect is a scalar value that can be mapped to surface

color, ﬁner granularities are then provided as propor-

tions on glyphs and, ﬁnally, as sediment layers that

also consider chronological order. The intuitive sedi-

ment metaphor provides easy access to relate time, lo-

cation, class, and item count to each other at the same

time in one sediment proﬁle. Furthermore, we decou-

pled local feature accentuation and glyph granularity

from the map’s zoom-level. Although the Gaussian

ﬁlter radius could still be adjusted with the zoom, we

think its more important not to lose features on the

map, e.g. to compare distant cities. Glyph granularity

is additionally determined by (sub-)hill components

above the polygon.

We implemented the proposed metaphor as a sur-

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

168

rogate map-widget to complement other frameworks

by quickly indicating relationships between data as-

pects that could then be analyzed individually with

speciﬁcally tuned widgets.

ACKNOWLEDGEMENTS

The authors thank anonymous reviewers for valuable

comments and assistance in revising the paper. The

work presented in this paper was supported by a grant

from the German Science Foundation (DFG), number

SCHE663/4-1 within the strategic research initiative

on Scalable Visual Analytics (SPP 1335).

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.

Dent, B. D. (1999). Carography: Thematic Map Design.

McGraw-Hill, 5th edition.

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.

Eccles, R., Kapler, T., Harper, R., and Wright, W. (2008).

Stories in geotime. Information Visualization, 7(1):3–

17.

Fabrikant, S. I., Montello, D. R., and Mark, D. M. (2010).

The natural landscape metaphor in information visu-

alization: The role of commonsense geomorphology.

J. Am. Soc. Inf. Sci. Technol., 61(2):253–270.

anicke, S., Heine, C., and Scheuermann, G. (2013).

GeoTemCo: Comparative Visualization of

Geospatial-Temporal Data with Clutter Removal

Based on Dynamic Delaunay Triangulations. In

Csurka, G., Kraus, M., Laramee, R., Richard, P.,

and Braz, J., editors, Computer Vision, Imaging and

Computer Graphics. Theory and Application, volume

359 of Communications in Computer and Information

Science, pages 160–175. Springer Berlin Heidelberg.

Kapler, T. and Wright, W. (2005). Geo time information vi-

sualization. Information Visualization, 4(2):136–146.

Kim, S., Maciejewski, R., Malik, A., Jang, Y., Ebert, D. S.,

and Isenberg, T. (2013). Bristle maps: A multivari-

ate abstraction technique for geovisualization. IEEE

Transactions on Visualization and Computer Graph-

ics, 19(9):1438–1454.

Kraak, M. J. (1988). The space-time cube revisited from a

geovisualization perspective. Proceedings of the 21st

International Cartographic Conference, 1995.

Maciejewski, R., Rudolph, S., Hafen, R., Abusalah, A. M.,

Yakout, M., Ouzzani, M., Cleveland, W. S., Grannis,

S. J., and Ebert, D. S. (2010). A visual analytics

approach to understanding spatiotemporal hotspots.

IEEE Transactions on Visualization and Computer

Graphics, 16(2):205–220.

Roberts, J. C. (2007). State of the art: Coordinated multiple

views in exploratory visualization. In Proceedings of

the 5th International Conference on Coordinated Mul-

tiple Views in Exploratory Visualization (CMV2007).

IEEE Computer Society Press.

Rogers, S. (2010). The Guardian - Wik-

ileaks Iraq war logs: every death mapped.

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

2010/oct/23/wikileaks-iraq-deaths-map (Retrieved

2013-09-17).

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 Purves, R. and Weibel, R.,

editors, Proceedings of GIScience.

Shneiderman, B. (1996). The eyes have it: A task by data

type taxonomy for information visualizations. In Pro-

ceedings of the 1996 IEEE Symposium on Visual Lan-

guages, pages 336–343. IEEE Computer Society.

Slocum, T. A., McMaster, R. B., Kessler, F. C., and Howard,

H. H. (2009). Thematic Cartography and Geovisual-

ization. Prentice Hall Series in Geographic Informa-

tion Science. Prentice Hall, 3rd, international edition.

Tominski, C., Schulze-Wollgast, P., and Schumann, H.

(2005). 3d information visualization for time depen-

dent data on maps. 2010 14th International Confer-

ence Information Visualisation, 0:175–181.

Tominski, C., Schumann, H., Andrienko, G., and An-

drienko, N. (2012). Stacking-based visualization of

trajectory attribute data. IEEE Transactions on Visu-

alization and Computer Graphics, 18(12):2565–2574.

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

Design. Morgan Kaufmann, 3rd edition.

Wise, J. A., Thomas, J. J., Pennock, K., Lantrip, D., Pot-

tier, M., Schur, A., and Crow, V. (1995). Visual-

izing the non-visual: spatial analysis and interaction

with information from text documents. In Proceedings

of the 1995 IEEE Symposium on Information Visual-

ization, INFOVIS ’95, pages 51–, Washington, DC,

USA. IEEE Computer Society.

Xu, K., Cunningham, A., Hong, S.-H., and Thomas, B. H.

(2007). Graphscape: integrated multivariate network

visualization. In Hong, S.-H. and Ma, K.-L., editors,

APVIS, pages 33–40. IEEE.

AGeologicalMetaphorforGeospatial-temporalDataAnalysis

169