An augmented reality study for public participation in urban planning

ABSTRACT Ongoing urbanisation processes invoke immense construction activities, for which citizens often participate in planning. Yet, imagining planned buildings based on visual representations is a highly demanding task. While traditional methods, such as construction spans, 2D, or 3D visualisation often fail to offer a complete picture, we propose Augmented Reality (AR) as a more adequate tool. We first present an evaluation of the suitability of AR compared to construction spans for a future building and assess which degree of abstraction of AR is most effective, as well as difficulty of interpreting them correctly. In a between-subjects field study we compare construction spans and a prototype AR application including three levels of detail (LOD) of the same building project. Participants solve two estimation tasks using the construction spans and six estimation tasks using the AR application, before answering a questionnaire on the different visualisation methods. We find participants are confident about the potential of AR, but no significant differences between the different LOD groups in subjective assessment. Results suggest that previous knowledge (e.g. in GIS) may have a positive impact on dimension estimation performance. Also, details, such as façade elements or windows, could facilitate estimation tasks because they allow inferences about a building’s size.


Introduction
With a forecasted population share of 68% living in cities by the year 2050, 1 urban development and building development projects become ever more important. In many countries citizens have the political rights to publicly -at least partially -participate in the planning and decision-making process about future city developments. Since the spatial imagination of buildings during the planning phase is a highly demanding task, appropriate visual representations are crucial for public opinion building and decision making. Unsurprisingly, visualisations have played and play a major role in public participation in urban development. Traditional and analogue ways of communicating new building projects, such as 2D plans, perspective 3D sketches, and physical plaster or wooden models have been complemented by their digital counterparts, including 3D models (Ilin 2019).
While digital 3D models seem to be easier to understand for lay people and facilitate informed decision-making, there is also critique in relation to opinion formation and decision-making as part of political, democratic, and legal processes. This is, for example, the case in Switzerland, where major construction projects are subject to referendums. Hyper realistic 3D image visualisations of construction projects included in official voting information often show the planned building in optimal conditions (Degen, Melhuish, and Rose 2017) and from a fixed perspective, which can mean a significant discrepancy between the visualisation and reality. Approaches that enable the assessment of the planned building in its real world context are therefore desirable.
A peculiarity in Switzerland is the erection of construction spans or profiles to visualise the dimensions of a planned building and how it fits into the environment ( Figure 1). These spans remain during the building application until a legally binding building permit is granted. Since these building spans are rather difficult to interpret by untrained citizens, novel methods are recently applied in urban planning.
The adaption and commercialisation of Augmented Reality (AR) and, moreover, the integration of ARKit and ARCore into standard GIS applications offers a new means for planners and city councils to communicate with citizens Figure 1. Construction spans staked out on the project development ground in Zurich, Switzerland (47.341 N,8.519 E) on 25 th June 2020. (Alulema et al. 2018;Devaux et al., 2018b) and to visualise construction projects. AR promises novel and engaging experiences and interactions with the real world (Ramos et al. 2018). It has great potential to resolve some of the issues mentioned above and to allow citizens to explore a planned building in its real environment, in various weather and lighting conditions, and from numerous perspectives. Eventually, this should allow citizens making better informed decisions. Although AR applications have a great appeal to people, it is still far from certain whether they are also suitable for an effective communication of planning information and how easy it is for users to interpret such information during their decision making. Also, the question of adequate representation of a virtual construction project, and the selection of appropriate level of details (LOD), and degree of realism are yet unanswered.
Construction spans are a common means in Switzerland to visualise a planned building in 3D at the building's future location. With this research we aim at evaluating the suitability of AR compared to construction spans as a method to visualise planned construction projects. Therefore, we aim to compare the two methods for visualising a future building in 3D in the field. We examine how effective construction spans are and if the population in Switzerland can interpret construction spans correctly. More specifically, we investigate how different LOD of a 3D visualisation in an AR application influence the ease of interpretation of a visualisation and, therefore, which LOD is best used for this purpose. We pursue the following research questions: RQ 1: How suitable is an AR application for visualising new building projects in an understandable and realistic way to assist decision-making compared to construction spans?
RQ 2: How difficult is it for users to interpret the visualisation of a building project in an AR application correctly? RQ 3: What is the most effective degree of abstraction for visualising a planned building project in an AR application?
In order to answer these research questions, we conducted a user study with two projects in different planning stages located in a development zone in the south of the city of Zurich (47.341 N,8.520 E). Thirty participants were split into three groups and exposed to three different LODs and had to solve various rating tasks. In addition, participants had to rate the perceived suitability of the visualisation methods, assessed in a questionnaire, which is compared to their task performance. Finally, it is investigated how the different LODs influence the participants' decision outcome. With our findings we expect to offer a better knowledge base for selecting suitable LODs for an AR visualisation for future construction projects.

Related work
While for a long time, most Location-Based Services (LBS) have been designed for gesture-based interaction with touch displays, novel visual, auditory, and tactile communication modes have gained increasing interest recently (Huang et al. 2018). Here, we focus on visual communication through AR. LBS research has explored the use of AR for different application scenarios, including pedestrian navigation (Walther-Franks and Malaka 2008), tourist guidance (Reitmayr and Schmalstieg 2004) and map interactions (Rudi et al. 2016). While early papers have mainly focused on the technical implementation of AR in LBS (Schmalstieg and Reitmayr 2007;Feiner et al. 1997), more recent work has investigated issues such as visualisation techniques (Cron et al. 2019), spatial knowledge acquisition (Huang, Schmidt, and Gartner 2012), and user experience. For instance, it has been demonstrated that digital maps and audio interfaces can lead to higher performance and user experience in navigation than AR (Rehrl et al. 2014). Here, we consider an AR-enhanced LBS that aims at supporting citizens in their decision-making process for urban planning.
Visualising 3D data is fundamental for urban planning processes as it can support the design of urban spaces, assists understanding (Devaux et al. 2018b;Fonseca et al. 2014), and leads to better decisions (Appleton and Lovett 2003;Rohrmann and Bishop 2002). A graspable visualisation of complex urban spaces supports visuospatial thinking and cognitive processes (Devaux et al. 2018a;Döllner, Baumann, and Buchholz 2006;Drettakis et al. 2007), and is ultimately key for a good understanding of a presented project to facilitate decisionmaking processes (Hu et al. 2015). Public participation is widely incorporated into planning processes (Davies et al. 2012). However, conventional public participation methods are unsuitable for many people since they are predominantly based on more abstract and static representations, such as 2D plans or pictures of rendered 3D models, and, therefore, there is a need for new methods (Foth et al. 2009;Wilson, Tewdwr-Jones, and Comber 2019).
Novel methods that make use of web or mobile applications allow for more engagement of citizens within their environment and for empowerment (Narooie 2014). Such applications have the potential to improve public participatory planning because they are more interactive (Bugs et al. 2010) and offer more choices of what content to see and from which perspective (Hanzl 2009). Modern visualisation methods enable citizens to better comprehend the consequences of planning choices as they allow for more interactive engagement and a more customised perspective-taking (Foth et al. 2009).
Among these novel methods, mobile AR applications are particularly interesting for public participation in planning, since they enable citizens to view planned projects any time and on site without the need to be present at a planning meeting (Allen, Regenbrecht, and Abbott 2011). Zhang et al. (2018), for instance, could show that visualising 3D city models in mobile AR applications is more advantageous compared to using traditional computer screens. Cirulis and Brigmanis (2013) found that AR applications can reduce financial resources before construction works are commenced or completed and can increase citizens' satisfaction, as the public interests are integrated in discussion processes and decision-making processes are made more transparent. The opportunity to intuitively engage with a planned project in the real urban context and obtain additional information has been identified as a major advantage of mobile AR applications (Kaji et al. 2018). Fonseca et al. (2014) mentioned the ability to easily see and compare alternative project proposals and scenarios before construction work starts as another benefit of mobile AR applications. Along the same lines, Baek, Ha, and Kim (2019) emphasise that AR applications may 'compensate for the weaknesses of ineffective verbal communication, time-consuming data accessibility, and distraction caused by domain switching'. Chu, Matthews, and Love (2018) could show that information retrieval processes improved as the mental workloads were lowered thanks to the AR applications and users were able to complete tasks with minimal error.
However, research also documents caveats and unsolved challenges of using AR. For one, current AR applications are still too abstract for most users (Schaffers et al. 2011). In studying the use of AR in education, Wu et al. (2013) reported that students might show a very high extraneous cognitive load introduced by technology, the large amount of presented information, or task complexity. For many applications AR devices (e.g. head mounted devices) still only recognise limited user gestures, which can cause user exhaustion (Zhang et al. 2018).
Although several researchers have tested their AR applications in real urban context (Tom Dieck, Jung, and Han 2016;Kaji et al. 2018;Kälin 2015;Lee et al. 2012;Ramos et al. 2018;Vlahakis et al. 2001;Zhang et al. 2018) and despite these AR applications being increasingly adopted and readily accepted, particularly among young people (Kaji et al. 2018;Vlahakis et al. 2001), their acceptance has barely been evaluated (Tom Dieck and Jung 2018). Evaluation results of AR systems in user studies are often based on self-reported data (Lee et al. 2012;Ramos et al. 2018). Therefore, experimentations with actual users and statistical results are needed (Devaux et al. 2018b;Ramos et al. 2018). Additionally, it is necessary to evaluate AR applications in a real urban environment and not only in a controlled environment (Kaji et al. 2018).
AR applications in urban planning fundamentally depend on the quality of visualisation. For computer simulated urban models, Radford et al. (1997) identified three distinct qualities: accuracy, abstraction, and realism. Accuracy describes the difference between a recorded value and its true value (Jones 1997). Abstraction refers to the amount of detail included within a scene. A common way of distinguishing level of detail of buildings is the LOD concept defined in CityGML. CityGML is an Open Geospatial Consortium standard 2 and stores 3D building models in Geography Markup Language (GML), an XML dialect for spatial data encoding, or as JSON files. The CityGML concept discerns five LODs, ranging from LOD 0 to LOD 4 (Biljecki et al. 2016). The lowest level of detail, LOD 0, is simply a 2D footprint of a building. LOD 1 represents a building as a simple 3D block. LOD 2 adds further elements to the block, such as a roof, and LOD3 represents the exterior hull, as well as features, such as windows and doors. The most detailed level, LOD 4, includes also interior structures (Löwner et al., 2018). LODs have very practical implications, since increasing the level of abstraction means less information to be represented and usually yields smaller file sizes which may be useful for mobile applications (Jahnke, Krisp, and Kumke 2011). Other popular file formats for 3D models to be visualised in AR apps are filmbox (fbx) from Autodesk, the DAE format from Collada's, or the Universal Scene Description ZIP (.USDZ) format developed by Pixar.
Realism, the third quality property identified by Radford et al. (1997), refers to how convincingly the objects are modelled and represent the real object. Elements of realism are typically colours and textures. Appleton and Lovett (2003) investigated levels of realism and let participants rate different visualisation alternatives of an environmental scene. They were not able to define sufficient level of realism, rather they found that a minimal degree of realism is indispensable in order to relate to a visual scene and form an opinion on it. Visual scenes of high realism-abstraction are inadequate (Daniel and Meitner 2001). However, highly realistic scenes might give the illusion of precision. In an urban planning context, rendering visual properties in optimal conditions may create highly subjective visualisations although visual scenes should be represented as objectively as possible (Day 2002).
Regarding realism, virtual objects in augmented scenes can either be rendered as realistically as possible, or just the opposite, in a non-photorealistic manner (Haller 2004). An important effect for realistic AR applications is illumination, which consists of shadowing and lighting effects (Haller 2004). In particular, shadows are essential for a 3D impression of a virtual scene as they give direct information about objects spatial relationships (Kolivand, Sunar, and Ji 2014;Naemura et al. 2002;Slater, Usoh, and Chrysanthou 1995). To achieve impressing photorealistic results, graphical and geometric detail is required. For example, photorealistic building façades can be generated with photo-textures (Döllner 2007). Durand (2002) suggested that the interpretation of virtual objects should rather be convincing than realistic. Döllner (2007), for instance, found that too much detail and realism can distract users and interfere with understanding visualisations. So called non-photorealistic depictions allow removing unnecessary details and emphasise relevant features of represented objects and, therefore, provide more purpose-oriented and task-oriented visualisations (Döllner 2007;Halper, Schlechtweg, and Strothotte 2002). How much detail and realism are beneficial for easily and correctly interpreting such visualisations, and how these factors might influence the outcome of decisions about planned construction projects remains unclear (Klausener 2012).

Methods
In order to answer the presented research questions in the introduction a field study had been conducted in the south of Zurich, Switzerland (47.341 N,8.520 E) between 26 th June and 31 st July 2020.

Participants
For the user study, 30 participants were recruited via mailing lists of University of Zurich and the Swiss Federal Institute of Technology in Zurich, and announcements on university webpages. The participants met the following inclusion criteria: they had to be entitled to vote in Switzerland, to have little or no previous knowledge about the planned project, and to have prior knowledge with iOS devices.
The participants of the study were composed of eighteen women and twelve men, aged between 21 and 58 years. Twelve participants grew up in a city, eighteen in the countryside. Sixteen participants have a university level degree, ten participants have a high school degree, and four an apprenticeship. Of the thirty participants, two are colour-blind and two have impaired vision.

Material
In order to get insights on the suitability of an AR application compared to construction spans, two projects in different planning phases were used. The first project (P1) was staked out with construction spans, in the south of Zurich. For the second project (P2), a prototype AR application visualising a building project located near P1, was implemented.
P2 was chosen by our collaborating partner because this project was visualised in three different LODs. P1 was chosen because of its vicinity to P2. P1 is a residential building complex consisting of four buildings with 270 flats in total ( Figure 2, left), which were visualised with construction spans (Figure 1) during our study (summer 2020). P2, used for our study and visualised in our prototype AR application, is a school building (Figure 2, right) which was at an advanced planning stage. Both projects are planned to be finished by the end of 2022.

AR prototype application
The prototype was implemented as a native AR application for iOS devices running on iOS 11 or higher, using Swift and ARKit 3 from Apple. This toolkit supports motion tracking using the device's camera(s), depth sensing including plane detection, and light estimation and perception. The building project was integrated into the AR app as a world-scale scene with ArcGIS Runtime SDK for iOS Version 100.7.0 including ArcGIS Runtime Toolkit (Esri Inc 2020a). The AR app tracks the GPS position of the viewer in a background process with a frequency of one second and an accuracy of less than ten metres, which is one of the best location update accuracies Apple's API provides. The location updates are stored with timestamps and other meta data (including the active LOD) as a log file, yielding an exact chronical sequence of the visited locations. Figure 3 depicts the prototype AR application with its interface components for the three implemented LODs refer to the definition from CityGML (Biljecki et al. 2016) of the planned school building (P2). These LODs include LOD 1 (3D block), LOD 2 (additional elements) and LOD 3 (exterior façade and texture).
Button A allows to select the LOD at the beginning of the application usage. Although an initial 3D viewpoint is set to display the content at the correct location, manual calibration needs to be applied for improving the accuracy of the virtual building object placement within the scene. Pressing the calibrate button (B) triggers two additional interface components on the image pane: a slider to change the height of the scene (D) and an additional house symbol for displaying surrounding buildings (E). These allow to move the entire scene up and down and adjust the physical (device camera) content from the camera feed by changing X and Y positions via gestures, respectively. Pressing the sun symbol (C) enables to adjust the ambient lighting of the scene based on a specified date and time and thus controlling for the shadows casted onto the virtual buildings. Finally, the switch button (F) displays or hides the basemap that serves only a reference for the current orientation of the virtual content. The visual scene of the AR application is based on the swissALTI3D digital elevation model (Swisstopo 2020). The World Imagery basemap (a high resolution satellite and aerial image data layer) from ArcGIS Online (Esri Inc 2020b) is laid over the elevation model ( Figure 4). The basemap has the coordinate system WGS 84 Web Mercator (Esri Inc 2020b) and was clipped to the study area to minimise storage size. The data for the planned school building were delivered by the city of Zurich based on the planning data models provided by the commissioned architectural company in Autodesk Filmbox format (Zürich and Für Hochbauten 2019). The data include the school building in LOD 1 and LOD 2, visualised in white, and LOD 3, displayed in colours in which the complete building will shine (Figure 4).  The final element in the virtual scene are the buildings surrounding the planned school building. These buildings were extracted from the freely available 3D city model of Zurich (Zürich 2020) and are visualised in LOD 1. Eventually, the digital elevation model, basemap, school building, and the surrounding buildings were projected into WGS 84 coordinate system and packaged into a mobile scene package, which is needed for using the AR application in offline mode. The AR application was locally installed on an iPhone 6s (iOS 13.5).

Questionnaires for qualitative feedback
An online questionnaire for collecting qualitative feedback was created using ArcGIS Survey123 from Esri Inc. (Esri Inc 2020c). The questionnaire was downloaded onto the Survey123 Connect App installed on an iPad (6th generation; iOS version 13.7), which was used for offline data collection.
The questionnaire consists of 21 questions grouped into following categories: general questions about visualisation, accuracy of visualisation, comparison of AR and construction spans, AR and participation, and individual opinion on the project.

Study design
The goal of this user study was to investigate the suitability of an AR application and to explore the different LODs of a virtual construction project. To answer these questions, we followed a between-subjects design. As independent variable we defined the LOD of the planned school building, which takes the three different levels corresponding to LOD 1, LOD 2, and LOD 3 (Figure 4). The LODs were assigned to three groups, meaning that each participant saw only one LOD. This has the advantage that the individual sessions are shorter and danger of distraction, tiredness, fatigue can be minimised. As dependent variables we measured performance of participants in several estimation tasks and the time for task completion, as well as their ratings and answers to the questionnaire. We controlled for the order of the tasks and, thus, for the order of the methods used for visualising the planned buildings (i.e. construction spans and AR visualisation). Study participants were randomly assigned to the three LODs.

Experimental procedure
Prior to the study session, participants had to fill in and send a digital consent form. The study started at a nearby train station where participants were introduced to the aim of the study and were told background information about the school building project ( Figure 5; marked as M). Next, participants had to fill in a digital questionnaire asking for demographic data and prior knowledge in related fields, such as urban planning, architecture, geographic information systems, augmented and virtual reality technology ( Figure 5; marked as A). Then, participants were guided to the first study location where Figure 5. Study locations at the construction sites for P1 and P2. The participants were welcomed near the train station (M), then had to fill in a consent form and received some information nearby the train station (A), were guided to the first task location near P1 (C) and lastly were guided to the final project site P2 where estimation tasks (I) and occlusion tasks (II) had to be performed. a construction project was staked out with construction spans (Figures 1 and 5; marked as C) and asked to estimate the height of the construction spans and estimate how many buildings are staked out.
Afterwards, participants were guided to the second site, the construction site of the school building. There, participants were asked to estimate the length of a concrete block and the height of a residential building southeast of the construction site, which served as a baseline for the participants' estimation skills. Then, the AR application was first shown to the participants. On start-up only the basemap was displayed. The LOD assigned to the participant was selected from the application interface and the application was calibrated for the participants by the experimenter. Once calibrated, the participants were asked to solve another three different estimation tasks using the AR application ( Figure 5; marked as I). The participants were free to use all interface elements apart from the calibration elements. However, for the estimation tasks, the participants did not use these elements.
The first of these tasks involved estimating height, width and length of the virtual school building and the distance between the school building and the residential building to its south. This task was solved southwest of the construction site, in front of an opening in the barrier walls fencing the construction site ( Figure 5; marked as I).
The final two tasks had to be solved at another location south of the construction site (Figures 5 and 6; marked as II). There, participants had to mark on an aerial photograph all buildings which will be occluded by the finished school building ( Figure 6). Finally, at this second location, participants were shown the same construction spans they had looked at earlier, and they had to decide whether the staked-out building will be occluded by the finished school building or not. For all tasks, the elapsed time during task fulfilment was recorded.
Once participants had solved all the tasks, they were free to explore the application by walking around the construction site (P2). After a maximum of 10 minutes of free exploration, participants had to fill in the final questionnaire.

Results
The participants were randomly split into three groups according to the LOD they would view during the study. The groups are referred to as group LOD 1, group LOD 2 and group LOD 3. Prior knowledge of fields relevant to AR (computer games, 3D visualisations, GIS, architecture, and urban planning) was rated on a scale of 1-6, with 1 -no experience or prior knowledge, and 6 -daily use or professional knowledge (Figure 7). GIS was the field in which the average previous knowledge of the participants was highest (M = 3.0, SD = 2.05). Participants had the least previous knowledge of architecture (M = 1.87, SD = Figure 7. Aerial photograph used in the study to mark the buildings that will be occluded by the new building (undefined). 1.07) followed by urban planning (M = 2.21, SD = 1.32). Participants stated that they had little previous knowledge of computer games (M = 2.8, SD = 1.86) and 3D-visualisations (M = 2.87, SD = 1.46). Eleven out of 30 participants had no experience with AR and 14 participants had no experience with VR. Both technologies are mostly used on an annual basis (11 participants), two participants use AR on a weekly basis. In contrast, VR is not used more frequently than on a monthly basis: two participants stated they use VR on a monthly basis, 11 participants use it on an annual basis and the rest has never used this technology.
Although participants were free to use all interface elements (see Figure 3) except for the calibration feature, they made no use of the app features for the tasks they had to solve.

Baseline estimation tasks
The estimation skills of the participants were assessed by estimating the length and height of two real-world objects: The average error of estimating the length of the concrete block is 0.55 m. The estimation error for the height of the residential building nearby is 2.3 m. Neither of results of the baseline tasks show statistically significant differences between the three groups.

Estimation performance based on construction spans
On average participants estimated the number of staked out buildings quite precisely, with an error of (M = > −0.17; SD = 1.8). The mean estimation error for the height of the construction spans is − 9.6 m (SD = 6.15 m). Only three participants estimated the height of the construction spans correctly. One participant overestimated the construction spans' height, while all other participants underestimated them, by about 34% on average.

Estimations of the dimensions of the school building in the AR visualisation
The average error of estimating the height of the virtual school building by using the AR application is 6.27 m (Figure 8). Participants overall overestimated the height by about 31%. A nonparametric two-way ANOVA of aligned rank transformed estimation errors (Type III Wald F tests with Kenward-Roger df) for the factor GIS experience shows a significant difference for the factor LOD (F(26) = 5.787 p = 0.02). A post-hoc contrast test yields a significant difference between LOD 1 and LOD 2 (p = 0.02). On average participants seeing LOD 1 were fastest (23 s), followed by LOD 3 (26 s) and LOD 2 (33 s) in estimating the height. The differences are not significant (Figure 8).
With 6.57 m, the average error for estimating the width of the school building is similar to the average height estimation error (Figure 9). While overall the LOD has no significant effect on the estimation error of the width, the interaction between the LOD and the pre-knowledge of GIS has a significant effect (p = 0.01). A post-hoc test shows that the main contrast is between LOD 1 and LOD 2 for the group of participants with high (>3) knowledge of GIS, although this contrast is not significant (p = 0.09). A similar contrast was found between the groups of high and low GIS pre-knowledge in LOD 1 (0.1134). For estimating the width participants seeing LOD 1 needed on average 25 seconds, those seeing LOD 2 33 seconds. Participants seeing LOD 3 needed on average 37 seconds. The differences are not significant.
Although the magnitude of the estimation error for the length of the building was larger than for the height and width (mean −26.9 m) a two-way ANOVA revealed no significant differences between the three LOD levels ( Figure 10). Compared to estimating the height and width of the building, participants overall underestimated the true length of the virtual building (negative values indicate underestimation). Participants seeing LOD 2 were fastest in estimating the length with an average time of 19 seconds. The other participants needed 31 seconds for the LOD 1 group, and 29 seconds for the LOD 3 group respectively. Participants who underestimated or overestimated the length of the building by 1%-10% spent between 6 and 59 seconds for their estimates, those who underestimated the length by 50%-80% between 6 and 48 seconds.

Identifying occluded objects
In addition to estimating the dimensions of the school building, participants also had to identify the buildings that will be occluded by the school building. From the five hidden buildings participants on average correctly identified 3.5 buildings, misidentified 0.53, and missed 1.57. Although no significant differences could be found between the three LOD conditions, in LOD 1 participants were most successful despite the higher rate of misidentification (Table 1). For LOD 1 the variability was also highest.

Perceived suitability of AR application and construction spans
In addition to tasks performed by the study participants, we also collected subjective ratings for the visualisation of the school building with construction spans in comparison to the AR application (ease, effectiveness, accuracy).
Participants overall rated the ease of interpretation of the AR visualisation high, but the ratings showed no significant differences between the LOD conditions. Participants also had to rate (on a 5-point Likert scale, 1 -strongly disagree, 5 -strongly agree) the recognisability and realism of the school building. The average rating scores are similar for both questions, with slightly lower average values for the realism. Of interest is the low score on realism for LOD 3, compared to LOD 1 (Table 2). No significant differences could be found between the LOD conditions. Finally, participants had to assess the accuracy of the AR visualisation by rating how recognisable the real size of the school building is (Figure 12). Although no significant differences in the ratings could be found between the LOD conditions, participants in LOD 1 (M = 4.1, SD = 0.99) rated the recognisability of the real height of the building on average higher than participants in group LOD 2 (M = 3.8, SD = 0.63) and LOD 3 (M = 3.5, SD = 0.71) (Figure 11). The general rating of the potential of construction spans to visualise a construction project was lower than the ratings of the AR application above (Figure 14). Participants reported that the AR visualisation helped them to identify buildings occluded by the virtual school building. While there were no significant differences between the LOD conditions, participants in LOD 1 condition had on average performed best in the task (see Table 1) and distributed on average slightly higher scores for this question than LOD 2 and LOD 3.  Participants clearly rated the potential regarding better recognisability of the true size of the future building for an AR application over construction spans very high (Figure 11). No significant differences were found between the three LOD conditions. Very similar results were yielded for the ratings in relation the higher realism of a construction project visualised in an AR application over construction spans.
Finally, participants had to rate the perceived suitability of the visualisation of a building project in an AR application compared to construction spans ( Figure 13). The average scores are high for all LOD conditions, but no significant  difference between the LOD conditions was found. The variability of ratings in LOD 3 is echoed in qualitative feedback. Three participants think that AR applications provide a better impression of the building project, while their proportions are better recognised through construction spans. One participant mentioned that the exact positioning of a building is more accurate with construction spans, but for all other aspects an AR application is far more suitable.

Discussion
With the presented study we evaluated the suitability of an AR application to effectively visualise a construction project in comparison to the traditional visualisation method with construction spans. An assumption of our research is that an AR application is more suitable for visualising a building project, as it would be more effective for visualising the building in a realistic way. The study introduced a prototype AR application, which visualises a building project in the city of Zurich at three different levels of detail.

RQ1:
How suitable is an AR application for visualising new building projects in an understandable and realistic way to assist decision-making compared to construction spans?
The questions regarding the effectiveness and accuracy of the AR application show no statistically significant difference between the three LOD conditions. Surprisingly, participants seeing LOD 1 on average rated the realism of the visualisation highest (Table 1), even though the visualisation was the least detailed. Previous research found that the more realistic a visualisation, the higher the rating of the perceived accuracy (Klausener 2012). This discrepancy could be due to the fact that Klausener (2012) and Zanola, Fabrikant, and Çöltekin (2009) displayed their visualisations on a screen as opposed to a mobile AR in our study which suffered from occasional motion lag when walking around the virtual objects, as one participant reported. Thus, the lower rating of accuracy of participants seeing LOD 3 compared to the two other conditions could be explained by the visualised building sometimes randomly showing motion lag, which could be interpreted as inaccurate. Furthermore, the results could also be skewed towards the small screen size of the used device. Newer mobile phone and tablet devices with a bigger screen might shift this perception towards higher LOD levels, i.e. LOD 3. The lower average rating of participants in condition LOD 2 could be due to their prior knowledge in computer games (Figure 7), as these participants might have had much higher expectations regarding realism, not met by LOD 2 with its features, such as façade or windows. Although participants in condition LOD 1 expressed their wish for more details of the building, they rated the accuracy higher than participants in condition LOD 2. One reason for discrepancies between our results and findings from other studies may be in our between-subject study design where participants only saw one model and answered questions only based on the shown level of detail, without the option to compare different models.
Although participants in condition LOD 3 on average rated accuracy, recognisability of different building features, and realism of the visualisation lower than participants of the two other conditions, they rated their ability to imagine what the finished building would look like highest, since the visualisation was the most detailed one. Mainly the use of colours and depiction of external building features may have supported a better imaginability of the building. The lower rating of realism can be explained by the lack of true photorealism.
Overall, participants rated the recognisability of the real size of the building higher for the AR application than for the construction spans ( Figure 11). However, a few participants, mostly in conditions LOD 2 and LOD 3, stated that the proportions of a building are easier to assess from construction spans, and their mere physical existence are a strong and explicit indicator for a new building to be built and its dimensionality. This is in line with previous research that construction spans are a good first indicator for communicating about a construction project in a non-textual form (Ilin 2019). Nonetheless, most participants greatly underestimated the height of the building from construction spans. Perhaps, this is because construction spans are a linear representation and the lack of volume representation may give room for interpretation (Ilin 2019). AR applications could step in and complement traditional visualisation methods, such as construction spans, with more detailed information (Adascalitei and Baltoi 2018;Ilin 2019), while construction spans could act as markers for the AR applications. However, combining construction spans and an AR application could lead to an even lower rating of the accuracy of the AR application. The perceived accuracy might be lower because the precision of overlaying the construction spans with a virtual building might be low when not calibrated accurately.
Overall, the suitability ratings for the AR application were high, which is in line with previous research (Allen, Regenbrecht, and Abbott 2011;Ilin 2019). Some of the participants believe that AR applications show great potential for providing information on construction projects. In the study by Ilin (2019), the visualisation of the school building project with an AR application on a tablet was rated positively by 85% of his study participants, which was the highest ranking of all evaluated methods. In contrast, construction spans were rated lowest (Ilin 2019). In line with the results of our study we conclude that AR applications are more suitable than construction spans for visualising a building project.

RQ2:
How difficult is it for users to interpret the visualisation of a building project in an AR application correctly?
As the results of the first two baseline estimation tasks confirmed, there are no statistically significant differences between the individual groups in their estimation abilities. On average, participants in condition LOD 3 estimated the dimensions of the virtual building best. A reason might be the depiction of external features in LOD 2 and LOD 3 that, for instance, allowed to count the number of floors and, thus, deduce the height. Moreover, previous knowledge, particularly in GIS, could have had an impact on the task performance. In all three fields, participants in condition LOD 3 had the most previous knowledge. This may have helped them perform better.
Whereas height and width of the virtual building were overestimated, the length was underestimated in our study. Other research has found that egocentric distances are underestimated in virtual environments (El Jamiy and Marsh 2019; Grechkin et al. 2010;Jerome and Witmer 2005;Richardson and Waller 2005;Steinicke et al. 2010). Grechkin et al. (2010) observed that participants significantly underestimated distances in an AR environment. Furthermore, Cutting and Vishton (1995) suggested that the egocentric distance perception is not the same for the three subspaces of an observer's visual environment, i.e. personal space (≤2 m), action space (2 m-30 m) and distant space (≥30 m). People seem to underestimate egocentric distances in action and distance spaces, while overestimating them in personal spaces (El Jamiy and Marsh 2019). Egocentric distance estimations can be improved when providing feedback to observers, for example, by letting them walk half as the distance they are asked to estimate (Richardson and Waller 2005). Hence, the task performance in this study could have been improved by letting participants walk around while estimating the distances.
Since we found no correlation between the accuracy of and time spent for the estimates, time is probably not the reason for the underestimates and overestimates. A possible reason, however, could be the display size of the smartphone we used in our study. Some users felt that the display was too small to view the building model. Therefore, moving the smartphone around to view the virtual object could be a source of error for the dimension estimates. When moving around the device, a motion lag could occur because the objects were re-positioned (Woodward 2010). These drift-effects could lead to misinterpretations of the size of the presented building. However, this seems to be a technical issue which could be resolved by a different implementation, e.g. using a marker-based approach in the future to place the virtual objects.

RQ3:
What is the most effective degree of abstraction for visualising a planned building project in an AR application?
Although the ratings of various aspects of the AR application showed no significant differences between participants in the three conditions, participants in condition LOD 1 overall rated the AR application slightly higher and were more positive. A reason for this might be that participants in condition LOD 1 had least previous knowledge in urban planning, architecture, GIS, and 3D visualisation which in turn might lower their expectations of the AR application and as a consequence yield higher evaluation scores in the post-questionnaire. Previous research suggests that the effect of novelty of AR applications could have positively impacted the given answers in the post-questionnaire (Fenais et al. 2019;Olsson et al. 2013). Hence, participants with no previous knowledge of AR applications could have been biased by a temporary wow-effect which may have led to higher ratings in the post-questionnaire. However, when looking at task performance, as described above, participants in group LOD 3 scored better on average than the other participants. Participants in groups LOD 1 and LOD 2 are comparable in task performance. Therefore, a more detailed model and previous knowledge in relevant fields, such as urban planning and architecture, could assist users in tasks which involve estimating distances and volumes.
In conclusion, no specific level of detail for visualising a planned building is most effective for all purposes. Different steps in urban planning may require different visualisations. For example, a more simplistic and abstract visualisation might be more useful to inform citizens about a future construction project in the early stages of its planning process. Previous research suggested that sketches give people the impression that the model can still be modified (Klausener 2012). Additionally, Drettakis et al. (2007) suggested that an artistic visualisation might be appropriate when designers or authorities do not wish to make a definite commitment. Therefore, a simpler LOD, such as LOD 1, could facilitate the involvement of citizens in planning processes. A more detailed model, such as LOD 3, may be more useful for advanced stages of planning processes. For example, an AR application with a detailed visualisation of the project could be a useful method for providing citizens with additional information ahead of a public vote. As earlier research suggested, AR technology could be an important communication platform to make public participation more accessible to citizens and facilitate their decision-making process (Goudarznia, Pietsch, and Krug 2017).

Limitations
During data collection, construction works for the school building -in particular, foundation construction -had already started and the construction site was surrounded by white barrier walls. These white barrier walls interfered with the virtual object in the AR application also rendered in white (Figure 4) which made a distinction for some participants in condition LOD 1 and LOD 2 difficult. Moreover, some participants made use of the already existing foundation of the school building when estimating the dimensions of the virtual school building. Perhaps, the partial construction may have reduced the effect of the AR application. Ideally, the AR application should have been used before construction works have started, allowing a direct comparison between construction spans and AR application. Smaller errors in the calibration process that have led to inaccuracies and minor misalignment of the building model in the real scene, may have influenced our results. These limitations reflect the general issues found when running ecologically valid field studies. Experimental control is difficult, the environment is highly dynamic and likely to change over the time span of data collection.

Conclusion
With recent developments of AR technology, new methods for visualising and communicating construction projects are available in the planning process. The first aim of this research was to find out how suitable an AR application is for visualising a construction project and what level of detail of the virtual visualisation is the most appropriate. Visualising a construction project with AR was compared with the traditional construction spans, which are widely used to visualise construction projects in Switzerland. Participants considered the value of construction spans relatively low. Although effect and accuracy of the AR visualisation did not suggest any statistically significant differences between the three LOD conditions. However, participants who were assigned either LOD 1 or LOD 2 reported they wished to see external structures, such as windows. Our research suggests that visualising external structures helps viewers to imagine what a building will look like in the future and better evaluate whether a building fits into its surrounding environment. These observations are in line with Ilin (2019). Participants rated the suitability of construction spans as visualisation method lower than an AR application. According to the participants the suitability of both methods depends on context; construction spans are suitable as announcers, whereas an AR application can be used to communicate more details about the design of the project.
The second aim of this research was to assess participants' interpretation difficulty of a building project visualised in an AR application. In self-report measure, participants indicated that the visualisation was relatively easy to interpret. The ease of interpretation was also tested with several estimation tasks including estimating dimensions of construction spans and virtual building in an AR application. This study confirmed earlier findings by Grechkin et al. (2010) that real-world distances are underestimated when using AR. Participants who were shown LOD 1 or LOD 2 performed generally worse in estimating dimensions.
The third aim of this research was to find out what the most effective degree of abstraction is to visualise a construction project in an AR application. Our research suggests that depending on the planning stage different LODs should be used. A less detailed visualisation of a project may give viewers a less definite and less unchangeable impression and might facilitate the involvement of citizens in public participation processes. Whereas, a more detailed LOD may be used for purposes where authorities wish to make a definitive commitment.
As an overall conclusion, providing citizens with an AR application to visualise a building project could increase their willingness to be involved in public participation processes. Integrating AR technology provides an additional information source which could support citizens in their decisionmaking, for example with regard to a public voting. However, our findings may pertain to the particular use case in our study and further studies and applications need to be done. Moreover, long-term implications of the usage of AR in participatory processes remain to be investigated. There is a need to understand if participants' responses are influenced by the novelty effect of seeing an AR visualisation of a building project for the first time. Furthermore, future research should investigate the influence of other factors, such as age, on the perceived suitability of an AR application. Lastly, it remains to be investigated what participants then effectively vote for after seeing the project virtually, i.e. how strongly is the political process, and therefore democracy influenced by the technological process? Therefore, future research should also compare participants' opinion of the AR simulation and the finished building in reality.