Today, cartography products are mostly connected to large volume datasets. These datasets have to be evaluated and visualized correctly, and in the context of digital cartography, the heat map method has gained popularity. Heat maps can help us explore big data sets to visually identify single instances or clusters of important data entities . The method became popular with cartographers as well as users. Common usage of the heat maps raises questions about the correct settings and interpretation of the method, etc. This article deals with the usage of the heat map method, its visualization and technical aspects, suitability of the method for various datasets, and a description of the method itself. The aim of the paper is to create an overview of the heat map method from different perspectives in its correct use and full understanding.
During the one-year IGA 2017 project (Internal Grant Agency project of Palacký University Olomouc, Czech Republic), a complex analysis of heat-maps on different spatial datasets was created. The suitability of heat maps was analyzed via four case studies in diverse subjects: traffic accidents, elections, spatial placement of libraries, and socio-economic themes. This article is focused only on spatial data from traffic accidents in Olomouc, Czech Republic from 2011. Compared to the conventional cartographic and geoinformatics methods such as choropleth maps or proportional symbol map, heat-map could provide alternative impression on examined phenomena. Due to even spatial distribution, traffic accident data are best fitting case for heat-map implementation by Geographical Information System. Heat-map as a dynamic visualisation method allows multiple interpretation depending on individual user skills and method settings. Therefore, the exploratory study aims on both cartographical educated and non-educated respondents. Survey findings and the results of the exploratory study lead to a set of recommendations for both cartographically educated and lay users. Based on recommendations we can evaluate the method from a user perspective.
2 Heat map method
Since no exact rules or precise definitions of heat maps exist, widespread use of the method has created different interpretations of spatial data. The term heat map has been used only recently in the context of digital cartography and the rise of map mashups. According to Dempsey , the heat map is “a method for showing the geographic clustering of a phenomenon”. Yeap and Uy  describe heat maps as “geospatial data on a map using different colors to represent areas with different concentrations of points — showing overall shape and concentration trends”. From a technical point of view, it is a visualization of the areas of influence of each point and further summation in places where areas overlap. The color gradient represents the power of influence at a certain point. For a non-cartographer user, the map is attractive, easily readable and the visualization is more comprehensible. Meteorological maps use maps in the visible spectrum where red represents higher temperature and cooler colors lower temperatures. Different schemes can be used for different datasets . Many methods allow estimation, for example, of traffic accident data to identify hotspots (point density, kernel density, closest neighbor). According to Ivan and Horák  or Anderson , kernel density estimation is one of the main methods for identifying hotspots. This leads to visualization using the heat map method.
Software designer Cormac Kinney is considered the pioneer of heat maps . He first used the method and the term heat map in 1991. The use was for real time information about financial markets. In this case, the values were represented by grey to black pixels. The heat map usually works with a very large data matrix. It visualizes results using cluster analysis that uses permutation of columns and rows to place similar values close to each other based on clustering. A similar system of color coding is used in fractal maps and tree maps. The need to create rules of application and software for effective map creation came about, and in 1994, Leland Wilkinson created the first software called Systat which could create a heat map with large color resolution .
Currently, heat maps are widely used in different scientific disciplines for data presentation and exist in different visualization forms (color ranges, basemaps and opacity, software implementation). According to Trame and Kessler , heat maps let us explore large datasets and visualize important cases or clusters. The heat map has been used for many years in visualizations for biological and statistical analysis. Lopez-Bigaz and Parez-Llamaz , for example, offer the open-source tool Gitools. This tool imports data from various biological databases for analysis and visualization with heat maps and compares the experimental data with hundreds of other analyses. Moon et al.  used heat maps to visualize the evaluation of steroid hormones — quantitative indicators can be visualized using heat maps based on hierarchical cluster analysis.
Communicating transportation safety to the general public is an essential and challenging task. When visualised and interpreted correctly, these data may help to change local policies, driver behaviour and specific traffic signs. Examples of web-based GIS mashups are presented by Hilton et al. . The article reports that there is a strong interest in such tools mainly because of the customisation and possibility to explore local data. Another approach based on GIS modelling is described by Li et al. . In this work, authors created a line-based model calculating relative crash risk which was then visualised using 2.5D graph. The topic is expanded in Kušta et al.  using traffic intensity and animal activity data predicting the possibility of vehicle collisions. The paper proves that increased locomotory activity of-of wild animals leads to increased number of animal-vehicle crashes while the intensity of traffic has an only little effect. Plug et al.  researched another related topic of single vehicle crashes. According to the authors, these cause over 80% of all fatal car crashes. The paper regards both spatial and temporal patterns of the phenomenon. Results are visualised using various methods including circular graph and point-based kernel density estimation. The latter of the two is used to identify locations of high crash intensity. The same method but using line-based visualisation on point data is used by Benedek et al. . Authors use a large dataset of traffic crashes including injuries. Road traffic accident data including all mentioned before were analysed by Soltani and Askari . In their work, authors focus on spatiotemporal analysis in the city of Shiraz. However, authors concluded that the trends vary according to incident type. To summarise, the topic of traffic safety is essential for society. GIS provides a wide range of tools for data analysis and visualisation in this field. Kernel density estimation was successfully used for identification of spatial patterns in many of these studies.
In geoscientific fields, heat maps are deployed “to find the density of houses, crime reports, or roads and utility lines influencing a town or wildlife habitat” . Generally, crime statistics, any type of accident or socioeconomic aspect (population or price distribution, election results, emotional maps) are popular subjects that can be visualized with a heat map. Its popularity is based on fast and intuitive interpretation, because the resulting density surface is visualized using a gradient that allows the areas of highest density to be easily identified . By contrast, easy identification is associated with a certain amount of subjectivity. Ivan and Horák  characterize the basic disadvantage of a heat map as the subjectivity of it results. According to Dempsey’s study , “the end visualization which affects how data is interpreted by the viewer is a subjective one.” With the boom of internet-based visualization platforms, heat map methods are also employed in subjects to do with tourism, sports activity, media, or social-networks.
3 Heat maps in Geographic Information Systems
Geographic Information Systems (GIS) are suited to the analysis and visualization of map outputs via various cartographic methods . GIS provide effective tools for identifying hotspots in order to understand spatial processes such as traffic accidents. Using suitable data types, data sources, and appropriate software are crucial steps in applying the heat map method.
3.1 Input data types for visualization by heat map
Generally, data are the primary foundation for creating any cartography output. According to Longley et al. , GIS handles three main data groups — spatial data, attributes data, and metadata. This paper focuses on the visualization of map outputs using spatial data only. This chapter briefly introduces the basic spatial structures (points, lines, and polygons) related to heat maps.
Points in two-dimensional space (2D) do not have width, height, or depth. They are therefore described as zerodimensional objects. This geography element is represented only by two coordinates (x, y) in 2D and is too small to be visualized as a line or polygon. In the real world, points mainly represent height points, settlements, vegetation etc. . Approximately 80 - 85% of the sources referred to in this work use point data. Therefore, we can say that point data are the primary input data for heat map implementation. Possible application fields for point-based heat maps are crime-rate maps, emotional maps  or natural disasters such as forest fires or earthquakes.
A line is a spatial data type having width, but not height or depth. It is therefore considered a one-dimensional object. A line can also be described as a series of points (vertexes) belonging to the original line, each of them described by two coordinates. In geography, they may represent rivers, road networks, or borders . Visualizing lines as a heat map is less frequent than when using point data. Heat maps can be used to visualize traffic volume, public transport lines, or data from sport apps such as Endomondo, Strava, or Runkeeper , see Figure 1.
A polygon has both width and height, but doesn’t have depth. Therefore, they are talked about as twodimensional objects. It is a closed object with a border comprising a series of points – lines. It may represent, for example, water bodies, buildings, or forest areas . Visualizing polygons with heat maps is the least frequently used application. Interesting examples are visualizing buildings by age, risk maps, flood danger, sea water salinity, commuting distance, or election results .
The main problem with a polygon-based heat map is that in most cases an uncertain type of choropleth map is created. Choropleth maps visualize a phenomenon represented by relative data recalculated to an area (e.g., countries, regions, watersheds, counties, or census blocks). This method works with quantitative data and allows the user to compare different areas by showing the spatial variability of the phenomenon . Normalization of data in choropleth maps is usually done for an area. In the case of socio-economic variables, normalization can be done by other quantitative characteristics of the area (e.g., 1000 infected / 100,000 inhabitants); it is therefore a visualization of relative data within existing borders. If a normalized choropleth map uses a divergent color range (red-yellow-green is typical for a heat map), it is still a choropleth map. By contrast, polygon-based heat maps handle absolute data. This method, however, usually leads to poor perception of information by the user and is not used frequently. If data are not converted to a usual area unit and new units, areas, and borders subsequently arise independently, a dasymetric method should be used. In fact, heat maps capture a certain intensity of a phenomenon through color transition in areas with no basis for spatial comparison of individual sub-units. It only shows distribution [22, 23].
3.2 Softwares and tools for generating heat map
In addition to data and visualization aspects, the software environment can affect the quality of map results. A subjective comparative overview of GIS was performed before the exploratory study focusing on heat map parameter settings.
3.2.1 Esri products (ArcGIS for Desktop, ArcGIS Pro, ArcGIS Online)
The Esri company, a leading producer of GIS solutions, provides diverse products dependent on the platform. ArcGIS for Desktop is currently being replaced by the enhanced and more powerful ArcGIS Pro. ArcGIS Online is a platform for creating and sharing GIS projects via the Internet environment. Each of these Esri products allows heat map visualization.
Three tools are available that can create heat maps in ArcGIS for Desktop: Kernel Density, Line Density, and Point Density. The tools can be found in ArcToolbox - Spatial Analyst Tools - Density. The Kernel Density tool works with both point and line data. Based on the input data, it calculates the size of a territorial unit using a kernel function. The result is a smooth curved surface on each line or point. Line/Point density can be used only on the respective type of data . The difference to the previous method is that it calculates the size of a territorial area from lines/points in a given diameter around each unit. Output raster is divided into 9 intervals visualized using a convergent one-color range. This can be changed to a classified scale, stretched color gradient, or discrete color range. Transparency can be also applied to the results.
ArcGIS Pro offers similar possibilities for heat map creation with a few improvements. The main one is that a heat map visualization can be chosen directly in the layer properties. The basis of the visualization is similar to Kernel Density. ArcGIS Pro allows the weight of an attribute to be set up, the radius entered, rendering quality selected, and a color scheme applied, customized, and saved. ArcGIS Online offers the same tools with limited settings for the color schemes of point data. For line data, the settings are much broader and include line styles and much more detailed color schemes. Unlike previous solutions, ArcGIS Online offers also heat map options for polygon data. All products allow the value of a radius to be set (by default in map projection units), and predefined color ranges (color schemes) selected, see Figure 2.
3.2.2 Quantum GIS
Quantum GIS (QGIS) is a free, open source GIS application allowing users to visualize, manage, edit, and analyse data. In terms of a customized experience and user-friendly interface, Quantum GIS offers an intuitive “heat map” module. The input layer can be a point layer only. It offers many output formats including GeoTIFF. The diameter setting offers map scale units and absolute units – metres. The possibility to set up transparency and color schemes is similar to ArcGIS Desktop, see Figure 3.
The QGIS application can be enhanced by calling scripts from SAGA GIS and GRASS GIS. The GRASS GIS v.kernel function generates a standard raster density map from vector point data using a moving kernel. It can also generate a vector density map on a vector network in a special network mode. The v.kernel function provides seven kernel density functions: uniform, triangular, Epanechnikov, quartic, triweight, gaussian, and cosine. SAGA GIS has a Kernel Density Estimation module (KDE). KDE allows the user to only use quartic and gaussian kernel functions. Both modules can be customized for radius.
3.2.3 Libraries and web tools
One of the other tools which can be used for heat map or kernel density creation is the python library heatmap.py available with Python 2.5+. The tool provides multiple settings such as dot size parameter or opacity and the size of the resulting image. Results can be visualized using five unique color schemes .
HeatmapTool.com is a tool for quick, online visualization of point data as a heat map. The tool has a great range of settings which can be used to customize the given result. It has eight color schemes as well as the option to set up your own, and settings for radius, heat map opacity, and feathering. HeatmapTool.com only needs the URL of a CSV file with spatial data and is recommended for use with Google Maps API. Google Fusion Tables generates heat maps from spreadsheets into only two map outputs (feature map vs. heat map) and only has basic functionality.
Python Matplotlib is another common tool used for visualization. It enables certain methods of heat map creation, with multiple approaches to achieve them. These include the imshow function or histogram2d function. In language R, a heat map can be generated using the ggplot function. Both packages permit extensive customization limited only by the range of possibilities in its languages. In contrast to the web tools and special libraries mentioned above, these functions are not designed for such specific visualization techniques and customizing them is much less intuitive and more complex for the casual user .
4 Color visualization aspects
Since a heat map is a subjective cartographical method, some rules should be followed. The choice and setting of a color range is a priority parameter that reflects overall readability, a reader’s orientation, and (in)correct interpretation of a heat map. Color is one of the most important aspects for both analog and digital maps . It is a fundamental part of all cartographic methods and map elements. Color carries both the information and aesthetic function of the map. This is why the authors of this study emphasize color aspects.
4.1 Color ranges in Cartography and GIS
Two data types are distinguished in thematic cartography and GIS – quantitative data, showing the quantity (regardless of whether it is absolute or relative), and qualitative data, which acquires specific values of the phenomenon . Quantitative data are expressed in a quantitative range and qualitative data in a qualitative range (see Figure 5).
Quantitative color ranges are further divided into convergent and divergent scales. These ranges show the quantity of the effect by means of brightness and color saturation. The convergent range shows the increasing intensity of the phenomenon. In most cases, they are monochromatic (the transition from lightest to the darkest) or multicolored. The divergent range displays the data in a certain interval with a specific (break) value, from where the positive values are on one side and the negative values on the other. These color ranges are further divided by the location of the break value as symmetrical and asymmetric ranges . A qualitative color range shows data that gain specific values for the phenomenon, for example, political divisions of the world. Currently, ColorBrewer 2.0 and Sequential Color Scheme Generator 1.0 are the only two generators available that allow a certain color range according to cartographic rules to be created [30, 31].
According to Voženílek and Kaňok , two crucial principles for the right quantitative range exist: 1) the higher the intensity of the phenomena, the higher the intensity of the color (less desaturation of color as typical mistake [28, 32] - see Figure 6), and 2) opposing phenomena (negative/positive) have different, two-color tones.
Overview of 140 heat map studies
|Color range||Overview||Number of cases|
4.2 Heat map color ranges
In this study, 140 examples of heat map implementation were evaluated. Only digital maps from sources across the world were collected. Thematically, most (110+) of the gathered data sources targeted socio-economic themes — especially public services, transport, economic matters, social-networks, and accidents, accumulated many times over at multiple sources. This was another partial reason why the study examines traffic-accident data. Only few examples displayed natural subjects — typically natural disasters.
A variety of color ranges was used in evaluated examples. The choice of color (range) should always correspond to the data they visualize according to cartographic semiology. From a cartographic point of view, color range was not suitably chosen for the given data visualization in many cases. For example, environmental data were displayed by a color range typical for socio-economic data and vice versa. Based on the evaluation of 140 heat maps, the three most-used color ranges were identified – divergent symmetrical, divergent asymmetrical, and convergent monochromatic. Generally, the most used were divergent ranges, 38 examples used divergent symmetric ranges of green-yellow-red or an extended version of violet-blue-green-yellow-red, see Figure 7.
A common problem was identified – the application of a symmetric range without a break/average value. This resulted in incorrect data interpretation and user-misleading visualization. In many cases, a convergent monochromatic scale should be chosen for accurate interpretation of a phenomenon. From evaluating the examples, the first assessment is that heat map authors are putting more emphasis on user-friendly color ranges instead of cartographically correct visualizations.
5 Exploratory study
5.1 Study description
During the one-year IGA 2017 project (Internal Grant Agency project of Palacký University), a complex analysis (based on four different datasets, such as traffic accidents, elections, spatial placement of libraries, and socio-economic topics) of heat maps was created. This article describes an exploratory study focusing only on Traffic Accident data.
The survey was completed by 69 respondents (28 women and 41 men; age mean 26 years) including 42 cartographers (14 women and 28 men) and 27 non-cartographers (14 women and 13 men). The male/female ratio is not similar because of the predominance of male students in Geoinformatic and Cartographic courses. Based on the Dreyuf model of skill acquisition (expert/novice)  respondents were divided into two groups. Respondents with at least a bachelor’s degree in Cartography or Geoinformatics (mostly from Palacký University in Olomouc) were considered as Cartographers (“carto”). General users with no higher cartography skills were considered as the group of non-cartographers (“non-carto”) . 56% of non-carto respondents was less than 25 years old. User´s time for the questionnaire was 22-26 minutes, followed by 10-15 minutes for the think-aloud interview.
The priority of the questionnaire and survey was to determine the preferences of individual parameter settings during the creation and interpretation of heat maps. The evaluation was based on a comparison of the answers between cartographic and non-cartographic respondents. The essence of the comparison was the confirmation or rebuttal of the assumption of a significant variation in non-cartographic respondents’ answers. The deviation was thought to lack the specialist knowledge of cartographic principles in the creation of monitored heat map parameters.
5.1.2 Stimuli and experiment procedure
The first part of the study looked at color range diversity. Six different color ranges were created. All color ranges had the same settings: 1 : 90 000 scale, 20 px radius and a layer transparency of 25%. User color range preferences were verified on three convergent monochromatic ranges (used to illustrate the number of traffic accidents) and three divergent symmetrical ranges (as a comparison scale with the maximum, average, and minimum accident values). Three tasks were allocated to an area of interest using color range.
The second part focused on creating four different visualizations based on the layer transparency. The scale (1 : 90 000), color range (divergent symmetrical: green-yellow-red), and radius (20 px) settings were identical, whereas the transparency varied between 0% - 25% - 50% - 75% values. One task was allocated to an area of interest using transparency.
The third part of exploratory study examined a different value for the radius. Four heat maps with different radii (10 - 20 - 30 – 40, unit of map projection) but with the same scale (1 : 90 000), the same transparency (25%), and the same color range (convergent monochrome) were prepared. Three tasks were allocated to an area of interest using radius.
Three assumptions were stated before the research began:
- Assumption #1Cartographic education will affect personal perceptions, especially in color (range) selection
- Assumption #2The transparency will not affect a respondent’s perception
- Assumption #3The radius will significantly affect a respondent’s perception
5.1.3 Input data and software
This exploratory study handles point layers of Traffic Accidents, provided by the Olomouc Region Fire Services (6 000+ points of accidents in shapefile format; spatial delimitation: the Olomouc city area; time delimitation: year 2011) . The exploratory study is focused and divided into three areas of interests: color range, transparency, and radius.
Based on the previous (subject) software comparison, ArcGIS Pro software was used to visualize the heat maps of all datasets. ArcGIS Pro meets the required general (legend, title, basemap, scale) and specific cartographic (color, transparency, radius) aspects in a fast and user-friendly environment. It is also not necessary to postprocess map outputs made by ArcGIS Pro with any other graphics software.
As a fundamental part of the research, great emphasis was placed on a primary survey. The entire research took place in the Cognitive-lab at the Dept. of Geoinformatics, Palacký University, Olomouc. A combination of questionnaire and think-aloud methods were used. The questionnaire was completed digitally by GoogleForms framework, followed by think-aloud interviews for deeper understanding of user preferences and cognitive strategies. The questionnaire was divided into five parts. The first section focused on the respondent’s basic information (name, age etc.), the following four sections were divided according to the datasets used (= four case studies).
5.2 Traffic accidents in Olomouc
To access three Areas of Interests (AoI), the following tasks were given to all respondents in the Traffic Accidents exploratory study.
Task #1 (AoI Color range): Choose your preferred colors for representing traffic accidents
The importance lies in the rule of cartographic semiology and semantics, which primarily matches character or color relationships to the content of what they designate [18,21]. Typically, red for fires, blue for water etc. The priority was to find out whether respondents perceive this relationship or just choose colors they liked. The respondent had a choice of 10 colors, see Figure 8.
The choice of colors is typical for the correct interpretation of traffic accidents and the answers were diverse. The color preference from carto respondents is mainly red (35 respondents), then orange (19), purple (12), pink (12), black (11), etc. For complete results see Figure 9. The obvious choice of red and orange highlights that the respondents chose the color according to the accident’s intensity. As think-aloud confirms, they subconsciously assign red and color to the places with the highest intensity of any obstacle. Compared to red, green ranked sixth, and would be interpreted as places with the lowest intensity of accidents. The reason behind choosing purple and pink was based strictly on individual response, not cartographic skill. The color-scheme in the non-carto group is quite different, except the most preferred color: red (20), black (19), grey (19), brown (17), orange (8), etc.
Task #2 (AoI Color range): Sort the color range according to the best visualization of traffic accident distribution
This task emphasized the chosen color transition and the color range type (convergent / divergent). According to results in Figure 12 it is obvious that the first two rankings were the most advanced of the range E (convergent monochrome, wine color range). The second most frequently chosen range was range A, (also convergent monochrome, with pink colors). Range B placed third (divergent symmetrical, green-yellow-red/min-avg-max). During the think-aloud interview, most respondents mentioned a significant relationship between task #1 and #2. The choice of color range strictly corresponded to the choice of preferred colors in the first task. It followed a general, publicly adopted semantic rule: (dark) red indicating places with a higher level of danger, and green colors or lighter tints of red indicating safer areas.
Range B, the most used public heat map color range (see Figure 12) was preferred by non-carto respondents. The next ranges were C (divergent symmetric scale, blue-white-red) and F (divergent symmetric scale, violet-yellow-dark red). The selections from non-cartographic respondents suggests that the lay public strictly prefer divergent ranges with an average value compared to the convergent scales based on hue with only one color. This result corroborates a previous thesis that cartographic education affects personal perception.
Task #3 (AoI Color range): Which map represents quantity and which map compares the incidence of traffic accidents?
The third task deals with identifying color ranges based on the recognition of the type of scale. Three divergent symmetrical scales for (comparison) and three convergent monochrome (for quantity) were used. Whether the respondent perceived the difference in concept of interpretation can be deduced from the results. In all cases, both cartographers and non-cartographers, the individual scales were correctly classified.
Task #4 (AoI Transparency): Sort heat maps according to their readability
Four different layer transparencies (0% - 25% - 50% - 75%) on heat maps were shown to respondents. Respondents from the carto group preferred layers with 50% transparency (C) and 25% transparency (B); the 0% transparency (A) was placed last by most respondents, see Figure 14. Non-cartographic respondent responses were generally similar. However, they confirmed that 0% transparency (A) is not a suitable solution for heat maps. Task #4 of our exploratory study refutes Assumption #2. Respondents confirmed that an appropriate level of transparency (50%) enables easy map readability, while opaqueness does not allow orientation on the base map at all. The upper value preference of transparency is quite individual: a third of respondents were satisfied with 75% transparency during think-aloud, some preferred 25% instead of 75%, regardless of cartographic education.
Task #5 (AoI Radius): Sort heat maps according to their readability (relevant to the point layer of traffic accidents)
Default scale of each map was set to a 1 : 90 000. In the context of the radius setting and with the aim of correctly interpreting traffic accidents, the responses of both cartographic and non-cartographic respondents in task #5 were the same in all cases. The most relevant radius settings were a radius of 20 (B) and a radius of 10 (A). Both groups selected the highest radius 40 (D) as least favoured, see Figure 16.
Task #6 (AoI Radius): Sort heat maps according to their readability (relevant to the scale of traffic accidents layer)
Radii values in heat maps change with each scale level! In fact, radii fluctuate during zoom-in or zoom-out movements. The reader receives different outputs in different scale levels from an identical dataset. This is why more than one task focusing on the radius were given to respondents during the exploratory study. Tasks #5, #6, and #7 focused on AoI radius among diverse scale levels.
Task #6 has a scale of 1 : 50 000 and radii A –10, B – 20, C – 30, and D – 40. Radius 30 (C) and 40 (D) were chosen as the least relevant due to incorrect interpretation. The carto group preferred a radius of 20 (B) and then 10 (A), while the non-carto group preferred the same, but in reverse order, see Figure 18.
Task #7 (AoI Radius): Sort heat maps according to their readability (relevant to the scale of traffic accidents layer)
Similarly to the previous task, #7 deals with the same radii (A –10, B – 20, C – 30 a D – 40), but at a scale of 1 : 130 000 (zoomed-out). Respondents from the carto group mainly selected radius 20 (B) as their first choice. A surprising choice could be seen in second place (radius C – 30), as that interpretation was judged to be less appropriate during think-aloud interviews. The non-carto group selected radius 10 (A) as the best interpretation and a radius 20 (B) as the second. According to Figure 20, both groups did not prefer the higher radius values 30 (C) and 40 (D).
In this chapter, we summarize basic tips for correct heat maps implementation via GIS. The first part is a list of generally established recommendations, with the purpose of cartographically correct interpretations of a certain phenomenon. The second part is focused on user preferences, confirmed by respondents’ answers in our exploratory study. As this paper focuses only on traffic-accident data, the whole project comprising four studies and diverse data provides a complex analysis of heat map implementation. The results are intended for cartographers and non-cartographers alike interested in heat map visualization with GIS. It allows the strengths and weaknesses of heat maps based on real study to be understood. It provides basic and specific recommendations, including recommended settings, especially for users not able to prepare comparative studies (non-professional cartographers, state administration employees, students). The authors assume the results will be disseminated by professionals and students for the purposes of education.
6.1 General recommendations
Thematically, the choice of visualized topic is not restricted, but the choice of appropriate datasets is a priority for successfully visualized heat maps. Confirmed by the evaluation of 140 heat map examples, around 80% of heat maps were generated from a point layer. The exploratory study confirms that the heat map method is recommended for creating heat maps from point datasets. Due to possible confusion with choropleth maps, we do not recommend creating heat maps from polygon datasets. According to Esri , “use heat map symbology when many points are close together and cannot be easily distinguished.” It is crucial to mention whether data describe the qualitative or quantitative (regardless on absolute or relative units) attributes of a phenomenon. The overwhelming majority (94.3%) of examined samples have been used for quantitative phenomena. It is essential to know whether a heat map is intended as a fast preview to attract the user’s interest only, or further data analysis is expected. From user’s point of view, a heat map is an attractive kind of spatial data visualization. On the other hand, it is not useful for detecting accurate values because of its cartographical principles. Based on the results of the exploratory study, certain rules can be suggested:
- Heat maps are recommended for a fast-preview of spatial data
- Heat maps are recommended as a method for identifying min/max hotspots (not min/max values)
- Heat maps are recommended with maximum caution (according to the user’s target group!) for comparing diverse phenomena only if the same parameters are set (radius, color range)
- Heat maps are strictly not recommended as a method for detecting accurate values
The main cartographical expression of a heat map is the color range. According to cartographic semiology, colors should be related to the data they display [18,21]. This rule is governed by the choice of a green color for a nature, grey/black for networks, red/orange for fire or accidents, specific colors of political parties, etc. Generally, a typical color for a certain phenomenon allows faster interpretation by the reader. The correct radius setting is individual, mainly related to the scale of the map. The intensity of a phenomenon (that extends a certain distance from it) should not be too unambiguous, but at the same time should not merge into one large object without boundaries. The basic function of transparency is readability of both the topographic background and thematic content.
6.2 User preferences in traffic accident data
User preferences for carto and non-carto groups were set for different heat map settings according to the survey (chapter 5.1). Respondents from the carto and non-carto groups both preferred a red color scheme. While the next best colors sequences were orange-violet-pink-black for the carto group, non-carto users selected neutral colors such as black-grey-brown and then orange as their fifth choice. The reason for red/orange is based on the understanding of the interpretation of traffic accidents, where areas with the highest incidence of traffic accidents are generally recorded as red . Red could be generally recommended for negative data topics (dangers, crime, obstacles etc), while green or muted colors are more suitable for positive phenomena.
The exploratory study examined whether convergent monochrome ranges (amount of the phenomenon) or divergent symmetric ranges (min-avg-max value of the phenomenon) are more user-friendly. Color range preference between carto and non-carto groups is diverse. For carto educated users, a convergent monochromatic range (e.g. violet and pink colors, see Figure 21) is recommended because of professional skills and cartographic knowledge such as choropleth map rules, which was mentioned during think-aloud. Cartographers look at the value of a phenomenon as the progressively increasing intensity of traffic accidents. This is why experts prefer a gradually increasing saturation. Emphasis on an appropriate legend is relevant to convergent ranges — for example, white or very light colors for minimum values combined with transparency and light backgrounds could lead to the incorrect presumption that the minimum value is marked as light red, whereas the minimum value in the map may be invisible to a reader (minimum is located outside the visible heat map overlay).
The public perceive heat maps in another way. Color range from green to red indicates the greatest threat of traffic accidents in each location. Common users are strongly accustomed to the most divergent symmetrical range, see Figure 21. As most of the research examples and the exploratory study confirm, the most implemented color range is a transition from a minimum (green), through an average (yellow), to a maximum (red) value. A divergent range is far more user-friendly for the lay user. It is popular in a user-friendly and quick-to-read interpretation, the color design being related to a certain threat present in a large number of traffic accidents, red representing maximums while areas with less frequent traffic accidents are shown in a positive, green color.
According to the same opinions from cartographers and non-cartographers, transparency around 50% is the most appropriate and most feasible option in which both thematic contents and topographic bases are readable. Another important factor influencing user readability is the map’s background — the choice of a basemap. The primary role of a background on a heat map is to allow the user spatial localization across the whole map for the heat map’s subject. An imagery map (orthophoto) and thematic layers are therefore not recommended as basemaps because it makes orientation in the map more difficult. More suitable variants are topographic layers such as borders, rivers, cities, etc. Map labels make orientation easier based on its quantity and map scale. Currently, GIS provide specific basemap layers — typically a light or dark gray layer, shaded relief, generalized topographic map, or layer with low color saturation. Since all the basemaps generated during the study combined a light gray layer with the borders of urban areas, we fully recommend a grey basemap or layer with low color saturation. This allows simple orientation across the whole map without disrupting heat map readability. Because it is a not simple question, it could be an interesting topic for further studies.
Radius values in heat maps change with scale level! Zoom-in or zoom-out movements cause the radius values to fluctuate. Therefore, radius setting preferences depend on individual skills and education due to the possible differences in user interpretation. While non-carto respondents prefer a smaller radius, a medium radius is more suitable for the carto group. A smaller radius provides more separated “islands” and is more similar to point method visualization. Finally, it is easier to interpret. A larger radius provides a more complex representation of a phenomenon, especially where the overlap occurs, but it requires conceptual skills for correct interpretation. Completely confluent and overlapping areas (one area over whole map) is not recommended at all.
The disadvantages of these aspects could be partly solved in a digital environment with interactive functionality. Web-based GIS such as ArcGIS Online or Google Fusion Tables allow interactive editing of cartographic aspects in real time via web browsers. Radius, transparency, or color could be edited on-screen, immediately and without needing to regenerate/resave map outputs. However, this facility could immediately lead to worse results if it is designed without knowledge of all the recommendations and rules. The main limitation of this paper is its focus on heat maps only. Since it is a subjective method, incorrect use can lead to incorrect perceptions of maps. The authors aim to help eliminate these mistakes via this paper. As previously mentioned, heat maps are a suitable method for visualizing point datasets, but not frequently used for line or polygon data, which could be limiting factors. The main aspect of similar studies which profoundly impacts the study results are participants. Participants should always represent the envisioned user, and it is crucial that the sampled participants represent the targeted levels of expertise and motivation for controlled experiments . That condition is followed by diving participants into two groups (novice and expert).
The choice of evaluation method is fundamental. According to the study of Rohrer  “all projects would benefit from multiple research methods and from combining insights”. Because of map designers are not common map (especially heat maps could be interpreted differently) users aware of cognitive processes increased. Van Elzakker et al.  brings a complex overview of user research. While repelling reactions to the quantitative revolution in geography were on the increase, “the think-aloud method has become an important tool” . As Hall  added “in attitudinal studies, you’re trying to find out what people say about a subject, while in behavioural studies, you’re analysing what people are doing. Qualitative methods tend to be stronger for answering ‘why’ types of questions, while quantitative methods do a better job of answering questions like ‘how many”’. Quantitative background of the survey “helps prioritise resources, for example, to focus on issues with the greatest impact” . That was the reason for choosing the combination of questionnaire and think-aloud methods for this study. The similar sequence has been used by Janicki et al. .
The limit of our study is the aim of the study - set recommendations based on quantitative research. There are some additional quantitative methods such as A/B testing, scenarios testing or quantitative eye-tracking testing which could be implemented into further research. Kubíček et al.  examined map-reading tasks on linear feature visualisation. Their study confirms that different forms of visualisation may have different impacts on performance in map-reading tasks: “colour hue and size proved more efficient in communicating information than shape and colour value.… Individual facets of cognitive style may affect task performance, depending on the type of visualisation employed” . Especially scenarios or A/B tests should lead to the more proper design and measuring the effect of these assignments on user behaviour . Roth et al.  present a complex agenda for user studies in the field of interactive maps. Their study highlights that digital and interactive maps “has fundamentally changed how maps are designed and used”. While heat map parameters could be easily customised via GIS platform we fully follow the statement “map reader is no longer passive in the creation of the representation … and the map user is empowered to create a representation that best supports his or her use context”. Roth et al.  recommend qualitative and mixed-method research as a staple for user studies, especially with a focus on human-computer interaction. Methods such as focus group, eye-tracking or another usability lab studies are suitable for qualitative testing of heat maps – detecting how to understand users´ perception on the heat map, answering why the heat map is (non)suitable method or why the heat map is (not) popular than another method. To provide a more comprehensive description of cartographic methods, three other exploratory studies of heat maps were performed under the project. Currently, we are developing a system of classification for evaluation and followed by the process of user testing with eye tracking evaluation. Some extension with practical tasks (e.g., find min/max hotspots or compare values at the same place on two heat maps) will be definitely included in further research.
Currently, the heat map is a highly used method for visualizing both graphical and tabular phenomena. For the interpretation of large-scale datasets, it is crucial for data to be properly classified and visualized. Heat maps are one effective and efficient option to process diverse spatial datasets in GIS via clustering analysis that looks for and unifies similar values. As a relevant cartographical standard is absent and many color ranges with an inappropriately chosen color transition were discovered, the authors decided to prepare a complex exploratory study focusing on evaluation and analysis of heat maps in geographical information systems and digital cartography. The aim of the thesis was to discover preferences for color ranges, transparency, and the radius setting of differently oriented data sets, emphasizing the preferences of both cartographic and non-cartographic respondents, not the individual needs a user may require. This paper describes an exploratory study examining traffic accident data. The experiment’s aim was to set basic recommendations and illustrate the typical strengths and weaknesses of heat map implementation.
This paper was supported by project “Cloud-based Platform for Integration and Visualization of Different Kind of Geodata” (IGA_PrF_2017_024) of the Palacký University.
Author Contributions: Rostislav Netek conceived the outline of the article and is the first author of the article. He is responsible for main concept, he conceived and designed the experiments. Tomas Pour is the co-author of Sections 2–4. He also helped with the data evaluation. Renata Slezakova performed the experiments and analyzed the data.
Yeap E.; Uy I. Marker Clustering and Heatmaps: New Features in the Google Maps Android API Utility Library. 2014. Google Geo Developers. Available online: http://googlegeodevelopers.blogspot.com/2014/02/marker-clustering-and-heatmaps-new.html. (accessed on 2018-03-11)
Sainio, J.; Westerholm, J.; Oksanen, J. Generating Heat Maps of Popular Routes Online from Massive Mobile Sports Tracking Application Data in Milliseconds While Respecting Privacy. ISPRS Int. J. Geo-Inf. 2015, 4, 1813-1826.
Ivan I., HorákJ. Metodika identifikace anomálních lokalit kriminality pomocíjádrových odhadů. 2016. Symposium GIS Ostrava 2016 – Geoinformatika pro společnost. Ostrava.
Anderson T.K. Kernel density estimation and K-means clustering to profile road accident hotspots. 2009. Accident Analysis & Prevention, Volume 41, Issue 3, May 2009, Pages 359-364
Kinney C. United States Patent and Trademark Office, registration #75263259". 1993.
Perez-Llamas, C. & Lopez-Bigas, N. Gitools: Analysis and Visualisation of Genomic Data Using Interactive Heat-Maps. PLoS ONE 6, e19541 (2011)
Moon, J. et al. Heat-map visualization of gas chromatography-mass spectrometry based quantitative signatures on steroid metabolism. Journal of the American Society for Mass Spectrometry, 2009, 20.9:1626-1637.
Hilton B., Horan N., Thomas A. and Schooley, Benjamin, 2009, Making Traffic Safety Personal: Visualization and Customization of National Traffic Fatalities. Visual Information Communication. 2009. P. 265-282. . Springer US
Li L., Zhu L., Sui D., 2007, A GIS-based Bayesian approach for analyzing spatial–temporal patterns of intra-city motor vehicle crashes. Journal of Transport Geography. 2007. Vol. 15, no. 4, p. 274-285. . Elsevier BV
Kušta T., Keken Z., Ježek M., Holá M., and Šmíd P., 2017, The effect of traffic intensity and animal activity on probability of ungulate-vehicle collisions in the Czech Republic. Safety Science. 2017. Vol. 91, p. 105-113. . Elsevier BV
Plug Ch., Xia J. C. and Caulfield C., 2011, Spatial and temporal visualisation techniques for crash analysis. Accident Analysis & Prevention. 2011. Vol. 43, no. 6, p. 1937-1946. . Elsevier BV
BENEDEK J., CIOBANU S. M., MAN T. C., 2016, Hotspots and social background of urban traffic crashes: A case study in Cluj-Napoca (Romania). Accident Analysis & Prevention. 2016. Vol. 87, p. 117-126. . Elsevier BV
Soltani, A. and Askari, S., 2014, Analysis of Intra-Urban Traffic Accidents Using Spatiotemporal Visualization Techniques. Transport and Telecommunication Journal. 2014. Vol. 15, no. 3. . Walter de Gruyter GmbH
DeBoer M. Understanding the Heat Map. 2015. . Available online: http://cartographicperspectives.org/index.php/journal/article/view/cp80-deboer/1420. (accessed on 2018-03-12)
Dempsey C. What is the Difference Between a Heat Map and a Hot Spot Map? 2014. Available online: https://www.gislounge.com/difference-heat-map-hot-spot-map/. (accessed on 201803-11)
Longley P.A., Goodchild M., Maguire J.M., Rhind D.W. Geoographic Information Systems and Science, 3 edition. 2010, Wiley, United Kingdom. 560 p. ISBN: 978-0470721445.
Panek J, Benediktsson K. Emotional mapping and its participatory potential: Opinions about cycling conditions in Reykjavik, Iceland. Cities. 2017, 61(1), 65–73.
Voženílek V, Kaňok J. Metody tematické kartografie: vizualizace prostorových jeů. Univerzita Palackého v Olomouci, 2011. ISBN 978-80-244-2790-4.
Kumar, C.; Heuten, W.; Boll, S. Visual Overlay on OpenStreetMap Data to Support Spatial Exploration of Urban Environments. IS-PRS Int. J. Geo-Inf. 2015, 4, 87-104.
Silverman B. W. Density estimation for statistics and data analysis. New York: Chapman and Hall, 1986.
Shi Z. et al.. A web-based interactive 3D visualization tool for building data. IBPSA Proceeding, 2016.
Galili T, O’Callaghan A, Sidi J, Sievert C. heatmaply: an R package for creating interactive cluster heatmaps for online publishing. 2017. Bionformatics, Oxford University Press.
Vondrakova A. Non-technological aspects of map production. 2013. Conference Proceedings SGEM 2013. STEF92 Technology Ltd.
Obadalkova, V. Hodnocení vlivu barev na čitelnost digitálních map. Olomouc, Czech Republic, 2012. Bachelor thesis. Palacký University Olomou
Dobesova Z., Vavra A., Netek R. Cartographic aspects of creation of plans for botanical garden and conservatories. 2013. Conference Proceedings SGEM 2013. STEF92 Technology Ltd., 653–660. ISSN 1314-2704.
Harrower, M; Brewer, CA.: ColorBrewer.org: An online tool for selecting colour schemes for maps. The Cartographic Journal. 2003, Vol. 40, Iss. 1.
Brychtova A., Vondrákova A.: Green versus Red: Eye-tracking evaluation of sequential colour schemes. 2014. SGEM 2014 Informatics, Geoinformatics and Remote Sensing Proceedings Volume III STEF92 Technology Ltd., 8s.
Živković D. Matrix pixel and Kernel density analysis from the topographic maps. 2016. Natural Sciences, Vol. 6, No 1, 2016, pp. 39-43.
Dreyfus, H L and Dreyfus, S E. Mind over Machine: the power of human intuition and expertise in the age of the computer. 1986, Oxford, Basil Blackwell.
Ullah, R., Mengistu, E.Z., van Elzakker, C.P.J.M., Kraak, M.J.: Usability evaluation of centered time cartograms. 2016, Open geosciences, Vol. 8, No. 1, p.337-359. ISSN 2391-5447.
Netek R., Balun M. WebGIS Solution for Crisis Management Support – Case Study of Olomouc Municipality. Computational Science and Its Applications – ICCSA 2014. ICCSA 2014. Lecture Notes in Computer Science, vol 8580
Esri. Heat map symbology. Available online: http://pro.arcgis.com/en/pro-app/help/mapping/symbols-and-styles/heat-map.htm (accessed on 2017-09-28)
Netek R., Panek J. Framework See-Think as a Tool for Crowdsourcing Support-Case Study on Crisis Management. 2016. ISPRS-International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, pp. 13-16.
Roth R.E., Çöltekin A., Delazari L., Filho H.F., Griffin A., Hall A., Korpi J., Lokka I., Mendonça A., Ooms K., van Elzakker C.P.J.M. User studies in cartography: opportunities for empirical research on interactive maps and visualizations. 2017. International Journal of Cartography, 3:sup1, 61-89,
- Export Citation
Roth R.E., Çöltekin A., Delazari L., Filho H.F., Griffin A., Hall A., Korpi J., Lokka I., Mendonça A., Ooms K., van Elzakker C.P.J.M. User studies in cartography: opportunities for empirical research on interactive maps and visualizations. 2017. International Journal of Cartography, 3:sup1, 61-89,)| false 10.1080/23729333.2017.1288534
Rohrer Ch. When to Use Which User-Experience Research Methods. 2014 Available online: https://www.nngroup.com/articles/which-ux-research-methods/. (accessed on 2018-0408)
van Elzakker C.P.J.M. The use of maps in the exploration of geographic data. 2004. Utrech, Enschede, Netherlands Geographical Studies 326. Utrech University, 208p. ISBN: 90-68 09-357-6
Hall T. How to choose a user research method. 2017 Available online: https://uxplanet.org/how-to-choose-a-user-research-method-985112051d84. (accessed on 2018-04-08)
Janicki J., Narula N., Ziegler M., Guenard B., Economo E.P. Visualizing and interacting with large-volume biodiversity data using client–server web-mapping applications: The design and implementation of antmaps.org. 2016. Ecological Informatics 32 (2016) 185–193. http://dx.doi.org/10.1016/j.ecoinf.2016.02.006
Kubíček P., Šašinka Č., Stachoň Z., Štěrba Z., Apeltauer J., Urbánek T. Cartographic Design and Usability of Visual Variables for Linear Features. 2017. The Cartographic Journal Vol. 54 No. 1 pp. 91–102, February 2017.