International Journal of Geo-Information

Article Geovisualization and Geographical Analysis for Fire Prevention

Nicklas Guldåker

Department of Human Geography, Faculty of Social Sciences, Lund University, Sölvegatan 12, S-223 62 Lund, Sweden; [email protected]



Received: 30 April 2020; Accepted: 25 May 2020; Published: 27 May 2020


Abstract: Swedish emergency services still have relatively limited resources and time for proactive fire prevention. As a result of this, there is an extensive need for strategic working methods and knowledge to take advantage of spatial analyses. In addition, decision-making based on visualizations and analyses of their own collected data has the potential to increase the validity of strategic decisions. The objective of this paper is to critically examine how some different geovisualization techniques—point data, kernel density and choropleth mapping—actively can complement each other and be applied in fire preventive work. The results show that each technique itself has limitations, but that, in combination, they increase the scope for interpretation and the possibilities of targeting different forms of preventive measures. The investigated geovisualization techniques facilitate various forms of fire prevention such as identifying which areas to prioritize for outreach, home visits, identification and targeting of different risk groups and customized information campaigns about certain types of fires in risk-prone areas. Furthermore, fairly simple mapping techniques can be utilized directly to evaluate incident reports and increase the quality of geocoded fire incidents. The study also shows how some of these techniques can be applied when analyzing residential fire incidents and their relation to underlying structural and socio-economic factors as well as spatio-temporal dimensions of fire incident data. The spatial analyses and supporting maps can help find and predict risk areas for residential fires or be used directly to formulate hypotheses on fire patterns. The generic functionality of the visualization methods makes them also useful for visual analysis of other types of incidents, such as reported crimes and accidents. Finally, the results are applicable to a work process adapted to the Swedish legislation on confidential data.

Keywords: residential fires; data visualization; geovisualization; emergency services; fire prevention; Sweden

1. Introduction The Swedish Civil Contingencies agency has a stated goal for fire safety in Sweden through Vision Zero aiming to eliminate citizen casualties and deaths [1]. In practice, the task for Swedish emergency services, is to reduce the number of fires and increase the ability to handle them when they arise. Beyond technical fire protection, the Swedish emergency services work actively with different fire safety outreach strategies. Such efforts cover everything from spreading information through campaigns, websites and social media to collaboration with other professionals such as the police or social services. The outreach methods also include educational fire safety demonstrations in schools, workplaces and retirement homes [2]. An effective way to improve fire safety work and the identification of fire-exposed areas and risk groups is to make better use of the fire statistics the emergency services have available in their databases. In this context, georeferenced fire data and various forms of visualizations are becoming increasingly important.

ISPRS Int. J. Geo-Inf. 2020, 9, 355; doi:10.3390/ijgi9060355 www.mdpi.com/journal/ijgi ISPRS Int. J. Geo-Inf. 2020, 9, 355 2 of 19

Geographical visualizations of urban fire incidents are relatively common in scientific contexts [3–5]. Spatial techniques are also widely used to visualize and analyze crimes, fires or other incidents over a specific urban area [6,7]. The aims with these spatial methods can vary, but the spatial dimension often gives the user an opportunity to reveal possible spatial patterns of incidents. Such information can be used for both theoretical and practical purposes, such as risk modeling, linking to socio-economic variables, theorizing causes of crimes and fires and developing hypotheses and approaches. These techniques can also be applied to analyze specific types of fire incidents on different geographical scales and in various areas. The magnitude of different types of fire incidents tends to vary in different areas. This is particularly significant for intentional fires, which often seem to be overrepresented in disadvantaged urban areas [8,9]. To highlight and understand the spatial dynamics and complexity of spatial phenomena, such as residential fires, often more than one spatial technology is needed [10]. An overall concern is thus how different spatial visualization techniques can be combined and develop the analysis of fire incidents on different geographic scales, and how these techniques can be applied by researchers and practitioners in a more structured way. Thus, there is a need for more in-depth knowledge and structured approaches of combined (geo) data-driven methods and visualization techniques in fire prevention processes. In the scientific literature, there is a general lack of spatial methods and strategies targeting the emergency services’ preventive fire work. Some of the existing and related GIS-based approaches include spatio-temporal analysis of structural fire incidents for fire prevention planning and fire response [6] as well as analysis of spatio-temporal distribution of intentional fires and living conditions as a basis for fire safety policies and fire prevention measures [9]. Other related approaches encompass explorative spatial analysis methods for rescue services [11], space-time patterns of residential structure fires and theory-based fire prevention [12] and spatial modeling of community fire risks for fire prevention activities [13]. Additional relevant studies cover risk mitigation strategies to support safety and emergency planning in case of urban fire [14,15], and the development of a methodology for risk assessment to handle territorial data [16]. Many of these spatial methods and visualization techniques are data-driven, and the reliability of the results is often influenced by the quality of the input data [7]. The reliability of the result is also affected by different settings such as scale, bandwidth, spatial analysis method and choice of visualization techniques. Geovisualization can thus be seen as a powerful strategy to present large spatial datasets, and the visual analysis offers a fast and interactive insight into the data without prior theoretical skills [7,10]. Swedish emergency services personnel have access to large amounts of data on fires and accidents collected from several national systems such as Daedalos or IDA [17,18]. Some systems, such as Daedalos, contain simple support for visualizations of fire data, e.g., point and heat maps. However, there is limited knowledge among emergency services personnel of the influence of basics map settings, and consequently, how the maps can be interpreted and used in their fire prevention work. Despite good access to large amounts of temporal and spatial data, the use is often far from adequate, and the map applications are limited in their functionality. Another important and relevant aspect related to the data management and spatial visualization process is how personal data should be processed and presented. The General Data Protection Regulation (GDPR) regulates the processing and flow of personal data in order to maintain privacy and anonymity rights for citizens within the European Union [19]. In a spatial data context, it is therefore essential to consider location-based information, geo-location methods and presentations forms that reduce the possibilities to identify and reveal private information about individuals. However, the scientific debate about individual exposure to location-enabled technologies is far from new, and has grown with social media and other ubiquitous positioning devices [20]. In addition to legal issues, geo-privacy also covers technological, ethical and educational aspects [21]. The coordination of privacy laws within EU to GDPR strengthens the restrictions and strongly influences the way organizations and authorities can process personal data. For emergency services, the use of private data is largely legal by ISPRS Int. J. Geo-Inf. 2020, 9, 355 3 of 19 exercise of authority according to the Civil Protection Act [22]. When it comes to data and presenting data on maps to be shared within and outside the emergency services organization, an appropriate method is to apply different spatial strategies depending on the type of fire preventive action. Consequently, there is a need for co-organized approaches using geographical visualizations and spatial analytical techniques in order to prevent fire incidents in different residential areas. The aim of this paper is within the scope of a more structured and user-friendly spatial visualization approach. The objective of this paper is to critically examine how different geovisualization techniques—in this case point data, kernel density and choropleth (map with colored and bounded areas) mapping—can complement each other and be applied in fire preventive work. The paper focuses on residential fires, which are defined broadly as fire incidents in residential buildings leading to emergency service interventions [23]. The examination includes a brief discussion on different subtypes of residential fires, such as intentional fires, accidental (unintentional) fires and fires related to technical failures or work processes. Furthermore, the paper also discusses the appropriateness (or not) of using different visualization techniques in relation to the General Data Protection Regulation (GDPR) and the protection of individuals [19]. A main conclusion from the paper are that the combination of several spatial techniques for visualizations of residential fires tend to increase the scope for interpretation and the possibilities of targeting various forms of preventive measures to specific neighborhoods, or even residential buildings with higher risk of residential fires.

2. Materials and Methods This section presents the study area, material, methodological aspects and the process of applying different geovisualization techniques to fire prevention.

2.1. Study Area In this paper, the geographical visualizations of residential fire data are demonstrated using the adjacent municipalities of Malmö and Burlöv as examples. Malmö is the fourth most densely populated city in Sweden, with more than 2159 inhabitants per km2. Burlöv is the twelfth densest area, with more than 968 inhabitants per km2. The municipalities have about 359,000 inhabitants (2018) and 160,000 households. About 70% of these are 1–2 people’s households. Over 80% of the residences are co-operative or rented apartments in apartment buildings [24,25]. Since the 1990s, the municipalities have gone through a substantial population increase owing to relatively high birth rates and a high influx of immigrants and young people. The city of Malmö is also characterized by an increasingly spatial, social and economic division. Statistics on residential fires from 2007–2017 shows a downward trend, especially for Malmö, which goes from a peak of 1.03 residential fires per 1000 inhabitants in 2008 to 0.58 in 2017. As shown in Figure1, Malmö also goes from having a significantly higher average than all of Sweden and the other metropolitan areas of Gothenburg and Stockholm in 2008, to falling under the national average in 2017. The main reason for the declining trend is a significant decrease of intentional residential fires from 2008–2015. Intentional fires, both outdoor and in residential buildings, were particularly frequent in socioeconomically disadvantaged neighborhoods in Malmö during the period 2007–2012 [2,9]. Malmö and Burlöv are two out of five municipalities within the Emergency Services South’s geographical area of responsibility (Figure2). ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 4 of 18 ISPRS Int. J. Geo-Inf. 2020, 9, 355 4 of 19

ISPRS Int.1.2 J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 4 of 17

1

0.8

0.6

0.4

0.2

0 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017

Malmö Gothenburg Stockholm Sweden 136 136

137 137 FigureFigureFigure 1. Residential 1. 1. ResidentialResidential fires fires per firesper 1000 1000 per inhabitants inhabitants 1000 inhabitants inin SwedenSweden in andand Sweden the the three three and largest largest the threemetropolitan metropolitan largest areas metropolitan areas 138 138 2007-2017.areas2007-2017. 2007–2017.

139 140 Figure 2. Emergency Services South consists of five municipalities in south Sweden, in addition to the 141 exemplified municipalities in this study Malmö and Burlöv (in red), also Eslöv, Lund and Kävlinge. 139 142 140 FigureFigure 2. Emergency 2. Emergency Services Services South South consists consists of five of municipalities five municipalities in south in south Sweden, Sweden, in addition in addition to the to the 141 exemplifiedexemplified municipalities municipalities in this in thisstudy study Malmö Malmö and Burlöv and Burlöv (in red), (in red),also Eslöv, also Eslöv, Lund Lund and Kävlinge. and Kävlinge.

ISPRS Int. J. Geo-Inf. 2020, 9, 355 5 of 19

2.2. Data and Process Applying different geovisualization techniques to fire prevention is divided in the following stages: In-data and Sources, Pre-processing, Processing and Visualization and Interpretation and Application (Figure3).

2.2.1. Stage 1 In-Data and Sources The fire records provided by the Emergency Services South over Malmö and Burlöv consist of 2469 residential fire incidents from 2007 to 2015 (Figure4). The fire dataset includes locations of residential fire incidents (X- and Y-coordinates) as well as a variety of information about assessed causes of fire, starting places, temporal information and addresses. Population data and other geodata are collected from Malmö City and Statistics Sweden through the map service Geographic Extraction Tool (GET) available to most Swedish Universities and other higher education institutions [26]. The data is temporary stored in a geodatabase managed by the Department of Human Geography at Lund University [27]. The ESRI basemaps World Light Gray Canvas Base and Imagery are used as backgrounds (Stage 1 in Figure3).

2.2.2. Stage 2 Pre-Processing An important part of pre-processing the residential fire data was the data review including cleaning, verification, sorting and selection of fire data categories (Stage 2 in Figure3). The selection criteria are based on previous research on residential fires [23]. The dataset over residential fires from the Emergency Services South was divided into one main and three different subcategories: 1) all residential fires; 2) intentional fires; 3) accidental fires; and 4) fires related to technical failures or work processes. Intentional fires are often associated with motives or expressions of norm breaking behaviors such as conflicts or different types of crime. Other motives can be thrill-seeking, boredom among young people or mental illness (pyromania) [9,28,29]. Intentional residential fires usually start in common areas that are difficult to monitor such as basements, storage rooms, stairwells in apartment buildings or just outside (outdoor fires) residential buildings [23]. Accidental fires can be related to people’s inability to prevent, detect and act when fires occur. Common causes of accidental fires include stress, forgetfulness, inattention or mishaps often associated with other conditions, such as disabilities, alcohol or drug influence and medication side-effects [30]. Common activities behind accidental residential fires include cooking, smoking and the use of candles or open flames for heating. Fires can also occur because of technical failures during work processes such as construction or heating buildings, e.g., chimney fires [31]. Chimney fires also tend to increase during cold winters in Sweden [32]. In order to select data from the residential fire data set for each category, different SQL functions were constructed in ArcGIS (Table1). The pre-processing stage also includes geo-referencing of fire data to X and Y, spatial joining of fire data to subareas, vector grids in Malmö and Burlöv (Figure5), as well as joining fire data to population data sets. In the subareas and grids, the numbers of residential fires are normalized per 1000 inhabitants. To facilitate the spatial joining of fire data to subareas and grids and normalization with population data, a model was created in ArcGIS (Figure6). The model also presents the pre-visualization process, from selecting residential subareas/grids to calculating the number of residential fires per 1000 inhabitants and per year for the period 2007–2015. ISPRS Int. J. Geo-Inf. 2020, 9, 355 6 of 19

Table 1. Categorization of residential fires and their related causes (Data source: Emergency Services South).

Category of Residential Fire Selected (SQL) Assessed Fire Causes of Fire from the Fire Data Set All residential fires Fire causes = all Intentional fires Fire cause = Intentional fire setting Fire causes = cooking; forgotten stove left on; combustion of residual oils on the stove; stove started by mistake; cooking with oil; child, adult, or animal turned the knobs on the stove; wrong burner on the stove turned on; boiling dry pot; unattended cooking; fire in oil in a saucepan on the stove; bedside lamp; water boiler on the stove; forgotten fryer; plastic bowl on a stove burner; towel in the oven; things lying on the stove; microwave oven; toaster on burner; forgotten microwave; overheated food; leftovers in the oven; cardboard on the stove burner; cake in the oven; basket with cakes on the stove; forgotten pressure cooker; confused lady who started the stove; paper in the cook box; objects in the oven; plastic bag on burner; forgotten toaster; Accidental fires forgotten or knocked-over candles; smoking; something thrown into the trash that ignited; burning matchstick; unextinguished cigarettes or cigarette butts; children start the iron; floor lamp fallen on the couch; warm lamp; forgotten smoke stick in toilet bowl; bed smoking; barbecue; threw burning garbage without intent; vacuuming ash; overfilled/leakage of lamp oil; plastic flower on the sauna unit; incense; insect extermination with gasoline; burning of weeds; overloaded dryer; carelessness; confused lady who burns paper; sparklers; outdoor torch; textile/clothing on heat source; forgotten disposable grill; towel on top of spotlights; drunk person; ashtray; mistake; carelessness with fire; fire in fireplace. Fire causes = technical error, explosion, leaking chimney, self-ignition, chimney fire, hot work processes, heat transfer, lightning strikes, re-ignition, electrical box, steam, Fires related to technical failures or work processes sparks from electricity, electric fault, short circuit, heating system in the basement, heat generation in power cord, electrical coil. ISPRS Int. J. Geo-Inf. 2020, 9, 355 7 of 19

176

177 FigureFigure 3. The3. The process process for for developing developing diff differenterent GIS-based GIS-based visualization visualizatio techniquesn techniques and and their their applications applications in firein fire prevention. prevention.

ISPRS Int. J. Geo-Inf. 2020, 9, x; doi: FOR PEER REVIEW www.mdpi.com/journal/ijgi ISPRS Int. J. Geo-Inf. 2020, 9, 355 8 of 19

178 178 179 FigureFigure 4. Point 4. Point map map (left) ( leftand) heat and map heat (right) map (right of th)e ofmunicipalities the municipalities of Malmö of Malmöand Burlöv, and Burlöv,Sweden. Sweden. 180179 TheFigureThe residential 4. residential Point firemap firedata (left) data period and period heat is 2007–2015. map is 2007–2015. (right) of the municipalities of Malmö and Burlöv, Sweden. 180 The residential fire data period is 2007–2015.

181 181 182 Figure 5. Choropleth maps as subareas (left) and grids (right) of the municipalities of Malmö and 183182 Burlöv,FigureFigure 5.Sweden. Choropleth 5. Choropleth The residential maps maps as subareas fire as subareasdata (left)period (andleft was )gr and 2007–2015.ids grids(right) ( rightof the) ofmunicipalities the municipalities of Malmö of Malmöand and 183 Burlöv,Burlöv, Sweden. Sweden. The Theresidential residential fire data fire dataperiod period was 2007–2015. was 2007–2015. ISPRS Int. J. Geo-Inf. 2020, 9, x; doi: FOR PEER REVIEW www.mdpi.com/journal/ijgi ISPRS Int. J. Geo-Inf. 2020, 9, x; doi: FOR PEER REVIEW www.mdpi.com/journal/ijgi ISPRS Int. J. Geo-Inf. 2020, 9, 355 9 of 19 ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 3 of 18

195 Figure 6. Modeling of the spatial join of fire data to subareas and grids and normalization with 196 Figure 6. Modeling of the spatial join of fire data to subareas and grids and normalization with population data. The model is developed with the model builder tool in ArcGIS desktop. 197 population data. The model is developed with the model builder tool in ArcGIS desktop. 2.2.3. Stages 3 and 4 Processing and Visualization 198 2.2.3. Stages 3 and 4 Processing and Visualization The processing of residential fire data for visualization and analysis of spatial fire data encompasses 199 three diThefferent processing mapping of methodsresidential (Stages fire 3data and for 4 in visualization Figure3). and analysis of spatial fire data 200 encompasses three different mapping methods (Stages 3 and 4 in Figure 3). 201 1. The point1. The data point mapping data mapping of fire data of fire in Malmö data in and Malm Burlövö and is Burlöv preceded is preceded by a cluster by a test. cluster A Nearest test. 202 NeighborA test Nearest was conductedNeighbor intest order was toconducted determine in whether order to residential determine fire whether clusters residential are statistically fire 203 significantclusters or whether are statistically they are significant likely explained or whet byher other they reasonsare likely than explained random by chance.other reasons For the 204 purpose,than the Averagerandom chance. Nearest For Neighbor the purpose, tool in the ArcGIS Average is utilized Nearest (see Neighbor Section tool 3.1 below).in ArcGIS is 205 2. The kernelutilized density (see mapping Section 3.1 is below). based on the kernel density estimation (KDE), which is one of 206 2. various methodsThe kernel for density point mapping pattern analysis is based (PPT) on the [33 kernel]. KDE density is widely estimation used for(KDE), the estimationwhich is 207 one of various methods for point pattern analysis (PPT) [33]. KDE is widely used for the of an increased probability of an event to occur in a pre-defined cluster [9,34,35]. In this paper, 208 estimation of an increased probability of an event to occur in a pre-defined cluster [9,34– the density of different types of residential fires is calculated and visualized in raster layers with 209 35]. In this paper, the density of different types of residential fires is calculated and low to high values per square kilometer (km2). In other words, many adjacent raster cells with 210 visualized in raster layers with low to high values per square kilometer (km2). In other high values in one or many places in the map may indicate geographical concentrations of fires. 211 words, many adjacent raster cells with high values in one or many places in the map The KDE method measures the distance between coordinates or centroids of the raster cells. An 212 may indicate geographical concentrations of fires. The KDE method measures the average of the distance from a coordinate to all nearest coordinates is then calculated. If the 213 distance between coordinates or centroids of the raster cells. An average of the distance average appears to be less than the hypothetical value of random distribution of residential fires, 214 from a coordinate to all nearest coordinates is then calculated. If the average appears to the method indicates—as the Nearest Neighbor test proves—that there are clusters. In order to 215 be less than the hypothetical value of random distribution of residential fires, the 216 find thesemethod local patterns indicates—as and to the reduce Nearest the Neighbor risk of spatial test proves—that variation, the there search are radius clusters. (bandwidth) In order 217 was set asto default,find these the local output patterns cell sizeand to set reduce consistently the risk at of 50 spatial m, and variation, the area the unit search set in radius square 2 218 kilometers(bandwidth) (km ). For was the classificationset as default, to the be output valid incell a Swedishsize set consistently urban context, at 50 the meters, ordering and ofthe the 219 values intoarea classes unit set presented in square in kilometers the HEAT-maps (km2). For is basedthe classification on KDE-modeling to be valid of residentialin a Swedish fire 220 values fromurban a larger context, data the set ordering covering of 16 the municipalities values into inclasses Stockholm, presented Gothenburg in the HEAT-maps and Malmö is [2 ]. 221 3. The choropleth-mappingbased on KDE-modeling was performed of residential in two fire ways: values a) onfrom subarea-level; a larger data and set b) covering in statistical 16 222 250 250municipalities m grids in morein Stockholm, densely populatedGothenburg areas and Malmö and 1000 [2]. 1000 m grids in less densely × × 223 populated3. The areas. choropleth-mapping However, 1000 was1000 performed m grids were in tw noto ways: collected a) on in Malmösubarea-level; and Burlöv and b) owing in × 224 to the highstatistical population 250 × density 250 meter of thesegrids municipalities.in more densely In populated order to avoidareas distortedand 1000 × ratio 1000 values meter of 225 residentialgrids fires in per less 1000 densely inhabitants, populated only subareasareas. However, with more 1000 than × 5001000 inhabitants meter grids and were 250-meter not 226 grids withcollected more thanin Malmö 100 residents and Burlöv were selected.owing to Thethe limithigh ofpopulation 500 inhabitants density for of subareas these 227 is also amunicipalities. proven and appropriate In order to level avoid for distorted statistical ratio analyses values conducted of residential in a parallelfires per study 1000 of 228 Stockholm,inhabitants, Gothenburg only andsubareas Malmö with metropolitan more than 500 areas inhabitants [36]. The and classification 250-meter of grids residential with 229 fire valuesmore per than 1000 100 inhabitants residents intowere classes selected. used The in limit the choroplethsof 500 inhabitants is based for onsubareas values is from also 853a 230 subareasproven and 2986 and residential appropriate grids level in Stockholm,for statistical Gothenburg analyses conducted and Malmö in metropolitan a parallel study areas of [2 ].

ISPRS Int. J. Geo-Inf. 2020, 9, 355 10 of 19

The classification technique is standard deviations rounded to the nearest half-integer, with the exception of the lowest and highest values. All layouts are carried out in ArcMap version 10.5.1.

2.2.4. Stage 5 Interpretation and Application The interpretations and proposals for using these different geovisualization techniques in fire prevention are partly based on results and feedback from workshops with staff from emergency services in the three largest metropolitan areas in Sweden, and partly on results from the use of techniques and methods in research [32,36]. In total, four workshops have been held with 32 employees from Emergency Services South (Malmö), Emergency Services Stockholm, Emergency Services Södertörn (Stockholm) and Emergency Services Gothenburg. Participants in the workshops were analysts, statisticians, fire engineers, fire inspectors, chief fire officers, commanders, fire safety officers and area managers. The overall purpose of the workshops was to present, discuss and get feedback on different GIS-based and geo-statistical analyzes of residential fires within each emergency services area. The process during the workshops was that the various map visualizations were presented to the personnel from the emergency services. All participants were then given the opportunity to comment on and problematize both maps and the spatial distribution of fires within their areas of operation (e.g., districts within Malmö, Stockholm and Gothenburg). All comments were compiled into a document that was further used as input to the evaluation of strengths and weaknesses with the various geovisualization techniques and their specific applicability in the emergency service’s fire prevention work. The results of the presented geovisualization methods in this paper are based on all workshops and largely applicable to all three metropolitan areas. In order to not repeat presentations from all workshops, this paper only presents data and visualizations of the municipalities of Malmö and Burlöv.

3. Results and Discussion In this section, different spatial visualization techniques and their applicability in fire prevention work are compared and critically examined (see also Stage 5 in Figure3).

3.1. Point Maps Point map visualizations hold exact cartesian coordinates of residential fires in the Euclidean space. These visualizations are advantageous for fire prevention, but can be difficult to validate. Dynamic Point map visualizations are more useful than static ones in the sense that they permit emergency services to navigate to exact x- and y-positions. The point map can be used to locate residential fires and discuss anticipated spatial clusters in different parts of the mapped area. Coordinates linked to specific addresses facilitate local and regional analyses. However, a possible spatial cluster must be proved, such as in Figures4 and7 over Malmö and Burlöv. A Nearest Neighbor (NN) test reveals that the observed mean distance is much lower than the estimated mean for all residential fires in Malmö and Burlöv. The probability (p) value is less than 0.01 (confidence level 99%), which means that clustered patterns for residential fires with a high probability can be explained by reasons other than random causes. Table2 shows that all four subcategories of residential fires have spatial clustering. This is also confirmed by index values (average nearest neighbor ratios) significantly below 1. The lowest index value applies for intentional fires, which indicates the strongest clustering among all categories. Other important uses of point map visualizations might be identifying mistakes in the data collection through using event reports, which can easily be corrected by double checking the positions just after a fire incident. This is particularly important because some of the reported fires have misplaced or no coordinates at all [37]. According to the emergency services, there are gaps in the knowledge and routines for working with dynamic maps and evaluating their spatial fire statistics. Consequently, it can be difficult for emergency service personnel to interpret the large amounts of points over large geographical areas. There are different ways to display and reveal spatio-temporal patterns of fires. Point maps, especially with individual colors for different categories, may work well ISPRS Int. J. Geo-Inf. 2020, 9, 355 11 of 19

for overall spatial analysis and hypothesis formulation about underlying causes and preparation for point pattern analysis (PPT). Some caution should be taken with point maps when it comes to protecting personal data. In ISPRS Int. J. Geo-Inf. 2020, 9, xdynamic FOR PEER forms, e.g., REVIEW on web GIS platforms with open access that can identify specific addresses, point 5 of 18 maps are not compatible with the General Data Protection Regulation’s (GDPR’s) goal to maintain privacy and anonymity rights for citizens, e.g., those whom have had fire incidents at home.

274 Figure 7. Point map.

Table 2. Given the z-score, there is a less than 1 % probability that clustered patterns of different types 275 residential fires could be the effFigureect of random 7. chance. Point The area map. was fixed at 159 km2 and the distance method is Euclidean.

Observed Estimated Mean Index—Average Residential Fires Z-Score P-Value 276 The lowest index value appliesMean (Meters)for intentional(Meters) fires, Nearestwhich Neighbor indicates Ratio the strongest clustering All 38.1 134.6 68.2 0.28 0.00 − Intentional 48.7 229 41.5 0.21 0.00 277 among all categories. Other important uses of point map− visualizations might be identifying mistakes Accidental 87.6 228.2 32.6 0.38 0.00 − 278 in the data collection throughTechnical failures using event166.8 reports, 378.3 which17.8 can easily 0.44 be corrected 0.00 by double checking − 279 the positions just after a fire incident. This is particularly important because some of the reported 3.2. Heat Maps 280 fires have misplaced or no KDEcoordinates modeling visualizes at clusters all [37]. of residential According fires per area unit, to in the this case, emergency square kilometer services, there are gaps (km2). These heat-maps provide clear indications of where to find statistically significant clusters of 281 in the knowledge and routinesfires in Malmö andfor Burlöv. working In all categories with of residential dynamic fires in Malmö maps and Burlöv, and the observedevaluating their spatial fire average distance is below the estimated hypothetical mean distance (Table2). Bright red colors and 282 statistics. Consequently,high it values can in thebe heat difficult map indicate high for spatial emergency intensity of residential service fires. Yellow colorspersonnel correspond to interpret the large with lower concentrations. Concentrations below 20 residential fires per km2 have not been visualized 283 amounts of points overand large lacks color geographical (Figures4 and8). Statistically areas. tested heat There maps assist are emergency different services in selectingways to display and reveal 284 neighborhoods for further analysis and fire prevention. They may also indicate underlying spatial spatio-temporal patternsprocesses of and fires. phenomena Point that need tomaps, be examined moreespecially thoroughly, such with as spatio-temporal individual analysis colors for different 285 categories, may work wellof residential for overall fires or different spatial living conditions analys in certainis neighborhoodsand hypothesis [7,9]. formulation about underlying However, one weakness with heat maps is that they are not normalized. High concentrations of 286 causes and preparation firesfor per point km2 may pattern possibly be an analysis effect of many people (PPT). living in dense areas, e.g., several block tower apartments within a smaller area, which is partly the case for Malmö and other metropolitan areas in 287 Some caution should be taken with point maps when it comes to protecting personal data. In 288 dynamic forms, e.g., on web GIS platforms with open access that can identify specific addresses, point 289 maps are not compatible with the General Data Protection Regulation’s (GDPR’s) goal to maintain 290 privacy and anonymity rights for citizens, e.g., those whom have had fire incidents at home.

291 Table 2. Given the z-score, there is a less than 1 % probability that clustered patterns of different types 292 residential fires could be the effect of random chance. The area was fixed at 159 km2 and the distance 293 method is Euclidean.

Estimated Residential Observed Mean Z- Index - Average Nearest P- Mean Fires (Meters) score Neighbor Ratio Value (Meters) All 38,1 134,6 -68.2 0,28 0,00 Intentional 48,7 229 -41,5 0,21 0,00 Accidental 87,6 228,2 -32.6 0,38 0,00 Technical 166,8 378,3 -17,8 0,44 0,00 failures

294 3.2. Heat Maps 295 KDE modeling visualizes clusters of residential fires per area unit, in this case, square kilometer 296 (km2). These heat-maps provide clear indications of where to find statistically significant clusters of 297 fires in Malmö and Burlöv. In all categories of residential fires in Malmö and Burlöv, the observed 298 average distance is below the estimated hypothetical mean distance (Table 2). Bright red colors and 299 high values in the heat map indicate high spatial intensity of residential fires. Yellow colors 300 correspond with lower concentrations. Concentrations below 20 residential fires per km2 have not 301 been visualized and lacks color (Figures 4 and 8). Statistically tested heat maps assist emergency 302 services in selecting neighborhoods for further analysis and fire prevention. They may also indicate 303 underlying spatial processes and phenomena that need to be examined more thoroughly, such as 304 spatio-temporal analysis of residential fires or different living conditions in certain neighborhoods 305 [7,9].

ISPRS Int. J. Geo-Inf. 2020, 9, 355 12 of 19

Sweden. In order to reveal other underlying factors than population density, the heat-maps can be ISPRS Int. J. Geo-Inf. 2020, 9, xcomplemented FOR PEER with REVIEW choropleth maps normalized with the population. Heat maps do not reveal where 6 of 18 people live and can be used in communication about fire risks and fire prevention with other private or public actors.

306 Figure 8. Heat map.

3.3. Choropleth Maps 307 The choropleth maps displayFigure the residential 8. fireHeat and population map. ratio or the number of residential fires per 1000 inhabitants per year for the period 2007–2015 in Malmö and Burlöv. Choropleth visualizations are here represented using subarea and grid maps. 308 However, one weakness with heat maps is that they are not normalized. High concentrations of 3.3.1. Subarea Maps 2 309 fires per km may possiblyHigh be oran low effect values of residentialof many fires inpeople subareas of living Malmö and in Burlöv dense entail moreareas, or less e.g., several block tower fires relative to the population (Figures5 and9). Notably, the average for Malmö and Burlöv is 310 apartments within a smaller0.76 residential area, fires which per 1000 inhabitants, is partly which the is significantly case higherfor Malmö than the average and 0.49 forother the metropolitan areas in metropolitan areas of Stockholm, Gothenburg and Malmö together. The highest value in Sweden, 6.1 311 Sweden. In order to revealresidential other fires per underlying 1000 inhabitants during factor 2007–2015,s than is found population in a subarea within Malmö density, called the heat-maps can be Rosengård, a district badly affected by intentional fire setting both outside and inside buildings [9]. As 312 complemented with choroplethmentioned in Section maps3, it should normalized also be emphasized thatwith the number the ofpopulation. intentional fires in Malmö Heat maps do not reveal 313 where people live and canhas decreased be used considerably in communication during the period 2008–2015 [about32]. The normalization fire risks of residential and firesfire prevention with other per 1000 inhabitants is an essential first stage in the analysis because determining factors other than 314 private or public actors.population density may correlate to the spatial distribution of residential fires. The subarea level is advantageous for further geo-statistical analysis owing to a rich and partly open access to other data, such as socio-economic variables at the same geographic level. The subarea level has been used in several geo-statistical analyses of intentional and residential fires and underlying factors in Malmö, 315 3.3. Choropleth Maps Stockholm and Gothenburg [9,36]. Another advantage of choropleth maps is that normalized subarea values can easily be compared with other normalized values in the same city or with other cities. In a comparison of residential fires 316 The choropleth mapsbetween thedisplay three largest metropolitanthe residential areas in Sweden, eightfire out and of eleven population of the most affected areas ratio or the number of are located in Malmö, while only two can be found in Stockholm, and one in Gothenburg [32]. One 317 residential fires per 1000disadvantage inhabitants is that the size ofper a subarea year can be consideredfor the too bigperiod for more e ff2007–2015ective area-based fire in Malmö and Burlöv. 318 Choropleth visualizationsprevention. are Preventivehere re residentialpresented fire interventions using are moresubarea often targeted and at smaller grid geographical maps.

319 3.3.1. Subarea Maps 320 High or low values of residential fires in subareas of Malmö and Burlöv entail more or less fires 321 relative to the population (Figures 5 and 9). Notably, the average for Malmö and Burlöv is 0.76 322 residential fires per 1000 inhabitants, which is significantly higher than the average 0.49 for the 323 metropolitan areas of Stockholm, Gothenburg and Malmö together. The highest value in Sweden, 6.1 324 residential fires per 1000 inhabitants during 2007–2015, is found in a subarea within Malmö called 325 Rosengård, a district badly affected by intentional fire setting both outside and inside buildings [9]. 326 As mentioned in Section 3, it should also be emphasized that the number of intentional fires in Malmö 327 has decreased considerably during the period 2008–2015 [32]. The normalization of residential fires 328 per 1000 inhabitants is an essential first stage in the analysis because determining factors other than 329 population density may correlate to the spatial distribution of residential fires. The subarea level is 330 advantageous for further geo-statistical analysis owing to a rich and partly open access to other data, 331 such as socio-economic variables at the same geographic level. The subarea level has been used in 332 several geo-statistical analyses of intentional and residential fires and underlying factors in Malmö, 333 Stockholm and Gothenburg [9,36].

334

335 Figure 9. Choropleth map - Subarea.

ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 7 of 18

336 Another advantage of choropleth maps is that normalized subarea values can easily be 337 compared with other normalized values in the same city or with other cities. In a comparison of 338 residential fires between the three largest metropolitan areas in Sweden, eight out of eleven of the 339 most affected areas are located in Malmö, while only two can be found in Stockholm, and one in 340 Gothenburg [32]. One disadvantage is that the size of a subarea can be considered too big for more 341 effective area-based fire prevention. Preventive residential fire interventions are more often targeted 342 at smaller geographical areas such as neighborhoods, blocks or building with higher fire risks. 343 Accordingly, emergency services need to identify work with large-scale fire risk areas, or in other 344 words, smaller geographic units.

345 3.3.2. Grid Maps 346 Proposed grid maps show the number of residential fires per 1000 inhabitants in densely 347 populated residential areas of Malmö and Burlöv (Figure 5 and 10). In comparison to subarea maps, 348 grid maps render an increased geographical accuracy. Many more grids hold values above the 349 average residential fire to population ratio. This facilitates the emergency service's preventive work, 350 e.g., by prioritizing smaller and more homogeneous areas for home visits or other geographically 351 targetedISPRS Int. fire J. Geo-Inf. prevention 2020, 9, xmeasures. FOR PEER REVIEW In the grid map, several high priority grids appear over 7 Malmö of 18 352 and Burlöv (Figure 5). This diffused pattern does not appear in the choropleth map of subareas, which 336 Another advantage of choropleth maps is that normalized subarea values can easily be 353337 insteadcompared highlight with expectedother normalized fire exposed values areas in thesuch same as those city orin Rosengårdwith other cities.(three Indarkest a comparison red subareas of 354 in Figure 5). However, caution should be taken when interpreting grid values. For example, 338 residential fires betweenISPRS Int.the J. Geo-Inf. three2020, 9, 355largest metropolitan areas in Sweden, eight13 of 19 out of eleven of the 355339 populationmost affected values areas just are over located 100 ininhabitants, Malmö, while in combinationonly two can withbe found multiple in Stockholm, fires, can and give one rise in to areas such as neighborhoods, blocks or building with higher fire risks. Accordingly, emergency services 356340 disproportionatelyGothenburg [32]. highOne need disadvantagefire to identifyrisk workvalues with is large-scale thatin certain firethe risk size areas, grids. orof in othera subarea words, smaller can geographic be considered units. too big for more 341 effective area-based fire prevention. Preventive residential fire interventions are more often targeted 342 at smaller geographical areas such as neighborhoods, blocks or building with higher fire risks. 343 Accordingly, emergency services need to identify work with large-scale fire risk areas, or in other 344 words, smaller geographic units.

345 3.3.2. Grid Maps 346 Proposed grid maps show the number of residential fires per 1000 inhabitants in densely 347 populated residential areas of Malmö and Burlöv (Figure 5 and 10). In comparison to subarea maps, 357348 grid maps render an increased geographical accuracy. Many more grids hold values above the 349 average residential fire to population ratio.Figure This 9. Choropleth facilitates map—Subarea. the emergency service's preventive work, 3.3.2. Grid Maps 358350 e.g., by prioritizing smaller and moreFigure homogeneou 10. Choropleths areas map -for grids. home visits or other geographically Proposed grid maps show the number of residential fires per 1000 inhabitants in densely populated 351 targeted fire preventionresidential measures. areas of Malmö In and the Burlöv grid (Figures map,5 and 10 several). In comparison high to subarea priority maps, grid grids maps appear over Malmö render an increased geographical accuracy. Many more grids hold values above the average residential 359352 andVarious Burlöv principles(Figure 5). Thisfirefor to population thediffused interpretation ratio. Thispattern facilitates does the emergencyof no thte appear service’s data preventive canin the work,therefore choropleth e.g., by prioritizing be mapapplied of subareas, and tested. which For 353 instead highlight expectedsmaller andfire more exposed homogeneous areas areas for homesuch visits as or otherthose geographically in Rosengård targeted fire prevention (three darkest red subareas 360 example, isolated grids measures.with higher In the grid map, values several high often priority gridslie appearin more over Malmö sparsely and Burlöv (Figurepopulated5). This areas with detached 354 in Figure 5). However,diffused caution pattern does notshould appear in thebe choropleth take mapn when of subareas, interpreting which instead highlight grid expected values. For example, 361 houses (e.g., villas). Gridsfire exposedadjacent areas such to aseach those in other Rosengård often (three darkest cover red subareas multi-family in Figure5). However, houses in apartment blocks 362355 withpopulation higher population values justcaution density.over should 100 be Causes taken inhabitants, when interpreting of residential gridin values.combination For example,fires populationmay with al values somultiple justdiffer over 100 fires,between can moregive riseand to less 356 disproportionately highinhabitants, fire ri insk combination values with in multiple certain fires, can grids. give rise to disproportionately high fire risk values 363 populated areas. In villain certain areas grids. in Stockholm, Gothenburg and Malmö, technical, electrical, or 364 constructional failures such as chimney fires are overrepresented. In multi-family houses in more 365 densely populated areas, accidental residential fires, such as cooking fires, appear to be more 366 prevalent [32]. As stated above, intentional fires have been the dominant cause in some poorer areas 367 in Malmö [9].

368 3.4. Discussion 369 Swedish emergency services still have relatively limited resources and time for proactive fire 370357 prevention. As a result of this, there is an extensive need for strategic working methods and 371 knowledge to take advantage of spatial analyses.Figure 10. Choropleth In addition, map—grids. decision-making based on analyses of 372358 their own collected data has the Figurepotential 10. Choropleth to increase map -the grids. validity of strategic decisions. The 373 application fields are multidimensional. To broaden the scope, residential fire data can be analyzed 359 Various principles for the interpretation of the data can therefore be applied and tested. For 374 in relation to contemporary socio-economic data provided from other parts of the society, e.g., the 360 example, isolated grids with higher values often lie in more sparsely populated areas with detached 375 municipalities. 361 houses (e.g., villas). Grids adjacent to each other often cover multi-family houses in apartment blocks 362 with higher population density. Causes of residential fires may also differ between more and less 363 populated areas. In villa areas in Stockholm, Gothenburg and Malmö, technical, electrical, or 364 constructional failures such as chimney fires are overrepresented. In multi-family houses in more 365 densely populated areas, accidental residential fires, such as cooking fires, appear to be more 366 prevalent [32]. As stated above, intentional fires have been the dominant cause in some poorer areas 367 in Malmö [9].

368 3.4. Discussion 369 Swedish emergency services still have relatively limited resources and time for proactive fire 370 prevention. As a result of this, there is an extensive need for strategic working methods and 371 knowledge to take advantage of spatial analyses. In addition, decision-making based on analyses of 372 their own collected data has the potential to increase the validity of strategic decisions. The 373 application fields are multidimensional. To broaden the scope, residential fire data can be analyzed 374 in relation to contemporary socio-economic data provided from other parts of the society, e.g., the 375 municipalities.

ISPRS Int. J. Geo-Inf. 2020, 9, 355 14 of 19

Various principles for the interpretation of the data can therefore be applied and tested. For example, isolated grids with higher values often lie in more sparsely populated areas with detached houses (e.g., villas). Grids adjacent to each other often cover multi-family houses in apartment blocks with higher population density. Causes of residential fires may also differ between more and less populated areas. In villa areas in Stockholm, Gothenburg and Malmö, technical, electrical, or constructional failures such as chimney fires are overrepresented. In multi-family houses in more densely populated areas, accidental residential fires, such as cooking fires, appear to be more prevalent [32]. As stated above, intentional fires have been the dominant cause in some poorer areas in Malmö [9].

3.4. Discussion Swedish emergency services still have relatively limited resources and time for proactive fire prevention. As a result of this, there is an extensive need for strategic working methods and knowledge to take advantage of spatial analyses. In addition, decision-making based on analyses of their own collected data has the potential to increase the validity of strategic decisions. The application fields are multidimensional. To broaden the scope, residential fire data can be analyzed in relation to contemporary socio-economic data provided from other parts of the society, e.g., the municipalities. The spatial analyses and supporting maps can help find and predict risk areas for residential fires or be used directly to formulate hypotheses on fire patterns. Relatively simple mapping techniques, such as point mapping, can be utilized directly to evaluate incident reports and increase the quality of geocoded fire incidents. They can also serve as an input to risk analysis procedures that enhance the performance of the emergency services [7]. Different templates and examples of techniques need to be implemented in existing incident reporting systems in the emergency service organization. The discussions with the emergency services also indicate that there is a general need for a basic understanding of how spatial information in maps and charts can be perceived in different ways. The knowledge of interpreting map visualizations varies within and between each emergency service. If a visualization becomes too rich in details and complex, it can be too difficult for people to transform what they see into meaningful interpretations and hypotheses [38]. This was evident from the workshops with the different emergency services associations. This can also be linked to uncertainty when visualizing different map types [39,40]. Different methods for incorporating and mapping uncertainty may increase the possibility of more accurate interpretations and analyzes of the information in different map visualizations [41]. Some research also indicates the need to include methods for uncertainty assessments in visualizations of spatiotemporal datasets [42]. Although uncertainty is an extensive field in risk research, there are still few applications related to map visualizations of fires, e.g., visualization of wildfire hazard [43,44]. The presented process highlights how different map visualizations can support emergency services fire prevention work in various and complementary ways. A summary of the strengths and weaknesses of the visualization methods is presented in Table3. The table also describes possible application fields for the spatial application of residential fires. ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 8 of 18

376 The spatial analyses and supporting maps can help find and predict risk areas for residential 377 fires or be used directly to formulate hypotheses on fire patterns. Relatively simple mapping 378 techniques, such as point mapping, can be utilized directly to evaluate incident reports and increase 379 the quality of geocoded fire incidents. They can also serve as an input to risk analysis procedures that 380 enhance the performance of the emergency services [7]. 381 Different templates and examples of techniques need to be implemented in existing incident 382 reporting systems in the emergency service organization. The discussions with the emergency 383 services also indicate that there is a general need for a basic understanding of how spatial information 384 in maps and charts can be perceived in different ways. The knowledge of interpreting map 385 visualizations varies within and between each emergency service. If a visualization becomes too rich 386 in details and complex, it can be too difficult for people to transform what they see into meaningful 387 interpretations and hypotheses [38]. This was evident from the workshops with the different 388 emergency services associations. This can also be linked to uncertainty when visualizing different 389 map types [39,40]. Different methods for incorporating and mapping uncertainty may increase the 390 possibility of more accurate interpretations and analyzes of the information in different map 391 visualizations [41]. Some research also indicates the need to include methods for uncertainty 392 assessments in visualizations of spatiotemporal datasets [42]. Although uncertainty is an extensive 393 field in risk research, there are still few applications related to map visualizations of fires, e.g., 394 visualization of wildfire hazard [43,44]. The presented process highlights how different map 395 visualizations can support emergency services fire prevention work in various and complementary 396 ways. A summary of the strengths and weaknesses of the visualization methods is presented in Table 397 3. The table also describes possible application fields for the spatial application of residential fires. ISPRS Int. J. Geo-Inf. 2020, 9, 355 15 of 19 398 Table 3. Strengths and weaknesses of the visualization methods, and possible fields for spatial 399 application of residential fires. Table 3. Strengths and weaknesses of the visualization methods, and possible fields for spatial application of residential fires.

VisualizationVisualization StrenghtsStrenghts Weaknesses WeaknessesApplication fields - Application Fields—Example of techniques/methods example of residential fires Techniques/Methods Residential Fires Point map • Easy to map • Perception of the number of • Plotting of residential fires Easy to map Perception of the number of residential Plotting of residential fires • • • • Exact positionsExact positions residential fires difficult fires di•ffi Visualizationcult of different Visualization of different types of • • General comparison between areas Overlapping points are not visible residential fires in points and colors • General• comparison • Overlapping points are• not types of residential fires in Addresses and residential buildings Confidentiality—easy to identify single General spatial analysis of residential • • • betweencan areas be identified visible addresses—notpoints and compatible colors with fires—overall comparison Applied on subcategories of fires Swedish and European legislation on Point of departure for area-based • Addresses and residential • Confidentiality—easy to • General spatial analysis of • • the right to the protection of personal residential fire prevention buildings can be identified identify single addresses—not residential fires—overall data (GDPR) if the map is presented on Spatio-temporal analysis of • • Applied on subcategories compatible with Swedish anda dynamic comparison GIS platform residential fires ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 9 of 18 of fires European legislation on the • Point of departure for area- Heat map right to the protection of based residential fire Point map • Concentrations per area • No normalization per • Quick and efficient Concentrations per area unit (e.g., km2) No normalization per Quick and efficient visualization and • personal data (GDPR)• if the prevention • unit (e.g.,Can km provide2) informationinhabitant—skews about inhabitant—skewsvisualization visualization and in localization of residential fire clusters • statistically significantmap clusters is presented on a relation• Spatio-temporal to the population analysis of Analysis of the geographic density of • Can provide information visualization in relation to the localization of residential • Facilitates comparisons between No exact positions residential fires • dynamic GIS platform• residential fires about statisticallydifferent fire-exposed areaspopulation Requiresfire many clusters points as input Developed basis for area-based • • significantCan clusters easily be plotted• in No 3D exact positions • Analysis of the geographic residential fire prevention—hot spots • Applied on subcategories of fires and classifications per area unit (e.g., • Facilitates• comparisons • Requires many points as input density of residential fires km2) Spatio-temporal analysis of between different fire- • Developed basis for area- • residential fires exposed areas based residential fire Heat map • Can easily be plotted in 3D prevention—hot spots and • Applied on subcategories classifications per area unit of fires (e.g., km2) • Spatio-temporal analysis of residential fires

• Value per area—a • "Ecological fallacy"—the area • Geo-statistical analysis and quantitative perception and value does not correspond to comparisons of residential classification of fires all individual households and fires and other independent • Normalized values, e.g., individuals in the mapped area variables within and per 1000 inhabitants • Areas may be heterogeneous between cities and regions • Easy and relatively regarding different types of • Developed basis for area- inexpensive to find residents based residential fire subarea based statistics • No exact positions prevention—classifications Choropleth map • Does not reveal addresses • Area divisions may vary per normalized area and individuals between cities and regions division (e.g., subarea) • Applied on subcategories • Spatio-temporal analysis of of fires residential fires

• Data aggregation without • Certain "ecological fallacy"— • Developed basis for directly revealing the area value does not accurate area-based individual addresses, correspond to all individual residential fire households and households and individuals in prevention—classifications individuals—at least for the mapped area in normalized grid smaller grids • The grid does not follow • Identification of smaller • Normalized values, e.g., natural geographical fire risk Grid map per 1000 inhabitants boundaries areas/neighborhoods • Easy to compare between • No exact positions • Geo-statistical analysis and different grids and cities • Can be relatively expensive to comparisons of residential • Advantageous for finding acquire grid based statistics fires and other independent

specific fire risk from, e.g., Statistics Sweden variables within and areas/neighborhoods • Processing large data sets can between cities and regions (outside large be time consuming • Spatio-temporal analysis of administrative areas) residential fires • Applied on subcategories of fires

ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 9 of 18

• Concentrations per area • No normalization per • Quick and efficient unit (e.g., km2) inhabitant—skews visualization and ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 9 of 18 • Can provide information visualization in relation to the localization of residential about statistically population fire clusters • Concentrations per area • No normalization per • Quick and efficient significant clusters • No exact positions • Analysis of the geographic unit (e.g., km2) inhabitant—skews visualization and • Facilitates comparisons • Requires many points as input density of residential fires • Can provide information visualization in relation to the localization of residential between different fire- • Developed basis for area- about statistically population fire clusters exposed areas based residential fire significant clusters • No exact positions • Analysis of the geographic Heat map • Can easily be plotted in 3D prevention—hot spots and • Facilitates comparisons • Requires many points as input density of residential fires • Applied on subcategories classifications per area unit between different fire- • Developed basis for area- of fires (e.g., km2) exposed areas based residential fire • Spatio-temporal analysis of ISPRS Int. J.Heat Geo-Inf. map2020 , 9, 355 • Can easily be plotted in 3D prevention—hot spots and 16 of 19 residential fires • Applied on subcategories classifications per area unit • ofValue fires per area—a • "Ecological fallacy"—theTable area 3. Cont• (e.g.,Geo-statistical. km2) analysis and quantitative perception and value does not correspond to • Spatio-temporalcomparisons of residentialanalysis of Visualization Strenghts Weaknesses Application Fields—Example of Techniques/Methods classification of fires all individual households and residentialfires and other fires independent Residential Fires Choropleth map • Normalized values, e.g., individuals in the mapped area variables within and • Value per area—a • "Ecological fallacy"—the area • Geo-statistical analysis and Value per area—a quantitative "Ecological fallacy"—the area value Geo-statistical analysis and comparisons per• 1000 inhabitants • Areas may be heterogeneous• between cities and regions • quantitativeperception perception and and classification value does of fires not correspond todoes notcomparisons correspond of to residential all individual of residential fires and other • Easy andNormalized relatively values, e.g.,regarding per different types ofhouseholds • Developed and individuals basis for area- in the independent variables within and classification• of fires all individual households and fires and other independent inexpensive1000 to inhabitants find residents mappedbased area residential fire between cities and regions • NormalizedEasy values, and relatively e.g., inexpensiveindividuals to in find the mapped areaAreas mayvariables be heterogeneous within and regarding Developed basis for area-based subarea• based statistics • No exact positions • prevention—classifications • per 1000subarea inhabitants based statistics• Areas may be heterogeneousdi fferentbetween types of cities residents and regions residential fire Choropleth map • Does notDoes reveal not addresses reveal addresses • Area divisions may vary No exactper positions normalized area prevention—classifications per • • • Easy andand relatively individuals regarding different types ofArea divisions• Developed may basis vary for between area- cities normalized area division (e.g., subarea) and individuals between cities and regions• division (e.g., subarea) and regions Spatio-temporal analysis of inexpensiveApplied to find on subcategoriesresidents of fires based residential fire • • Applied• on subcategories • Spatio-temporal analysis of residential fires subarea based statistics • No exact positions prevention—classifications of fires residential fires ChoroplethGrid map map • Does not reveal addresses • Area divisions may vary per normalized area • andData individuals aggregationData aggregation without without• betweenCertain directly "ecological cities and regions fallacy"— Certain• "ecologicaldivisionDeveloped (e.g., basis fallacy"—the subarea) for area Developed basis for accurate area-based • • • revealing individual addresses, value does not correspond to all residential fire • Applieddirectly revealingon subcategories the area value does not • Spatio-temporalaccurate area-based analysis of households and individuals—at least individual households and individuals prevention—classifications in ofindividual fires for addresses, smaller grids correspond to all individualin the mappedresidential area firesfire normalized grid Normalized values, e.g., per The grid does not follow natural Identification of smaller fire risk households• and households and individuals• in prevention—classifications • • Data aggregation1000 inhabitants without • Certain "ecological fallacy"—geographical• Developed boundaries basis for areas/neighborhoods individuals—atEasy to least compare for betweenthe mapped different area No exactin positions normalized grid Geo-statistical analysis and comparisons directly• revealing the area value does not• accurate area-based • grids and cities Can be relatively expensive to acquire of residential fires and other smaller grids • The grid does not follow• • Identification of smaller individualAdvantageous addresses, for findingcorrespond specific to fire all individualgrid basedresidential statistics fire from, e.g., independent variables within and • Normalized• values, e.g., natural geographical fire risk between cities and regions householdsrisk and areas /neighborhoodshouseholds (outside and large individualsStatistics in prevention—classifications Sweden per 1000administrative inhabitants areas) boundaries Processingareas/neighborhoods large data sets can be Spatio-temporal analysis of Grid map • • individuals—atApplied least on subcategoriesfor the mapped of fires area time consumingin normalized grid residential fires • Easy• to compare between • No exact positions • Geo-statistical analysis and smaller grids • The grid does not follow • Identification of smaller different grids and cities • Can be relatively expensive to comparisons of residential • Normalized values, e.g., natural geographical fire risk • Advantageous for finding acquire grid based statistics fires and other independent Grid map per 1000 inhabitants boundaries areas/neighborhoods specific fire risk from, e.g., Statistics Sweden variables within and • Easy to compare between • No exact positions • Geo-statistical analysis and areas/neighborhoods • Processing large data sets can between cities and regions different grids and cities • Can be relatively expensive to comparisons of residential (outside large be time consuming • Spatio-temporal analysis of • Advantageous for finding acquire grid based statistics fires and other independent administrative areas) residential fires specific fire risk from, e.g., Statistics Sweden variables within and • Applied on subcategories areas/neighborhoods • Processing large data sets can between cities and regions of fires (outside large be time consuming • Spatio-temporal analysis of administrative areas) residential fires • Applied on subcategories of fires

ISPRS Int. J. Geo-Inf. 2020, 9, 355 17 of 19

4. Conclusions The aim of this study has been to critically examine how different geovisualization techniques—point data, kernel density and choropleth mapping—can complement each other be applied in fire preventive work. The combination of these visualization techniques facilitates various forms of fire prevention on different geographical scales, such as identification and targeting of different vulnerable groups, strategic selection of neighborhoods before home visits, fire safety supervisions for citizens, information campaigns, targeting of residential fire-prone areas to match different preventive strategies, and identifying certain types of fires in different areas. The study also shows how some of these techniques can be applied when analyzing residential fire incidents and their relation to underlying structural and socio-economic factors as well as spatio-temporal dimensions of fire incident data. Although visual spatial analysis does not necessary demand high technical skills among emergency service employees, there is a risk of misinterpretation and incorrect strategic decisions. However, the combination of different visualizations of residential fires increases the scope for interpretation and the possibilities of targeting various forms of preventive measures to specific neighborhoods, or even residential buildings with higher risk of residential fires. There are still issues that need to be resolved, e.g., how uncertainty in map visualizations of residential fires can be highlighted and further analyzed, especially in relation to spatio-temporal variations of fires. The weaknesses presented in Table3 can be used as a direct input to further discussions and mappings of uncertainty in spatial and temporal trends of residential fires. Other important issues include how the fire data how can be further explored, quality assured and protected, or how different convenient GIS-based methods and interpretation skills can be transformed and come to practical use in emergency services’ fire preventive work.

Funding: This work was funded by the Swedish Civil Contingencies Agency (grant number Dnr 2013-2657). Acknowledgments: This study has been conducted within the Swedish research project Residential Fires in Metropolitan Areas: Spatial Differences and Fire Safety Work in the Socially Fragmented city, 2014–2018. Special thanks to my colleagues Mona Tykesson, Per-Olof Hallin and Jerry Nilsson for their contributions and invaluable comments during the project. Special thanks also to the Emergency Services South (Malmö), Emergency Services Stockholm, Emergency Services Södertörn (Stockholm) and Emergency Services Gothenburg for their supply of data and professional comments. Thanks to Dalena Tran for providing language help during the completion of this paper. Conflicts of Interest: There is no conflict of interest. The funder had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

References

1. Kristianssen, A.-C.; Andersson, R.; Belin, M.-Å.; Nilsen, P. Swedish Vision Zero policies for safety—A comparative policy content analysis. Saf. Sci. 2006, 103, 260–269. [CrossRef] 2. Guldåker, N.; Hallin, P.O.; Tykesson, M.; Nilsson, J. Fire Safety Work in the Social Fragmented City, Residential Fires in Metropolitan Areas Series; In Swedish with the Title: Brandsäkerhetsarbete i den Socialt Fragmenterade Staden. Rapport 4 i serien Bostadsbränder i storstadsområden—BIS; Malmö University: Malmö, Sweden; Lund University: Lund, Sweden, 2018; p. 4. ISBN 978-91-7895-166-6. 3. Corcoran, J.; Higgs, G.; Brunsdon, C.; Ware, A.; Norman, P. The use of spatial analytical techniques to explore patterns of fire incidence: A South Wales case study. Comput. Environ. Urban Syst. 2007, 31, 623–647. [CrossRef] 4. Corcoran, J.; Higgs, G.; Higginson, A. Fire incidence in metropolitan areas: A comparative study of Brisbane (Australia) and Cardiff (United Kingdom). Appl. Geogr. 2011, 31, 65–75. [CrossRef] 5. Kc, K.; Corcoran, J. Modelling residential fire incident response times: A spatial analytic approach. Appl. Geogr. 2017, 84, 64–74. [CrossRef] 6. Asgary, A.; Ghaffari, A.; Levy, J. Spatial and temporal analyses of structural fire incidents and their causes: A case of Toronto, Canada. Fire Saf. J. 2010, 45, 44–57. [CrossRef] ISPRS Int. J. Geo-Inf. 2020, 9, 355 18 of 19

7. Špatenková, O.; Virrantaus, K. Discovering spatio-temporal relationships in the distribution of building fires. Fire Saf. J. 2013, 62, 49–63. [CrossRef] 8. Chainey, S. Using the vulnerable localities index to identify priority areas for targeting fire safety services. Fire Saf. J. 2013, 62, 30–36. [CrossRef] 9. Guldåker, N.; Hallin, P.-O. Spatio-temporal patterns of intentional fires, social stress and socio-economic determinants: A case study of Malmö, Sweden. Fire Saf. J. 2014, 70, 71–80. [CrossRef] 10. MacEachren, A.M.; Kraak, M.J. Research Challenges in Geovisualization. Cartogr. Geogr. Inf. Sci. 2001, 28, 3–12. [CrossRef] 11. Krisp, J.M.; Virrantaus, K.; Jolma, A. Using explorative spatial analysis methods in a GIS to improve fire and rescue services. In Geo-information for Disaster Management; Oosterom, P., Zlatanova, S., Fendel, E.M., Eds.; Springer: Berlin, Germany, 2005; pp. 1282–1296. ISBN 978-3-540-27468-1. 12. Wuschke, K.; Clare, J.; Garis, L. Temporal and geographic clustering of residential structure fires: A theoretical platform for targeted fire prevention. Fire Saf. J. 2013, 62, 3–12. [CrossRef] 13. Higgins, E.; Taylor, M.; Jones, M.; Lisboa, P. Understanding community fire risk—A spatial model for targeting fire prevention activities. Fire Saf. J. 2013, 62, 20–29. [CrossRef] 14. Ferreira, T.M.; Vicente, R.; Da Silva, J.A.R.M.; Varum, H.; Costa, A.; Maio, R. Urban fire risk: Evaluation and emergency planning. J. Cult. Heritage 2016, 20, 739–745. [CrossRef] 15. Zhang, X.; Yao, J.; Sila-Nowicka, K.; Jin, Y. Urban Fire Dynamics and Its Association with Urban Growth: Evidence from Nanjing, China. ISPRS Int. J. Geo-Inf. 2020, 9, 218. [CrossRef] 16. Martín-Gómez, C.; Vergara-Falces, J.; Elvira-Zalduegui, A. Geographic information system software application developed by a regional emergency agency. Case Stud. Fire Saf. 2015, 4, 19–27. [CrossRef] 17. IKAROS Daedalos. Available online: www.daedalos.org (accessed on 27 April 2020). 18. IDA MSB. Available online: https://ida.msb.se (accessed on 27 April 2020). 19. Regulation (EU) 2016/679 of the European Parliament and of the Council. Available online: https://eur-lex. europa.eu/legal-content/ENG/TXT/PDF/?uri=CELEX:32016R0679&rid=1 (accessed on 27 April 2020). 20. Ataei, M.; Degbelo, A.; Kray, C.; Santos, V. Complying with Privacy Legislation: From Legal Text to Implementation of Privacy-Aware Location-Based Services. ISPRS Int. J. Geo-Inf. 2018, 7, 442. [CrossRef] 21. Keßler, C.; McKenzie, G. A geoprivacy manifesto. Trans. GIS 2017, 22, 3–19. [CrossRef] 22. Civil Protection Act 2003:778. Available online: https://www.riksdagen.se/sv/dokument-lagar/dokument/ svensk-forfattningssamling/lag-2003778-om-skydd-mot-olyckor_sfs-2003-778 (accessed on 27 April 2020). 23. Hallin, P.-O.; Guldåker, N.; Tykesson, M.; Nilsson, J. Residential fires in metropolitan areas—Theoretical perspectives. In Residential Fires in Metropolitan Areas Series; In Swedish with the title: Bostadsbränder i Storstadsområden—Teoretiska Utgångspunkter. Rapport 2 i serien Bostadsbränder i storstadsområden—BIS; Malmö University: Malmö, Sweden; Lund University: Lund, Sweden, 2018; p. 2. ISBN 978-91-7895-165-9. 24. Facts about Malmö. Available online: https://malmo.se/Fakta-och-statistik.html (accessed on 27 April 2020). 25. Population Statistics on the Municipality of Burlöv. Available online: https://burlov.se/kommunpolitik/ omburlov/%20befolkningochstatistik.4.4a8b2c6a159cac1506d4a161.html (accessed on 27 April 2020). 26. GIS-Centre—Geodata Available for Lund University. Available online: https://www.gis.lu.se/geographical- data/geografical-data-for-university-students-and-staff (accessed on 27 April 2020). 27. M-Database—A Research Database at Malmö University and Lund University. Available online: https://www.keg.lu.se/en/research/research-projects/current-research-projects/m-database-a-research- database-at-malmo (accessed on 27 April 2020). 28. Jennings, C.R. Social and economic characteristics as determinants of residential fire risk in urban neighborhoods: A review of the literature. Fire Saf. J. 2013, 62, 13–19. [CrossRef] 29. Putnam, C.T.; Kirkpatrick, J.T. Juvenile Firesetting: A Research Overview. Juvenile Justice Bulletin. 2005. Available online: https://www.ncjrs.gov/pdffiles1/ojjdp/207606.pdf (accessed on 27 April 2020). 30. Kobes, M.; Helsloot, I.; De Vries, B.; Post, J.G. Building safety and human behaviour in fire: A literature review. Fire Saf. J. 2010, 45, 1–11. [CrossRef] 31. FEMA. Residential Building fires (2013—2015). Top. Fire Rep. Ser. 2017, 18, 1–15. 32. Guldåker, N.; Hallin, P.O.; Tykesson, M.; Nilsson, J. Fire safety work in the social fragmented city. In Residential Fires in Metropolitan Areas Series; In Swedish with the tile: Brandsäkerhetsarbete i den socialt fragmenterade staden. Rapport 5 i serien Bostadsbränder i storstadsområden—BIS; Malmö University: Malmö, Sweden; Lund University: Lund, Sweden, 2018; p. 4. ISBN 978-91-7895-168-0. ISPRS Int. J. Geo-Inf. 2020, 9, 355 19 of 19

33. Mitchell, A. The ESRI Guide to GIS Analysis. Volume 2: Spatial Measurements; Statistics; Esri Press: Redlands, CA, USA, 2009, ISBN 9781589481169. 34. Anderson, T.K. Kernel density estimation and K-means clustering to profile road accident hotspots. Accid. Anal. Prev. 2009, 41, 359–364. [CrossRef] 35. Xie, Z.; Yan, J. Kernel Density Estimation of traffic accidents in a network space. Comput. Environ. Urban Syst. 2008, 32, 396–406. [CrossRef] 36. Nilsson, J.; Hallin, P.-O.; Tykesson, M.; Guldåker, N. Differences in fire occurrences within and between different Swedish metropolitan areas—A statistical analysis. In Residential Fires in Metropolitan Areas Series; In Swedish with the title: Skillnader i brandförekomst inom och mellan olika storstadsområden—En statistisk analys. Rapport 4 i serien Bostadsbränder i storstadsområden—BIS; Malmö University: Malmö, Sweden; Lund University: Lund, Sweden, 2018; p. 4. ISBN 978-91-7895-167-3. 37. Tykesson, M.; Nilsson, J.; Guldåker, N.; Hallin, P.-O. A quality review of the reporting of residential fires in the metropolitan areas of Malmö, Gothenburg, Södertörn and Stockholm. In Residential Fires in Metropolitan Areas Series; In Swedish with the title: Kvalitetsgranskning av insatsrapportering av bostadsbränder: Storstadsområdena Malmö, Göteborg, Södertörn och Stockholm. Rapport 1 i serien Bostadsbränder i storstadsområden—BIS; Malmö University: Malmö, Sweden; Lund University: Lund, Sweden, 2016; p. 1. ISBN 978-91-7895-164-2. 38. Demšar, U. Investigating visual exploration of geospatial data: An exploratory usability experiment for visual data mining. Comput. Environ. Urban Syst. 2007, 31, 551–571. [CrossRef] 39. MacEachren, A.M. Visualizing Uncertain Information. Cartogr. Perspect. 1992, 13, 10–19. [CrossRef] 40. MacEachren, A.M.; Robinson, A.; Hopper, S.; Gardner, S.; Murray, R.; Gahegan, M.; Hetzler, E. Visualizing Geospatial Information Uncertainty: What We Know and What We Need to Know. Cartogr. Geogr. Inf. Sci. 2005, 32, 139–160. [CrossRef] 41. Lucchesi, L.R.; Wikle, C.K. Visualizing uncertainty in areal data with bivariate choropleth maps, map pixelation and glyph rotation. Stat 2017, 6, 292–302. [CrossRef] 42. Alberti, K. Web-based Visualization of Uncertain Spatio-Temporal Data. Master’s Thesis, Utrecht University, Utrecht, The Netherlands, 2013. Available online: http://karssenberg.geo.uu.nl/_static/Alberti_K_Thesis_ Webbased_Uncertainty_Visualization.pdf (accessed on 22 May 2020). 43. Cheong, L.; Bleisch, S.; Kealy, A.; Tolhurst, K.; Wilkening, T.; Duckham, M. Evaluating the impact of visualization of wildfire hazard upon decision-making under uncertainty. Int. J. Geogr. Inf. Sci. 2016, 30, 1377–1404. [CrossRef] 44. Amatulli, G.; Pérez-Cabello, F.; De La Riva, J. Mapping lightning/human-caused wildfires occurrence under ignition point location uncertainty. Ecol. Model. 2007, 200, 321–333. [CrossRef]

© 2020 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).