REGULAR PAPER

Jie Li • Zhao-Peng Meng • Mao-Lin Huang • Kang Zhang

An interactive visualization approach to the overviewof geoscience data

Received: 14 September 2015 / Revised: 9 November 2015 / Accepted: 10 December 2015 / Published online: 17 February 2016� The Visualization Society of Japan 2016

Abstract Geoscience observation data refer to the datasets consisting of time series of multiple parametersgenerated from the sensors at fixed locations. Although a few works have attempted to visualize such data,none of them views these data as a specific type and attempts to show the overview in all the space, time andattribute aspects. It is important for domain experts to select the subsets of interest from huge amounts ofobservation data according to the high level patterns shown in the overview. We present a novel approach tovisualize geoscience observation data in a compact radial view. Our solution consists of three visualelements. A map showing the spatial aspect is in the center of the visualization, while temporal and attributeaspects are seamlessly combined with the spatial information. Our approach is equipped with interactivemechanisms for highlighting the selected features, adjusting the display range, as well as interactivelygenerating a fisheye view. We demonstrate the usability of our approach with three case studies of differentdomains. Eye tracking records and user feedbacks obtained in a small experiment also prove the effec-tiveness of our approach.

Keywords Spatiotemporal visualization � Geovisualization � Geoscience observation data � Radial layout �Visual analytics

1 Introduction

Observation is one of the most common means for experts in different geoscience domains, such asatmosphere, ocean, environment, etc., to monitor the variations of natural environment or physical phe-nomena, as in Fig. 1. Due to the automatic and continuous observation capabilities, huge amount ofobservation data can be collected, indispensable for domain studies and result evaluation. In a typical usagescenario, an observation station is built at a representative location to obtain the long-term temporal

J. Li � Z.-P. Meng � M.-L. HuangSchool of Computer Software, Tianjin University, Tianjin, ChinaE-mail: jie.li@tju.edu.cn

Z.-P. MengE-mail: zpmeng@tju.edu.cn

M.-L. HuangE-mail: mlhuang@tju.edu.cn

M.-L. HuangSchool of Software, The University of Technology, Sydney, Australia

K. Zhang (&)Department of Computer Science, MS EC31, The University of Texas at Dallas, Richardson, TX 75080-3021, USAE-mail: kzhang@utdallas.edu

J Vis (2017) 20:433–451DOI 10.1007/s12650-016-0352-z

variation trend of multiple environmental parameters. A large amount of stations have been built throughoutthe world to obtain the data for overall distributions.

Though the techniques used in domain studies could well capture the evolutionary processes of spa-tiotemporal phenomena, they are costly and have other well-known limitations (Faghmous and Kumar2014). Furthermore, because a large number of domain problems have no reliable ‘‘ground truth’’ data, it iscrucial to develop objective evaluation methods for comparing the outcomes of different analysis tech-niques. We thus resort to visualization techniques, and attempt to design a visualization framework that canintuitively and interactively reveal the evolutionary nature of climate change patterns at different spa-tiotemporal scales.

Recently, an increasing number of scientists attempt to use information visualization techniques toanalyze such data in domain studies (Nocke et al. 2008; Qu et al. 2007; Drocourt et al. 2011). An overviewof the observation data is usually the start to help users quickly gain high-level insight into the data(Hornbæk and Hertzum 2011). Because the observation data are intrinsically multi-dimensional, time-oriented and geo-related, it is hard to show all the information facets in a compact view (Li et al. 2014a, b).In the visualization process, space, time and data attributes should all be considered together due to thestrong coupling of the three aspects. To comprehensively show the spatial, temporal and multi-dimensionalfeatures of observation data, visual designers often jointly use a map and other visualization techniques.These approaches depend on interactions and cross-reference among different representations, rather thanshowing all the data aspects in a single view.

We have previously developed a visualization approach to study climate changes, called Vismate (Liet al. 2014a, b), which includes three interrelated visualization views. Vismate focuses on analyzing climatechanges, without viewing the geoscience observation data as a specific data type, and evaluating its usabilityin other application domains. This paper focuses on the major enhancements to the main view of Vismate toassist users to quickly identify high-level spatiotemporal patterns of datasets and specify interesting subsetsfor further in-depth studies. The enhancements include three aspects:

• Interactive operations, such as highlighting interrelated objects between different visual elements, mapoperations typically used in GIS platforms, and focus?context views.

• Applications to the datasets of different domains to demonstrate its applicability. The new case studiesinclude air quality observation data and income statistics of 180 countries around the world (Extension tothe proceeding paper of Vinci’15).

Fig. 1 Four typical observation activities. a Meteorological observation. b Air quality observation. c Oceanographicobservation. d Fixed traffic flow monitoring

434 J. Li et al.

• A comprehensive usability evaluation, in which eye tracking data and user feedbacks are collected andanalyzed.

The remaining part of this paper is organized as follows. Section 2 reviews related work. The approachoverview is described in Sect. 3, followed the visualization technique, interactive design and implemen-tation details in Sects. 4, 5 and 6. Sections 7 and 8 describe three analytical case studies and a usabilityexperiment, respectively. Finally, we conclude the paper in Sect. 9.

2 Related work

2.1 Visualizing geoscience observation data

There exist numerous classical visualization techniques (Li et al. 2014a, b), such as contour line (Watson1992), standard coloration (Slocum et al. 2009), rose diagram, etc., used to represent several facets ofgeoscience observation data. These methods, however, are usually simple and can only show analysisresults, lacking interactions and support for discovering the potential knowledge from big observation data.

With the advance of display technologies, researchers have applied the state-of-the-art visualizationtechniques to domain specific studies. Qu et al. (2007) utilized observation data generated from multiple airquality observation stations to analyze the pollution sources in Hong Kong. Drocourt et al. (2011) designed avisualization to intuitively show the advance and retreat of calving glaciers based on the data generated fromthe observation stations distributed along the coastline of Greenland. Nocke et al. (2004) and Tominski et al.(2011) used multiple information visualization techniques to analyze meteorological observation data.Landesberger et al. (2012) visualized the cluster features of the weather data of 111 weather stations overGermany and Slovakia.

The temporal series of data generated from numerous sensors installed at fixed locations serve similarpurposes as the observation data. For example, Wang et al. (2014) used multiple transportation cells inNanjing (a city in South China) for macro traffic analysis. Guo et al. (2011) visualized the road intersectiondata captured by roadside laser scanners. Although the above works have resolved different domainproblems, they do not consider a proper data type and offer a generic and scalable approach that can be usedin multiple domains. These approaches either involve limited number of stations and temporal points, or areonly used to find the spatiotemporal patterns in a specified field. In contrast, we view geoscience observationdata as a new standalone data type and offer an overview visualization approach that can accommodatealmost any number of stations, cover any temporal span, and clearly show the boundaries with differentgeometric shapes of countries or regions.

2.2 Radial visualization

Radial visualization, or displaying data in a circular or elliptical layout, is a common technique in infor-mation visualization (Draper et al. 2009), which often encode time dimension by several concentric rings(Drocourt et al. 2011; Burch and Diehl 2008). However, one often ignores its advantage of representingorientation and position due to its non-directional shape. For example, Qu et al. (2007) designed an s-shapedaxis used in parallel coordinates to represent wind direction, and Malik et al. (2012) used multiple regularpolygons for comparative visualization. To improve the effective space usage, one could utilize the insideand outside of a radial visualization. For example, Draper and Riesenfeld (2008) put search inputs in thecentral space, while Wu et al. (2010) utilized the internal space to place a triangular ScatterPlot. A humanbody chart was laid in radar plot by Zhang et al. (2013) to visualize a person’s health condition. Burch andDiehl (2008) drew many thumbnails outside the outermost ring of TimeRadar. Inspired by these methods,our approach also uses a radial layout. We encode time dimension by different concentric rings along radialdirection with a map in the center, so that our approach can visualize spatial and temporal dimensions at thesame time. Moreover, the outside area of the radial plot is utilized to represent clustered information.

2.3 Space, time, and attributes visualization

Although many visualization techniques have been proposed to show space (Cressie and Wikle 2011; Freyet al. 2012), time (Faghmous and Kumar 2014; Javed et al. 2010) and multi-dimensional attributes(Inselberg and Dimsdale 1991; Carenini et al. 2014; Yuan et al. 2013) separately, showing numerous data

An interactive visualization approach to the overview of geoscience data 435

features in a single view is still a challenge for visual designers. To simultaneously show the space andattribute aspects, Kim et al. (2013) improved the classic thematic map by drawing short lines along the roadsin a map to represent spatial uncertainty, and called it BristleMap. Similarly, Gratzl et al. (2013) proposed anapproach to visually analyzing multi-attribute rankings and used it to visualize the world university ranking,showing time and attribute information together.

To simultaneously show the space, time and attribute aspects, most researchers attempted to jointly use amap and other basic visualization techniques. Malik et al. (2012) used map, bar chart, line chart and piechart to analyze the correlation between urban crime activities and spatiotemporal dimensions. Landes-berger et al. (2012) designed a dynamic categorical data view (DCDV) to visualize human position tran-sitions in 1 day, and associated it with a geography view. Parallel coordinates (Inselberg and Dimsdale1991; Johansson and Jern 2007; Tominski et al. 2004), ThemeRiver (Havre et al. 2000, 2002), stacked barcharts (Nocke et al. 2004), and perhaps all the existing time-series visualization techniques could becombined with a map to make effective spatiotemporal visualization. However, these loosely coupledspatiotemporal visualization methods do not provide an overview of spatial and temporal dimensions. Theylack intuitive and compact methods for visualizing spatiotemporal data in a single view. Different from suchmulti-view methods, our approach offers a global yet compact view, integrating both spatial and temporalfeatures.

3D visualization is another way for simultaneously showing space, time and attribute information.Tominski et al. (2012) and Andrienko et al. (2014) proposed a stack-based trajectory wall method forexploring multiple trajectories in a 3D context. Two similar methods, Great Wall of Space–Time (Tominskiand Schulz 2012) and Space–Time Cube (Kraak 2003), have been proposed by different researchers. Theseapproaches are, however, only suitable for trajectory data visualization and have inherit the drawbacks of 3Dvisualization, such as clutter, overlapping, and slow interaction.

In summary, although various approaches have been used to visualize multi-dimensional spatiotemporaldata, few works are capable of showing the overview of space, time, and attribute information in a compactand intuitive form. This is the main motivation of our work in this paper.

3 Approach overview

3.1 Data characteristics

One of the primary characteristics of geoscience observation data is their multiple types by nature. First,data are continuously generated from observation stations having fixed spatial coordinates, therefore aregeo-related and time-oriented. The data values only represent the temporal variations of environmentalparameters at the locations where the data are generated. Data generated at a station can be viewed as a timeseries, while multiple time series generated at different stations with specified geographic coordinates formthe spatial dimension. By selecting and analyzing multiple time series in the spatiotemporal range of aphenomenon, scientists can better understand the variation process of the phenomenon and investigatefurther. Second, records of the same data structure are output from observation devices at a fixed temporalinterval, each often containing multiple parameters. Multi-dimension is therefore another characteristic ofgeoscience observation data. Finally, geoscience observation data are inherently hierarchical, because eachrecord is associated with an observation station positioned in a hierarchically administrative area. Forexample a country can be divided into multiple administrative levels, such as province (or state), city,county, etc. Table 1 illustrates typical attributes and the corresponding data characteristics of geoscienceobservation data, although the parameters of different types of such data vary greatly.

Table 1 Characteristics of geoscience observation data

Attribute Data characteristic

Station code, continent, country, province, county HierarchicalLongitude, latitude, altitude SpatialDate, time TemporalPrecipitation, air pressure, wind speed, air temperature, temperature anomaly, sunshine hours Multi-dimensional

436 J. Li et al.

3.2 Visualization requirements

‘‘Overview first, zoom and filter, then details-on-demand’’ is a widely accepted principle of visual infor-mation-seeking (Shneiderman 1996). We focus on how to effectively generate the overview of a givendataset consisting of big observation data, for which intuitively visualizing all the characteristics is animportant criterion. Our objective is to help analysts better understand the variations of attributes (A) withrespect to the spatial (S) and temporal (T) dimensions (Andrienko and Andrienko 2006). We classify themost important requirements into three major categories:

• Exploration Show attribute (A) distribution at different spatiotemporal scales. This requirement isimportant for analysts to identify different spatiotemporal patterns, such as constant, gradual or abruptchanges, outliers, and repetitions in space and time. Because the analysts may use our approach as thepreliminary step of data analysis to select interesting subsets from a target dataset, our approach shouldaccommodate the data of the overall S and T.

• Retrieval This requirement contains two aspects, i.e. identifying the S and T of particular spatiotemporalpatterns (A ? S ? T) and detecting the attribute distribution of a subarea or a subinterval (S ? T ? A).To support the bidirectional visual reasoning, our approach should be able to highlight the specifiedattribute values and demonstrate the attribute distribution at a selected spatiotemporal scale.

• Comparison Compare the attribute distribution in different S and T, or in different subsets of the dataset.Our approach allows users to interactively generate the visualization by adjusting the parameters.

3.3 Design decisions

To detect geoscience objects, scientists attempt to locate interesting areas and identify the correspondingtemporal profiles generated at the stations in the areas, which are characteristic for the processes understudy. This is a challenging task due to the volume of the data and the complexity of the temporal three-dimensional structure of the objects under study. Even for a local area on land surface, the data compriseshuge amounts of time series. To address this challenge, scientists often aggregate the time dimension bycomputing a single statistical measure for each temporal profile in the selected dataset. Although themeasure is selected based on the expert knowledge and provides an overview of the corresponding temporalprofile, the aggregation inevitably results in loss of information. To understand the detailed temporalvariations, scientists have to select a small number of stations for a detailed analysis. Important details maypossibly be missed due to the limited number of the selected stations.

This paper introduces a visual analytics approach that provides both an overview and detailed temporalvariations. We believe that full awareness, investigation, and understanding of both time and spatial patternsat the same time is extremely important, and therefore design our approach to visualize spatiotemporalpatterns in a compact view. As in Fig. 2, the framework of our approach consists of:

• A map in the center conveying the geospatial information.• A ring-band outside the map, encoding time series changes.• Multiple concentric cluster rings outside the ring-band, each showing a cluster generated by a clustering

algorithm over multiple attributes of the dataset.

Three highly interrelated components are used to encode spatial, temporal and clustering information.All the stations having spatial coordinates are drawn (as black dots) on the map (see Fig. 2), and also on thecircumferences of rings at equal angular intervals representing spatiotemporal clusters according to aselected measure (see the colored dots on the rings in Fig. 2). We use a line to connect the correspondingdots on two components, along which a ring-band encoding the observation values at each time point isplotted. To avoid overlapping, the line is placed under the ring-band. We do not make the line semi-transparent, since it will cause adverse effects of unintended color blending. This arrangement aims toeffectively condense a large number of widely distributed stations in a single view, while emphasizing timeseries and clustered changes. Through interactive operations, such a layout facilitates the exploration andcomparison of different stations, regions and clusters over time. The detailed descriptions of the threecomponents are illustrated in Sect. 4.

An interactive visualization approach to the overview of geoscience data 437

4 Visualization components

4.1 Map

To embed a map into a circle, the map’s bounding box should fit into the circle. Using the traditionallongitude/latitude equidistance grid projection, a world map could be projected onto a rectangle with thewidth twice of its height, which would waste much space, as shown in Fig. 3(left). We use the Mercatorprojection (Snyder 1987) featured with shape-preserving after converting a world map to a square within aspecified bounding box (longitude: -180� to 180�, latitude: -85.05� to 85.5�), as in Fig. 3 (right). Thephenomenon of area enlargement becomes more serious as latitude increases, and the regions higher than85.05� and lower than -85.05� cannot be mapped. Such effects on our visualizations are, however, limiteddue to the area-preserving feature of the Mercator projection and few stations in such a high-latitude area.

To accommodate any geographic shapes of different countries, map drawing has to compute the map’scentral coordinate. Drocourt et al. (2011) used a modified standard formula to compute the map gravitycenter of Greenland, however, it is more effective to use the bounding-box’s center as the map center.Furthermore, to facilitate manual browsing of different subareas of a map and distinguish the attributevalues of target areas, our approach also supports two types of interactive operations, to be discussed inSect. 4.3. To keep the map style intuitive and minimize the viewer’s cognitive effort, we use the online toolColorBrewer (Brewer and Harrower 2009) with a recommended color solution to assign a color to each area.The color solution is also used to color the stations the same as the areas they belong to.

Fig. 2 Scheme of our approach consisting of three visual elements, i.e. a map, a sector-based ring band and several clusterrings. This visualization shows the climate changes in China during the period of 1981–1989

438 J. Li et al.

4.2 Cluster rings

The stations are divided into multiple clusters using the K-means algorithm (Hartigan and Wong 1979),although our framework does not depend on any specific clustering algorithm. Each cluster ring (see Fig. 2)represents a cluster. The thickness and color of a ring indicate the value and direction (positive or negative)of the corresponding cluster, respectively. Using Fig. 2 as an example, pink and grey represent the warmingand cooling trends, respectively, while a ring’s thickness represents the absolute value of the averageclimate variation trend of the cluster. The K-Means clustering algorithm (see Sect. 4.3) divides stations intoseveral clusters that are drawn on the corresponding cluster rings. To clearly show the overall state, thecluster rings are also sorted outward on average values of the clusters.

4.3 Sector-based ring band

We use a sector-based ring band to encode the parameter values of all stations over a period of time.Outward direction indicates the time axis (see the Sector-based ring band in Fig. 2). Each sector indicatesthe time series of one station, while a radial bin is colored to represent the attribute value at a time point. Anintuitive and rational color legend is the key to temporal mapping. We design the color legend according tothe domain research standard and also support color filtering (Tominski et al. 2012) to allow the user toemphasize or de-emphasize selected intervals, as in Fig. 4. This effect can be triggered by clicking on thecorresponding interval on the color legend. For easy understanding of our visualization, we use the colorscheme based on the existing industrial standard.

4.4 Station mapping

In the basic view, all the stations with (longitude, latitude) Cartesian coordinates are mapped to (h; r) polarcoordinates, which creates extra space to represent time and clustering information. Using a fixed angularinterval, h of a station can be determined by simply adding an equal angular interval to the angularcoordinate of the previous station. Therefore, obtaining the angular coordinate of the first station is the keyto this step. To avoid two geographically distant stations to be too close to each other on the ring, we use ahierarchical mapping strategy. We first divide a map into several subareas and assign an angular span toeach subarea in proportion to the number of the stations in that area, as in Fig. 5. Then the stations areprojected in each area onto the ring according to a reference. Longitude, latitude, and any an attribute can beselected as the reference according to analysis tasks. The purpose that we use a line to connect the centermap and the dots on the cluster rings is to obviously show the subarea division. We select latitude as thereference for determining the sequence of the stations in a subarea, while other references can also be chosenupon the domain requirements. Parameter r of the polar coordinate of a station can be obtained by deter-mining the cluster ring that the station should be drawn.

(-180,-90)

(180,90)(-180,90)

(180,-90)

R

AfricaAsia

South America

Antarctica

North America Europe

Oceania

(-180,85.05) (180,85.05)

(-180,-85.05) (180,-85.05)

Fig. 3 Comparison of traditional longitude/latitude equidistance projection (left) and Mercator projection (right)

An interactive visualization approach to the overview of geoscience data 439

5 Interactivity

To clearly display the detailed information of any area with different geometric shapes, our approachprovides three types of interactive operations.

5.1 Highlighting interrelated visual elements

Simultaneously showing the spatial, temporal and attribute aspects of a selected subset is a common taskwhen performing analysis tasks. We therefore design an interactive mechanism to help the user simulta-neously understand the three aspects of the selected subset. The three visual elements of our approach (Map,Sector-based Ring Band, and Cluster Ring), respectively, show space, time and attribute aspects of a givendataset. When the user focuses on an area of one of the three visual elements, our approach can highlight thecorresponding areas on the other two visual elements. For example, when the mouse hovers on a circle onthe cluster rings, the name of the station represented by this circle appears at the mouse location and thesector for showing temporal information and the geographic location on the map is also highlighted on theother two visual elements. Similarly, when we put the mouse on any other element, the corresponding areason the remaining two elements are also highlighted. Furthermore, if the mouse hovers on a cluster ring, allthe corresponding stations are also highlighted to show the overall distribution.

5.2 Polar coordinate based fisheye

Although the user can quickly find the areas of all the three aspects by highlighting interrelated visualelements, it is still a challenge to accurately judge the color and location of each geographic shape ondifferent elements. This is especially difficult when the view accommodates too many stations or covers along temporal interval. Therefore, a focus?context model is used to generate the fisheye view to enlarge thetarget area while maintaining the contextual information. Because our approach has a radial shape, thefocus ? context model is based on polar coordinate system, which features shape-preserving for map

Fig. 4 Color filtering effect. The width of the selected interval of legend (see Fig. 2) is decreased for reducing observationload

Fig. 5 An example of station mapping which contains 4 subareas and 2 cluster rings

440 J. Li et al.

transformation (Jakobsen and Hornbæk 2011). Figure 6a shows an example of the fisheye view, in whichXinjiang province is selected as the focus area.

When a focus point is selected, each point of the three visual elements (Map, Sector-baser Ring Bandand Cluster Ring) x; yð Þ is transformed to the fisheye coordinate xfeye; yfeye

� �by re-computing the distance d

Fig. 6 Interactive operations supported by our approach. a Polar coordinate system based fisheye view. b Interactiveoperations used in GIS, such as Zoom in, Zoom Out, Pan, etc

An interactive visualization approach to the overview of geoscience data 441

between the focus point xfocus; yfocusð Þ and this point. After obtaining the new distance, fisheye coordinatescan be computed using a simple trigonometric function. The three visual elements use the same transfor-mation equation with different distortion factors (Yang et al. 2003). For map transformation, each geo-graphic vertexes move toward or away from the focus point along the line connecting the focus point andthis geographic vertex (angle is unchanged), while both angular interval and radius interval are changed forthe Sector-based Ring Band and Cluster Ring. Keeping the focus point, old vertex and new vertex along aline could maximally retain the geometric shape of each map feature.

5.3 Interactive operations of GIS platform

Apart from visualizing the datasets generated from a fixed area, our approach is generic, such as showingover a long time with any number of stations, which may be generated from any countries or regions of theworld. Inspired by GIS platforms, we use several interactive operations, such as pan, zoom-in, zoom-out,etc., to operate maps containing any countries and areas, as in Fig. 6b. Through these interactions, userscould flexibly interactively explore various geographical areas at any scales, which is particularly useful forvisualizing big data obtained from a huge amount of stations.

In summary, our approach provides three complementary interactive operations to support seamlessexploration.

6 Implementation

The current implementation of our approach is written in WPF (Windows Presentation Foundation), uti-lizing the client–server architecture. The server program is responsible for loading datasets and performingautomatic clustering, while the client program presents the visualization and handles user interactions.Separating the computationally intensive tasks can minimize the CPU and memory usage of the client-sidecomputer, and the client program can smoothly run on a laptop.

Our approach uses the K-means (Hartigan and Wong 1979) clustering algorithm, while any otherclustering algorithm can be selected according to the domain requirements. To support clustering, we usethe slope of the linear regression line of a station as the clustering reference. Let X = {x1, x2,…, xN} be theset of time points in the selected time interval, Y = {y1, y2,…, yN} the set of attribute values at thecorresponding time points, and �x and �y the average values of X and Y. The slope is calculated as follows:

slope ¼PN

i¼1 xi � �xð Þ � yi � �yð ÞPN

i¼1 xi � �xð Þ2ð1Þ

7 Case studies

This section reports the evaluation of our approach through three case studies of different domains.

7.1 Meteorological observation data

China surface meteorological observation data was used to the evaluation. This dataset contains the long-term observation records of 206 stations distributed throughout the country. Within the China meteoro-logical observation network, these stations could automatically and continuously obtain 23 types ofmeteorological attributes on the land surface. The dataset covers 1951–2012, containing about 4 milliondaily average records, 135,000 monthly average records and 47,000 yearly average records. The entire sethas been checked for consistency by the meteorological authority, and thus is reliable to use.

We first divided China into seven areas using the customary geographical division method in climatestudies, and assign a color to each area. Furthermore, a Chinese meteorological industry standard (Zhang2009) is used to define the color codes of multiple meteorological attributes, as depicted in Fig. 2.

Using the example of Fig. 2, the interval from 1981 to 1989 is recognized as the period when China’sclimate change happened. We divided the stations into nine clusters according to the temperature changerates using the K-means algorithm. The fact that most of the stations are on the pink rings indicates that thewarming trend is a national phenomenon during that period. By observing the sector-based ring band, wefind the bins on the outer part of the band are dark red and a bluish ring is in the middle of the band. These

442 J. Li et al.

findings illustrate the national warming trend in China began in the mid-1980s and there was no significantclimate change before then. Our findings are consistent with the research results of Ren et al. (2012).

We continue to analyze the overall climate change in the winter during the period of 2001–2012. Whenall the stations were mapped on the cluster rings, spatiotemporal patterns were clearly shown, as in Fig. 7a.Beyond our expectation, all the cluster rings showed cooling trends (grey colors). By observing the colorfiltering view of the sector-based ring band, we found the warming process mainly appeared in the first halfof the 2000s. Therefore, we selected the period of 2001–2008 as the target interval and generated thevisualization again, in which almost all the cluster rings have reddish colors, representing a strong nationalwarming trend, as in Fig. 7b.

Seasonal climate changes are also of interest to researchers. Although the seasonal change patternsbefore 2000 have been published (Sun et al. 2011), few have analyzed the seasonal variation patterns inrecent years. To fill this gap, we used our approach to visualize the temperature anomalies in differentseasons, as in Fig. 8. Warming trend in the winters was very strong, especially in Northeast China (see thedark red sectors in Fig. 8d). Spring and autumn also had obvious warming trends, and in East China suchtrend was even more obvious than in the winter. The warming trend in the summer was the weakest. Wefound that the overall seasonal variations in the 2000s were the same as the prior years. Climate academicsconfirmed our findings by stating that seasonal climate changes have stable characteristics and do notsuddenly change.

Fig. 7 Climate changes in different interval of the 2000s

Fig. 8 Seasonal climate changes during the period of 2001–2012

An interactive visualization approach to the overview of geoscience data 443

7.2 Per capita income data

To verify the generality of our approach, we visualize the per capita income data of 180 countries around theworld during the period of 1990–2009 (see http://bost.ocks.org/mike/nations/).

We use the same color coding scheme as in Fig. 3, while nine clusters are established according to theincome variation rates of the countries during the entire interval. Pink and grey cluster rings represent theincome increasing and decreasing, respectively. Each bin of the sector-based ring band is colored accordingto the ratio of the income anomaly of that year to the average income of the entire interval. Red bins indicatethat the income of that year was higher than the average income, and blue bins the opposite. We candetermine the temporal variation trend of a country by observing the distribution of its bins.

Most of the countries show income increasing trends, particularly obvious in Asia (see the dark redregion in Fig. 9). Europe also has a strong increasing trend, and the fact that few decreasing cases appear inthis region represents its relatively balanced and stable development. On the contrary, the economicdevelopments of the countries in Africa are uneven. The bin distribution in Equatorial Guinea is distinctive,which represents an extreme case. The literature records that at the beginning of the 1990s, the discoveriesof large oil fields made the economic development of Equatorial Guinea rapid, and the GDP has beenincreasing from about one hundred million to over 15 billion, which proves our findings. The similar rapiddevelopments can also be found in China and Ireland (see the selected red ring in Fig. 9), highlighted in thecenter map. The United States also has a stable increase in per capita income. However, the subprime crisisoccurred in 2009 affected this trend, reflected by the lighter color of its bin distribution. There also existmultiple countries having decreasing trends, such as North Korea and Somalia. The distributions of thesecountries have opposite variation (from red to blue), and the colors of the corresponding cluster rings aregrey.

Fig. 9 Visualization of the per capita incomes of 180 countries during the period of 1990–2009. Cluster 8 is selected, and thecorresponding countries are highlighted on the map

444 J. Li et al.

7.3 China air quality observation data

China’s Ministry of Environmental Protection and Environmental Monitoring Centre Station has operated anational air quality observation network containing thousands of observation stations throughout thecountry. These stations can continuously observe and record air qualities at a fixed temporal interval (1 h)and automatically transfer these data to the specified data centers. Through an online data service, wecollected the observation obtain air quality observation data of 846 stations in each hour.

Because the locations of all the stations are not released, we aggregate these stations according to thecities they locate. In a preliminary data analysis phase, we find the stations in one city have almost the samechanging trends, as in Fig. 10. Therefore, we aggregate the data of all the stations in a city using the Averageaggregation function, and perform visual analysis using a city as the smallest unit.

We divided China into seven areas and assign a color to each area using the same method as in Sect. 5.1.The parameter AQI (air quality index) represents the overall pollution status of different cities, derived fromthe other parameters of the observation data (Ministry of Environmental Protection 2012). A smaller AQIvalue indicates a better air quality of a city.

We select the smog data and meteorological data in December 2013, during which smog happenedfrequently and seriously. Since variation trends in a month is obvious, each cluster ring only utilizes itswidth to represent the absolute value of the AQI, while all the cluster rings are in the same color. As inFig. 11a, we divided the cities into 9 clusters based on the average AQI. We found that the numbers of citiesin different clusters are almost identical and the average AQIs of all the clusters are higher than 50, implyingthat smog in China is a nationwide phenomenon.

By selecting the rings that represent the cluster of the most serious smog, we find most of the cities inthose clusters are in the west of Hebei province, suffered smog during the entire month. East China isanother area suffered from smog, while the smog in Southwest, South and Northwest China are relativelymoderate. Inspecting the time series of all the cities in East China, we find the extreme smog occurred onlyat the beginning of the month.

To fully understand the smog status in Beijing-Tianjin-Heibei region, which is the economic andpolitical center of China, we selected that region as the focus area and generated a fisheye view, as inFig. 11b. As the regional center, Beijing has high population density and industrial production activity.Smog in this city is, however, not as serious as its surrounding cities. One reason may be that the small citiesaim to rapidly improve their GDPs and indiscriminately build new factories. To the contrary, big cities havebetter urban planning and invest more on environmental protection. Similar cases can also be found inShanghai and Guangzhou regions.

Due to the lack of observation stations in central and western China, the analysis in many areas cannot beperformed, but the overall smog distribution is clearly shown in Fig. 11. In general, smog in China isserious, the pollution gradually moves from the east to the west, and the coastal cities have the worst airqualities. These patterns are consistent with the regional economic development activities, proving thatsmog is most likely caused by industrial activities.

Fig. 10 Daily-average AQI of 4 stations in Beijing

An interactive visualization approach to the overview of geoscience data 445

Fig. 11 The global distribution of the air quality observation data in December 2013. a General view. b Fisheye view in whichHebei province is selected as the focus area

446 J. Li et al.

8 Experiment

8.1 Objectives and tasks

Having verified the feasibility of our approach through the case studies, we conduct a usability experimentto evaluate the effectiveness and interactivity of our approach. Our approach is embedded into multi-viewvisual analytics tool Vismate, presented in our previous paper (Li et al. 2014a, b), which does not report anyexperiment on its usability, nor applications in other domains. This section analyzes the influence of eachindividual visual element on the overall cognitive effect, and the effectiveness of embedding our approachinto an analysis tool consisting of multiple views for domain users. The experiment results could also guideusers to better set visualization parameters to meet their application needs.

To perform the experiment, we proposed three interrelated tasks regarding the climate changes in EastChina, which require comprehensive uses of all the three visual elements:

• Find the cluster containing most of the stations of East China.• Identify two turning points of climate changes in East China.• Determine the stations having different variation trends.

In addition to the routine experimental parameters, such as accuracy and completion times, as animportant means of evaluating visualization techniques, eye tracking has also been utilized in the experi-ment. Eye movements are recorded continuously throughout the visualization task and can provide insightinto the process of working with a visualization environment (Kurzhals et al. 2014). Therefore, the mea-surement of spatiotemporal eye movement data may be more diagnostic than summative experimentalparameters. To analyze the cooperative effects of the four views (our approach is one of them) in the tool,we set six AOIs (Areas of Interest) on the interface, each on a visual element (Map, Sector-based Ring Band,Cluster Ring), as in Fig. 12. By observing the direct transitions of the subjects between the AOIs, multipleusability patterns can be identified.

8.2 Preparation

8.2.1 Stimuli

As in Fig. 12, the tool used in the experiment contains 4 views, in which our approach offers the overview(AOIs 1, 2 and 3), while other three views are specialized in finding turning points (AOI 4), analyzing

Fig. 12 GazaPlot with 6 AOIs. Each subject is mapped to a different color generated by the integrated eye tracking softwareand all gaze trajectories are overlapped

An interactive visualization approach to the overview of geoscience data 447

temporal variation trends (AOI 5) and detecting anomaly cases (AOI 6). The three AOIs marked on ourapproach are on the three visual elements, respectively, each covering the corresponding area of the threeelements of East China. To support interactive operations, an executable program with the predefinedparameters was used, which was shown to each subject after a calibration procedure. We used the dataset ofCase Study 1, selected the period of 2001–2012 as the target interval, and set 9 cluster rings.

8.2.2 Subjects

We chose a within-subjects study design with 10 subjects. Eight subjects are students from the School ofComputer Software, Tianjin University, and two others have art backgrounds. All the subjects were graduatestudents and had experiences in visualization. Seven of them were male and three were female, aged24.6 years on average (between 19 and 31). None of the subjects had used our approach. All the subjectswere confident with mouse and keyboard interaction and had heard about global warming.

8.2.3 Environment

All the trials were conducted in a laboratory environment during a vacation to minimize distractions fromoutside. We used a Tobii T60 XL eye tracking system with a TFT screen resolution of 1920 9 1200 pixels.The fixation duration is chosen to be 60 ms and the fixation size 10 pixels on the eye tracking software.

8.2.4 Trials

Each subject conducted a trial on all the three tasks, resulting in a total of 30 trials.

8.3 Procedure

We first gave the subjects a brief introduction to this tool, and then spent 10–15 min training them on usingthree sample programs specialized in the four views. We also answered the questions raised by the subjectsto ensure the experimental environment to be functional and the subjects capable of completing the tasks.Subjects’ accounts were created in the eye tracking system with age, major, gender, and prior knowledge invisualization being selected as independent variables. Any technical problems during the training procedurewere solved before the experiment started.

At the beginning of the experiment, the eye tracker was calibrated to the subject’s eyes using a 5-pointcalibration. Then, the test program was shown to each subject. To keep the subjects facing to the eyetracking screen, the experimental administrator recorded the solutions they orally pronounced, and thenclicked any key to switch to the next screen. The accuracy and completion time of each question weremanually recorded. Though there was no time limit for the tasks, the subjects were instructed to completethe tasks as quickly and accurately as possible. Emphasis on a fast solution would potentially results in higherror rates and possibly chaotic gaze trajectories, because the subjects would be forced to guess an answer,which was not the intention of this eye tracking study.

8.4 Results

The overall accuracy of task completion was 93.33 % and each session (a subject) including three trials took70–120 s. The high accuracy is in accordance with our exceptions, since the parameters were carefullyselected to ensure the visualization patterns to be obvious. Due to the ceiling effects of the accuracies, wemainly focused on analyzing the exploration behaviors of all the subjects.

The GazePlot of the experiment is shown in Fig. 12. By observing the distribution and transitions offixations, we can evaluate the role of our approach among all the four views. To quantitatively analyze theGazaPlot, we also generated a statistic table consisting of two important metrics, as in Table 2. The resultsare summarized as follows:

• The importance of our approach in the tool is obvious. The subjects rely heavily on this view to completedomain related tasks, because it (AOIs 1, 2, and 3) has more fixations than other three views.

• The frequent fixation transitions between our approach and other three views prove the coherentrelationships of the former to the latter. However, other three views depend on the overview and needcross-referencing to determine the spatial and temporal distributions of a station, which is clearly shown

448 J. Li et al.

in our approach. In contrast, other three views have a loose coupling relationship, since there are fewtransitions between them (transitions between AOI 4 and AOI 5 are also frequent, because both of themare used for temporal series analysis.).

• AOI 1 has the smallest fixation count and the shortest duration among the three AOIs of our approach.This implies that the subjects mainly use other two regions to complete the tasks and the function of thecenter map needs to be improved.

• Among the three AOIs of our approach, AOI 2 has the most fixations, indicating the sector-based ringband is critical for users to perform different types of domain research.

8.5 User feedback

An open-ended discussion session was held with all the subjects after the experiment. We encouraged thesubjects to talk about their experiences with the experiment, and took notes on their feedbacks.

The feedback of the subjects was mostly positive. All subjects answered that the interaction wasbeautiful and easy to learn. Furthermore, they also agreed that our approach was the most important of allthe views they use and was convenient for them to explore the space, time and clustering information. Thedesign of the cluster ring was interesting and useful for them to view the spatiotemporal distributionaccording to the clustering results. Several subjects highlighted the particular importance of interactions,especially the polar coordinate based fisheye. Although the fisheye view has an asymmetric shape, none ofthe subjects had difficulties using it.

Most subjects thought that our method was preferably used in combination with other views to form adomain related analysis tool. The capability of accommodating almost any numbers of stations in differentspatiotemporal scales makes our approach be considered a data reference to select objects having differentoverall features. Since all the views are interrelated, through interactive operations, selected objects can beimported into other views for more in-depth analyses, such as temporal trend, anomaly detection, etc.

The subjects also made constructive suggestions for improvements. Four subjects advised us to add afunction of temporarily locking the highlighted cluster rings, since it is difficult to use the mouse to preciselyfollow a ring with small thickness. Three subjects mentioned that the angular range of an area sometimesdoes not strictly cover the region on the map when the differences in the numbers of stations are big. Wehave used a hierarchical station mapping strategy (see Sect. 4.1) to minimize this effect. It is, however,impossible to completely solve this problem when the number of stations in each subarea is disproportionalto the angular span of that region. With this ratio less constrained, more spatial information during stationmapping (see Sect. 4.1) will be lost. Further research is needed, for example, we may add an interface formanually dividing area and setting angular interval of each subarea We can also use an adjustable angularinterval, and the region has more stations will have a smaller angular space, of course, this mechanismscarifies the symmetry. One subject mentioned that our approach could also be used to show geo-relatedstatistical data, such as yearly economic data in different countries of the world. In fact, our visualizationframework could visualize any type of multi-dimensional spatiotemporal data.

In summary, both the eye tracking data and the user feedbacks confirm the effectiveness and importanceof the proposed radial approach in analyzing geoscience observation data.

9 Conclusion and future work

We have developed a novel visualization approach offering insight into geoscience observation datasimultaneously with geo-related, time-oriented and multi-dimensional features. By integrating spatial and

Table 2 Eye tracking results of AOIs

AOI Mean fixation duration (s) Mean fixation count

AOI 1 0.23 10.25AOI 2 0.33 60.25AOI 3 0.37 42.00AOI 4 0.32 44.75AOI 5 0.27 27.50AOI 6 0.29 20.00

An interactive visualization approach to the overview of geoscience data 449

temporal displays into a compact view, our approach supports exploratory spatiotemporal analysis. Ourvisualization design is based on a radial layout, in which three visual elements are harmoniously embedded,aiming at showing the high level characteristics of the entire dataset. Furthermore, three types of interactionshave been integrated in our solution to help the user clearly observe any details of the visualization. Theusefulness and usability of our solution have been demonstrated by several cases studies and a usabilityexperiment. Having verified the patterns found in case studies, we consider our approach to be effective anduseful in real-world scenarios. In the future, we plan to improve our approach in two aspects. First, we willconduct a thorough evaluation environment by using more visualization techniques provided by eyetracking. Second, we attempt to make our approach more scalable to support a huge number of observationstations. Finally, we plan to apply our approach in more visual analytics tools to fully test its effectiveness inreal scientific research of different domains.

Acknowledgments The authors wish to thank Zhao Xiao and Huan Yang for their discussions. We are also grateful to theanonymous reviewers for their insightful comments that have helped us in improving the final presentation. This work wassupported by Tianjin ‘‘Big Data Algorithms and Applications’’ project (13CZDGX01099).

References

Andrienko N, Andrienko G (2006) Exploratory analysis of spatial and temporal data: a systematic approach. Springer Scienceand Business Media, Berlin

Andrienko G, Andrienko N, Schumann H, Tominski C (2014) Visualization of trajectory attributes in space–time cube andtrajectory wall. Cartography from Pole to Pole, pp 157–163

Brewer CA, Harrower MA (2009) ColorBrewer 2.0: color advice for cartography. http://colorbrewer2.org/, Penn StateUniversity

Burch M, Diehl S (2008) TimeRadarTrees: visualizing dynamic compound digraphs. Comput Graph Forum 27(3):823–830Carenini G, Conati C, Hoque E, Steichen B, Toker D, Enns J (2014) Highlighting interventions and user differences: informing

adaptive information visualization support. In: Proceedings of the 32nd ACM Conference on Human Factors inComputing Systems, pp 1835–1844

Cressie N, Wikle CK (2011) Statistics for spatio-temporal data. Wiley, USADraper G, Riesenfeld RF (2008) Who votes for what? A visual query language for opinion data. IEEE Trans Vis Comput

Graph 14(6):1197–1204Draper G, Livnat Y, Riesenfeld RF (2009) A survey of radial methods for information visualization. IEEE Trans Vis Comput

Graph 15(5):759–776Drocourt Y, Borgo R, Scharrer K, Murray T, Bevan SI, Chen M (2011) Temporal visualization of boundary-based geo-

information using radial projection. Comput Graph Forum 30(3):981–990Faghmous JH, Kumar V (2014) Spatio-temporal data mining for climate data: advances, challenges, and opportunities. Data

mining and knowledge discovery for big data, pp 83–116Frey S, Sadlo F, Ertl T (2012) Visualization of temporal similarity in field data. IEEE Trans Vis Comput

Graph 18(12):2023–2032Gratzl S, Lex A, Gehlenborg N, Pfister H, Streit M (2013) Lineup: visual analysis of multi-attribute rankings. IEEE Trans Vis

Comput Graph 19(12):2277–2286Guo H, Wang Z, Yu B, Zhao H, Yuan X (2011) TripVista: triple perspective visual trajectory analytics and its application on

microscopic traffic data at a road intersection. In: Proceedings of IEEE PacificVis’11, pp 163–170Hartigan JA and Wong MA (1979) Algorithm AS 136: A k-means clustering algorithm. Appl Stat 100–108Havre S, Hetzler B, Nowell L (2000) Themeriver: visualizing theme changes over time. In: Proceedings of the IEEE InfoVis

2000, pp 115–123Havre S, Hetzler E, Whitney P, Nowell L (2002) ThemeRiver: visualizing thematic changes in large document collections.

IEEE Trans Vis Comput Graph 8(1):9–20Hornbæk K, Hertzum M (2011) The notion of overview in information visualization. Int J Hum Comput Stud 69(7):509–525Inselberg A, Dimsdale B (1991) Parallel coordinates. In: Proceedings of Human-Machine Interactive Systems, pp 199–233Jakobsen MR, Hornbæk K (2011) Fisheye interfaces—research problems and practical challenges. In: Proceedings of Human

Aspects of Visualization, pp 76–91Javed W, Mcdonnel B, Elmqvist N (2010) Graphical perception of multiple time series. IEEE Trans Vis Comput

Graph 16(6):927–934Johansson S, Jern M (2007) GeoAnalytics visual inquiry and filtering tools in parallel coordinates plots. In: Proceedings of the

15th Annual ACM International Symposium on Advances in Geographic Information Systems, pp 33Kim SY, Maciejewski R, Malik A, Jang Y, Ebert DS, Isenberg T (2013) Bristle Maps: a multivariate abstraction technique for

geovisualization. IEEE Trans Vis Comput Graph 19(9):1438–1454Kraak MJ (2003) The space–time cube revisited from a geovisualization perspective. In: Proceedings of the 21st International

Cartographic Conference 1988–1996Kurzhals K, Fisher B, Burch M, Weiskop D (2014) Evaluating visual analytics with eye tracking. In: Proceedings of the Fifth

Workshop on Beyond Time and Errors: Novel Evaluation Methods for Visualization, pp 61–69Landesberger T, Bremm S, Andrienko N, Andrienko G, Tekusov M (2012) Visual analytics methods for categoric spatio-

temporal data. In: Proceedings of the IEEE VAST’12, pp 183–192

450 J. Li et al.

Li J, Meng ZP, Zhang K (2014) Visualization of oceanographic applications using a common data model. In: Proceedings ofthe 29th ACM Symposium of Applied Computing (SAC’2014), pp 933–938

Li J, Zhang K, Meng ZP (2014) Vismate: interactive visual analysis of station-based observation data on climate changes. In:Proceedings of the IEEE VAST’14, pp 133–142

Malik A, Maciejewskiet R, Jang Y, Huang W (2012) A correlative analysis process in a visual analytics environment. In:Proceedings of the IEEE VAST’12, pp 33–42

Ministry of Environmental Protection (2012) HJ633-2012. People’s Republic of China National Standard-TechnicalRegulation on Ambient Air Quality Index (on trial). Beijing: China Environmental Science Press

Nocke T, Schumann H, Bohm U (2004) Methods for the visualization of clustered climate data. Comput Stat 19(1):75–94Nocke T, Sterzel T, Bottinger M, Wrobel M (2008) Visualization of climate and climate change data: an overview. In: Digital

earth summit on geoinformatics, pp 226–232Qu H, Chan WY, Xu A, Chung KL, Guo P (2007) Visual analysis of the air pollution problem in Hong Kong. IEEE Trans Vis

Comput Graph 13(6):1408–1415Ren G, Ding Y, Zhao Z, Zheng Z, Wu T, Tang G, Xu Y (2012) Recent progress in studies of climate change in China. Adv

Atmos Sci 29:958–977 (In Chinese)Shneiderman B (1996) The eyes have it: a task by data type taxonomy for information visualizations. In: Proceedings of IEEE

Symposium on Visual Languages, pp 336–343Slocum TA, Mcmaster RB, Kessler FC, Howard HH (2009) Thematic cartography and geographic visualization, 3rd edn.

Pearson EducationSnyder JP (1987) Map projections—a working manual (US geological survey professional paper 1395), United States

Government PrintingSun JQ, Wang HJ, Yuan W (2011) Decadal variability of the extreme hot event in China and its association with atmospheric

circulations. Clim Environ Res 16(2):199–208Tominski C, Schulz HJ (2012) The great wall of space–time. In: Proceedings of Vision, Modeling and Visualization,

pp 199–206Tominski C, Abello J, Schumann H (2004) Axes-based visualizations with radial layouts. In: Proceedings of the 2004 ACM

Symposium on Applied Computing (SAC’04), pp 1242–1247Tominski C, Donges JF, Nocke T (2011) Information visualization in climate research. In: Proceedings of the International

Conference on Information Visualization (IV’2011), pp 298–305Tominski C, Schumann H, Andrienko G, Andrienko N (2012) Stacking-based visualization of trajectory attribute data. IEEE

Trans Vis Comput Graph 18(12):2565–2574Wang Z, Ye T, Lu M, Yuan X, Qu HM, Yuan J, Wu Q (2014) Visual exploration of sparse traffic trajectory data. IEEE Trans

Vis Comput Graph 20(12):1813–1822Watson DF (1992) Contouring. Pergamon PressWu Y, Wei F, Liu S, Au N, Cui W, Zhou H, Qu H (2010) OpinionSeer: interactive visualization of hotel customer feedback.

IEEE Trans Vis Comput Graph 16(6):1109–1118Yang CC, Chen H, Hong K (2003) Visualization of large category map for Internet browsing. Decis Support Syst 35(1):89–102Yuan X, Ren D, Wang Z, Guo C (2013) Dimension projection matrix/tree: interactive subspace visual exploration and analysis

of high dimensional data. IEEE Trans Vis Comput Graph 19(12):2625–2633Zhang M (2009) Color code for products of weather forecast and service GB. Chinese meteorological industry standard (in

Chinese)Zhang Z, Wang B, Ahmed F, Ramakrishnan IV, Zhao R, Viccellio A, Mueller K (2013) The Five W’s for Information

Visualization with application to healthcare informatics. IEEE Trans Vis Comput Graph 19(11):1895–1910

An interactive visualization approach to the overview of geoscience data 451