The impact of spatial scale on layout learning and individual evacuation behavior in indoor fires: single-scale learning perspectives

Abstract The detail and representation of a spatial layout varies with scale. This affects an individual’s learning effectiveness and understanding, in turn directly influencing their behavior in a fire evacuation. However, the impact of layout learning methods with different spatial scales on fire evacuation behavior, and the relationship between spatial cognition and evacuation effects, remains unclear. We conducted spatial layout learning across three scales with 81 participants and simulated a fire evacuation scenario in a mobile virtual reality for groups. We collected evacuation decision-making and user experience questionnaires as supplementary data. The results demonstrate that small-scale learning objects are the easiest for participants to understand in terms of spatial layout and relationships, but their performance in fire evacuation is poor. Large-scale learning objects significantly improve participants’ evacuation efficiency. Spatial layout learning plays a crucial role in fire evacuation outcomes, but traditional spatial knowledge acquisition measurement methods cannot predict fire evacuation performance. This study sheds light on how spatial cognition influences fire evacuation behavior and provides a more reliable fire evacuation simulation method based on mobile virtual reality (MVR).


Introduction
Fire often exhibits characteristics of concealment and abruptness, leading to serious damage to life safety and public facilities (Cheng and Hadjisophocleous 2011).With the progression of urbanization, living and working environments are becoming increasingly dense.This densification has led to a rise in the number of fires and, consequently, an increase in fire-related casualties in recent years.The effectiveness and efficiency of evacuation during a fire directly impact the number of casualties.
Statistics show that individuals who perish in fires often fail to choose the correct evacuation method (Ronchi 2021).Therefore, studying human behavior during evacuations is of great significance to improve evacuation efficiency and promote the selection of appropriate evacuation methods.
Individual evacuation behavior is primarily influenced by four factors: building space structure, individual characteristics, environmental factors, and crowd interaction.Each of these factors includes multiple sub-factors, making individual evacuation behavior complex and variable (Fu et al. 2021).Specific aspects of the building's spatial structure affect individuals' positioning within the space and their search for exits.Individual factors include evacuees' emotional and physiological conditions; environmental factors encompass environmental temperature, smoke, signage, and so on; and crowd factors involve the impact of other evacuees on individuals (Zhiming et al. 2020).Since evacuation behavior is affected by multiple factors, its essence is making decisions to move towards a destination while comprehensively considering those factors (Philpot and Levine 2022).In this process, individuals dynamically adjust their mobility strategies due to changes in the surrounding environment (Xie et al. 2022).The changing locations of evacuees mean that they must utilize their spatial perception, spatial orientation, and wayfinding capabilities (Lin et al. 2023).Therefore, evacuation behavior is considered to be related to spatial abilities, and the wayfinding efficiency of evacuees in emergency situations largely determines the evacuation outcomes.Numerous studies have investigated how individuals navigate during the evacuation process, such as the impact of building space structure on wayfinding in evacuation (Zhang et al. 2021, Natapov et al. 2022), the effect of environmental familiarity on wayfinding (Bouguetitiche et al. 2019, Cao et al. 2019), the influence of environmental factors on wayfinding (Wang et al. 2022), and wayfinding behavior under time pressure (Zhou et al. 2020).
Wayfinding behavior is regarded as the performance and implementation of individual spatial cognition (Kim et al. 2023).Wayfinding during an evacuation is influenced by more factors than in safe environments (Løvs 1998, Merhav andFisher-Gewirtzman 2023).It is subject to both internal and external factors of the individual.Internal factors include the individual's perception and memory relating to space (Ekstrom and Hill 2023).The wayfinding process involves perceiving spatial information through various senses, analyzing the information obtained in real-time, and integrating it with memory across time and space dimensions to arrive at a final decision, a process also known as cognitive mapping (Peer et al. 2021).External factors encompass wayfinding by means of maps, signage, and other objects.
Spatial memory plays a crucial role in wayfinding.It has long been considered a manifestation of intelligence and is often associated with cognitive ability, which is also applicable to spatial memory and spatial ability (Winocur et al. 2005).Spatial memory refers to the long-term memory of the evacuation environment's layout, as opposed to short-term memory, with its content being spatial knowledge.Siegel and White (1975) proposed that the acquisition of spatial knowledge can be divided into landmark knowledge, route knowledge, and survey knowledge.Landmark knowledge includes prominent signs in the environment that are distinct from their surroundings.Route knowledge comprises sequences of landmarks or the relationships between landmarks.Survey knowledge represents a comprehensive understanding akin to maps.These three types of knowledge are progressive, with individuals who master survey knowledge considered to have higher spatial capabilities.Montello (1998) posits that the acquisition of spatial knowledge occurs through a continuous and quantitative process rather than a discrete one, and there are differences in the ability and accuracy of spatial knowledge obtained by individuals from direct experience.In fact, these differences are manifested in various ways of acquiring spatial knowledge (Van Asselen et al. 2006).During fire evacuations, the cognition of activity space by individuals is primarily determined by the spatial layout, describing the location and size of accessible paths and obstacles distributed in the evacuation environment (Dang et al. 2023).Mastery of this spatial layout by individuals determines their movement routes and evacuation efficiency, with those more familiar with the space exhibiting higher evacuation efficiency.Hegarty et al. (2006) posit that based on their relationships and dimensions, objects can be divided into figural, vista, and environmental scales for spatial layout learning, which varies by spatial scale.Although there are also national, continental, and other scales, they are predominantly understood via maps and thus provide limited utility in the evacuation process; therefore, they are not the focus of this discussion.The figural scale pertains to a small spatial scale relative to the body, such as with floor plans or sand tables, allowing individuals to easily observe the entire space from a certain angle.The vista scale is applicable when the space is equal to or larger than the body, but the entire space, for instance, a room, can still be observed without substantial movement.The environmental scale, as described by Montello (1993), necessitates a large space for the body, in which individuals cannot observe all areas from a single perspective and must rely on movement for observation.Processing spatial information at different scales engages distinct regions of the brain, with small-scale spatial information processing activating more parietal lobe structures (Kosslyn and Thompson 2003), and large-scale spatial information acquisition and processing involving the hippocampus and medial temporal lobe (Morris and Parslow 2004).Hegarty et al. (2006) further claim that spatial capabilities at different scales are partially distinct, implying that spatial layout learning at these scales also affects evacuation efficiency.However, as mentioned earlier, evacuation represents a more complex form of wayfinding behavior, so the relationship between spatial layout learning at different scales and evacuation efficiency remains unclear.In this study, we hypothesize that smaller-scale spatial learning is associated with higher evacuation efficiency.To test this assumption, we establish spatial layout learning at three scales and design these learning methods to incorporate objects commonly encountered in daily life as much as possible, aiming to provide insights into the everyday evacuation process.Research by Dong et al. (2022) demonstrates that wayfinding in virtual reality (VR) exhibits no significant difference in spatial orientation or sketching compared to reality.Therefore, we use MVR to simulate the fire evacuation scenario and record individuals' movement data during the evacuation process.Additionally, we employ a post-experience questionnaire to obtain more comprehensive information.

Materials and methods
The overall framework of the experiment is illustrated in Figure 1, primarily divided into three processes: recruitment and grouping, layout learning, and evacuation experiments.

Recruitment
This study recruited participants from a university campus via social media.All participants included in the study possessed normal vision, demonstrated the ability to walk or run normally, had not experienced fire accidents in the past, were familiar with virtual reality technology, and could adapt to the display and interaction of VR without discomfort.A total of 81 participants, comprising 32 males and 49 females with an average age of 21.5 years, met the requirements and enrolled.Upon completion of the experiment, each participant received a compensation of 30 CNY.The participants were divided into three groups: figural, vista, and environmental, with 27 individuals in each group.
Spatial cognitive ability has a notable impact on evacuation performance, with factors such as age and sex influencing spatial ability and wayfinding strategy (Montello et al. 1999, Hegarty et al. 2023).Abilities such as perspective-taking and mental rotation can be used to predict individual spatial navigation performance (Kozhevnikov et al. 2006).Given that the participants' spatial orientation and mental rotation abilities are the primary factors affecting the results in subsequent experiments, we have selected specific benchmarks for classification.These include observing three-dimensional scenes, using maps etc.Therefore, we have chosen the Santa Barbara Sense of Direction Scale (SBSOD), Object Perspective Test (OPT), and Mental Rotation Test (MRT) as benchmarks to classify the participants.Using multiple tests to classify participants helped this study to focus more on assessing the impact of individual spatial cognitive differences on the experimental results, while avoiding the interference of other individual differences such as gender, age, and educational background on the outcomes.
The SBSOD is a self-report test consisting of 15 statements used to test individuals' judgments of their own spatial orientation ability (Hegarty et al. 2002).Participants are required to choose a number between 1 and 7 to indicate their level of disagreement to agreement.Higher scores are considered to be correlated with better spatial orientation abilities.OPT is a test for spatial orientation developed by Kozhevnikov and Hegarty (2001).In the OPT, participants are asked to imagine themselves in the position of one image, face another image, and indicate the direction to a third image.
where, AVG i is the answer given by the participant, and D 0 j is the true angle deviation.The MRT, based on the work of Shepard and Metzler (1988), is used to assess participants' ability to mentally rotate two-dimensional or three-dimensional objects.After the test, participants were divided into groups according to their results.The guiding principle was to ensure no significant differences in SBSOD, OPT, and MRT scores among the three groups, and to minimize any intergroup differences as much as possible.Results of the grouping are presented in Table 1.
Following the aforementioned grouping method, no significant differences were observed between the groups regarding Age, SBSOD, OPT, and MRT, as shown in Supplementary Material Table 1.

Equipment
In this study, we employ eye-tracking technology within a virtual reality (VR) environment to investigate the participants' processing and comprehension of visual information (Li et al. 2023).Compared to MVR, the eye tracking technology of PC-based VR (PCVR) is more mature, so we collect learning information from participants in PCVR.Conversely, MVR enables the synchronization of participants' movements in the real environment with their positions and postures in the virtual environment, allowing the brain's vestibular system to perceive actual movements (Mestre et al. 2011, Dang et al. 2021).This not only significantly reduces motion sickness induced by PCVR but also records participants' real positions, movement speeds, and other relevant data.Consequently, the evacuation simulation experiment is conducted using MVR.PCVR uses HTC VIVE and Droolon F1 eye tracking devices, and MVR uses Pico 4.

Experimental scenario
Since the current positioning accuracy of MVR cannot meet the requirements of multilevel evacuation, we chose the first floor of a building as the virtual evacuation scenario, with an area of approximately 38 m � 40 m that was unfamiliar to all participants.It is worth noting that in some fire evacuation scenarios, evacuees need to move across multiple floors.For the purposes of the experiment and safety considerations, we only used a single floor for the evacuation experiment.Multi-level evacuations are sufficiently complex to merit separate investigation and discussion.To reduce unnecessary variables, we conducted experiments using a single-level setting.We use SketchUp software to design an evacuation scene and conduct 3D modeling (Figure 2).The real evacuation experiment scene was set up as approximately 60 m � 60 m square, and the participants wore the MVR to move without obstacles.Distinct from the wayfinding experiment, fire evacuation constitutes an emergency situation, inducing tension and psychological distress in participants.To simulate the emotions experienced during evacuation, we enhance the sense of realism through both visual and auditory means.Visually, we introduce flames and smoke at Exit 1, simulating their diffusion using the Fire Dynamics Simulator (FDS) software.To set more appropriate parameters in FDS, we simulated crowd evacuation in Unity based on the crowd movement parameters from Dang et al. (2021) and used NavMesh for navigation.The simulation results indicate that in the scenario of this study, the evacuating crowd can reach one of the two exits within 1-2 minutes from the starting point.According to the simulation results, we accelerate the FDS calculation results by increasing the diffusion speed of flames and smoke fivefold, selecting the simulated intermediate period for visualization.We set the heat release rate at 2500 kW, and the smoke exhaust rate was 2 m/s.We enhance participants' sense of tension by rendering smoke, flames, and lights, which contributes to the validity of experimental results (Figure 2).
At the same time, considering the existence of the herding effect, that is, participants following other individuals, we also set up other agent evacuees controlled by programs in the scenario.These evacuees had their own fixed routes and moved at a speed of 0.5 m/s −1 m/s in the evacuation scenario.A total of 60 agent evacuees were set up.Of these, 30 people initially moved to the wrong exit (exit 1) but redirected to the correct exit (exit 2) upon arrival.The remaining 30 people proceeded directly to exit 2 after the start of the evacuation.To simulate auditory conditions, we added a loud alarm sound to the scenario.The participants would hear the alarm during the entire evacuation process.

Experimental design
The three groups of spatial layout learning tasks were all conducted in a PCVR environment.The figural scale represents a relatively small space compared to the body, allowing individuals to easily observe the entire environment from a specific angle.To create a realistic context, we designed a virtual evacuation map (38 cm � 40 cm, 1:100) featuring common objects from daily life as spatial layout learning materials for the figural group.In the VR environment, relevant information such as room names and exits are clearly marked.Participants can rotate the virtual map using a handle, allowing them to view it from multiple angles.The learning of the Vista scale space layout presented the participants with a 3D model as large as the real environment, but the participants were required to stand at a fixed point (both in the real and virtual environments) to browse the scenario and could make slight angle changes but could not move.We set 11 points for the participants to observe, which covered the entire evacuation scenario and were arranged in order (Figure 3).Participants can discreetly and instantaneously control the navigation to the next or previous observation point through the handle buttons.Participants in the Vista group traverse all observation points at least once.
In the environmental scale space layout learning, the participants could use the handle to control the movement of the camera in the virtual scene (the body remains stationary in the real environment).The helmet-mounted display direction determined the movement direction, and the movement speed was 1 m/s.The participants could move freely in the virtual scene but could only walk on the ground.The 3D scene used for learning was the same as that for the vista group.
Each of the three groups was required to complete the spatial layout learning within a 5-minute time frame.Given that sketch drawing is regarded as an effective method for assessing cognitive mapping (Billinghurst and Weghorst 1995), participants were instructed to draw room layouts and exit locations upon completion of the learning phase.The scoring criteria for the sketch drawing are presented in Table 2.Although sketch drawing may influence participants' memory and potentially introduce deviations, this study accounts for such effects by having all participants engage in sketching and analyzing the relative scores among the three groups, thereby mitigating any impact on the study's conclusions.To better align with realworld conditions, participants were asked to partake in the fire evacuation experiment on the day following the learning phase, which facilitated the utilization of long-term memory to simulate evacuation.For participants unable to provide an answer due to complete forgetfulness of the room's location and size, their answers will be replaced with the largest error in that group.During recruitment, participants were informed that the experiment was safe.This could potentially reduce the ecological validity of the study due to participants' psychological expectations.To further enhance the sense of realism and unpredictability, participants were not made aware of their specific tasks.They were instructed to move according to prompts within the virtual environment and to act based on inscenario instructions.Upon reaching a predetermined position, the MVR system unexpectedly announced a fire occurrence within the scenario, and participants were required to evacuate to Exit 2.
Two exits were established in the evacuation scenario.Due to the diverse movement routes of the participants in the scenario, we have summarized these possible routes.From the starting point to the exit, there are a total of four paths that can be reached, including the optimal route (shortest distance) and the worst route (longer distance, facing the flames).Route 1 represented the optimal choice, as it was not only the shortest distance but also not directly exposed to the flames.Route 4, which transitioned from Route 2 to Route 3, was considered the least favorable choice.Participants following this route not only traversed the longest distance but also directly faced the fire.Route 3, although of considerable length, also faced the flames.Route 2, although longer than Route 1, did not directly face the fire, making it a slightly better option.Each route is shown in Figure 4.
During the evacuation, we enabled the video recording function of the MVR.The video recording was saved on the MVR equipment.After the experiment, these videos were exported to a computer for analysis.After the evacuation experiment, the participants were asked to complete a questionnaire about the experiment.

Data quality check
A total of 81 participants took part in the experiment, each completing the SBSOD, OPT, MRT, and VR evacuation experiments.The data from all participants were deemed valid and were included in the analysis.During data preprocessing, values more than two standard deviations from the mean were identified and removed as  outliers.We utilized one-way analysis of variance (ANOVA) and the Mann-Whitney U test to ascertain whether there were significant differences between the groups.Additionally, in the OPT, the angle data for four participants were found to be anomalous due to confusion between clockwise and counterclockwise directions.We adjusted these anomalous angle data by subtracting the values from 360 � and used the corrected values in the final results.

Data analysis
To comprehensively examine spatial layout learning, evacuation behavior differences, and outcomes among groups during fire evacuation, we developed an analytical framework and used VR eye-tracking equipment for a 5-min observation.Eye tracking metrics include saccades, rapid movements about three times per second, and fixations, which are usually before and after saccades and typically last 200-300 ms, varying by the observed object, along with microsaccades, smooth pursuit, and vergence (Antes et al. 1985).Objects with longer eye gaze may signify areas of interest or complexity (Kardan andConati 2012, Gibaldi et al. 2017), reflecting scenario understanding.
For 3D scenes, we recorded observation time in milliseconds, quantifying visual search efficiency with average saccade duration (ASD) and amplitude (ASA); longer ASD means more search time, and larger ASA means a wider search range.
In evaluating evacuation behavior, we recorded participants' position and posture at a 50 Hz frequency and the field of vision center for analysis during evacuation.Due to the lack of mature eye-tracking hardware and software for MVR, some studies have used the virtual scene's camera center point as a gaze point substitute (Dang et al. 2023).We determined focus areas and visual search efficiency similarly, calculating the region of interest by cumulatively measuring gaze duration, modifying the point into the shape of a frustum (a truncated cone), treating all objects within as observed, and specifying a 30 � viewing cone angle (Figure 5).We take the intersection of view cones at two times and, and the area of this intersection is considered the area of interest.The interval between the 2 determines the accuracy of the recognition of the object of interest.If the interval is too small, each object in the scene becomes the object of interest, which is no different from the fixation point.However, if the interval is too large, it is difficult to identify the object of interest.Through multiple tests, we sampled at a frequency of 10 Hz in this study, and the adjacent two moments were 100 ms.To show the efficiency of the participants' information collection in the evacuation process, we used the difference between the movement direction and the observation direction of the participants at two times to show the changes in the participants' cameras at the X, Y, and Z angles (Figure 6).
where P 1 , P 2 , and P 3 are the positions of the participants at three times, and there is an included angle between the gaze direction and the movement direction at each time a 1 , a 2 : a 0 1 is the angle change between two moments.a 0 1 , a 0 2 , . . ., a 0 n are summed to obtain the overall angle change a 0 , and the visual search efficiency is E a , where E a ¼ a 0 =T: E a represents the visual search efficiency in the X direction, E b represents the visual search efficiency in the Y direction, and E c represents the visual search efficiency in the Z direction.
For evacuation outcomes, we evaluated participants' movement distance and evacuation strategies.Shorter evacuation distances indicate a higher probability of successful escape.Participants moved in a plane, so we recorded the Euclidean distance in the X and Y directions.Evacuation strategies reflect individuals' behavior under different spatial cognitive levels, thereby constituting a crucial metric in fire evacuation.Since objective data alone cannot accurately indicate participants' decisions, we administered a post-evacuation questionnaire to capture the timing and nature of the strategies employed (Supplementary Material Table 2).To assess the realism of our evacuation scenarios and experiments, participants also completed an evacuation experience questionnaire (Supplementary Material Table 3).

Visual attention in spatial layout learning
Given that the vista and environmental groups utilized the same learning objects, the results of their analyses were comparable.The virtual 3D scene mainly includes building structures (such as walls, floors, ceilings, and doors), text messages (such as doorplates, evacuation signs, and advertisements), and facilities (such as tables, chairs, vending machines, and potted plants).Figure 7 shows the duration for which the vista and environmental groups observed objects in the scene.
The most significant difference between the two groups, environmental and vista, lay in the observation duration of the building structure and facilities.The environmental group's observation time was 155.34 ± 19.82 s for the building structure and 47.44 ± 23.02 s for facilities, while those of the vista group were 85.96 ± 10.02 s and 95.98 ± 18.63 s, respectively.These differences were highly significant (F (1, 52) ¼ 263.62, 69.58, both with p < 0.01).The environmental group observed the floor 2.66 times longer, moving within the scene for a cognitive map and annotating walls or floors more frequently.The focus of the vista group on objects standing out led to the differences in observation duration.
Regarding other objects like tables, chairs, potted plants, and cabinets, significant differences in observation time were noted (F (1, 52) ¼ 93.78, p < 0.01), with the vista group spending more time on unique physical structures and textures, helping them connect observation routes (Liu et al. 2021).Differences were also observed in understanding doorplates, evacuation signs, and advertisements (environmental group: 97.21 ± 17.90 s; vista group: 119.07 ± 21.06 s), though not markedly different compared to building structures and facilities.The fixed perspective of the vista group led to fewer observed posters, and the wider visual search range of the environmental group (ASA: 14.39 ms and ASD: 41.33 ms) compared to the vista group (ASA: 7.47 ms and ASD: 69.82 ms) was expected due to attention-worthy objects' variation.The vista group observed landmark objects more often.
The figural group's performance was analyzed using eye-tracking density map (Figure 8a), with the plane divided into room, doorplate, and route.The observation times were 128.44 ± 31.12 s, 97.96 ± 20.44 s, and 73.59 ± 14.82 s, respectively.The combined observation time for the room and doorplate constituted 75.46% of the total time, suggesting that participants could learn the spatial layout by observing the room's size, relative position, and name.After completing the study, the participants were asked to draw a sketch (Figure 8b).
The learning materials of the figural group were more in line with the requirements of sketch drawing, and the figural group performed better on the three projects.In the room name project, the distribution of room names is very regular, which can be seen intuitively through the plan, so the figural group presented well.It was very difficult for the vista group to establish the relative position of each room and estimate the room size.In three of the eight rooms, significant differences were observed between the vista group and the other two groups, F (2, 78) ¼ 36.52,p < 0.01.There was no significant difference between the figural group and the environmental group in the area estimation of some rooms.Even in the area estimation of Classroom 1 and Classroom 3, the environmental group was more accurate than the figural group.

Visual attention during evacuation
The visualized result of the center of a participant's visual cone is shown in Figure 9.
We divided the scene into four regions: Region 1 is diagonally opposite to the incorrect exit, Region 2 as the correct exit, Region 3 as the starting point, and Region 4 as the incorrect exit.Figure 9 reveals that the figural group was more active in Region 2 and Region 4, accounting for 77.41% of all fixation points (Vista: 62.29%, Environmental: 57.92%).Particularly in Region 4, where they more often considered Exit 1 correct, resulting in more fixation points.This may be related to a symmetry error that occurred during the mental rotation of the map.The fixation point count in Region 4 was 14,255 for the figural group, 9327 for the vista group, and 8603 for the environmental group, indicating the figural group spent more time in this region.The visual results of the vista and environmental groups were similar but notably different in Region 2. The vista group tended to choose the right side of Office 1 to reach the exit, while the environmental group preferred the left side, selecting a shorter path.The cause of this difference warrants further investigation.Similarly, the statistics on the visual attention toward various objects in the scene for the three groups are shown in Figure 10.
As evacuation duration varied among participants, we analyzed the proportion of time spent observing specific objects to the total time.During learning, the vista group observed the building structure (wall, floor, ceiling) for 0.286 ± 0.033 of the time, while during evacuation, this increased to 0.744 ± 0.08.A significant difference was observed between the learning and evacuation phases, with F (1, 52) ¼ 81.27 and p < 0.01.For the environmental group, the proportions were 0.517 ± 0.066 during learning and 0.817 ± 0.099 during evacuation.A significant difference was observed between the learning and evacuation phases, with F (1, 52) ¼ 24.27 and p < 0.01.This indicates a shift in attention between learning and wayfinding processes.However, there was no significant difference among the groups concerning content understanding, such as doorplates, evacuation signs, and placards.The proportion of fixation time in the figural group was 0.109 ± 0.010, vista group 0.113 ± 0.012, and environmental group 0.098 ± 0.012.The vista group's fixation on readable content during learning was 0.396 ± 0.07.A significant difference was observed between the learning and evacuation phases, with F (1, 52) ¼ 9.52 and p < 0.05.The environmental group's fixation proportion during learning was 0.322 ± 0.059.For facilities, the vista group's learning time proportion was 0.319 ± 0.062 and 0.072 ± 0.017 during evacuation; in the environmental group, it was 0.158 ± 0.076 and 0.028 ± 0.069 during evacuation.
In terms of visual search, although the vista group and environmental group had the same learning and evacuation scenes, their visual search indicators were not comparable due to differences in the equipment.Therefore, we compared visual search only during evacuation.The differences between the three groups are shown in Figure 11.
The overall angle changes for the figural, vista, and environmental groups were 119.22 ± 21.88 rad, 85.84 ± 17.48 rad, and 57.34 ± 9.72 rad, respectively.A significant difference was observed between the figural and environmental groups, with F (1, 52) ¼ 41.76 and p < 0.01.This was due to search duration, not visual search efficiency.Contrary to learning, a significant difference in visual search between the vista and environmental groups was observed during evacuation, with F (1, 52) ¼ 26.47 and p < 0.01.However, search efficiency was similar among all groups at 1.27 ± 0.10 rad/s, 1.35 ± 0.14 rad/s, and 1.25 ± 0.068 rad/s.This was because the goals differed: during learning, participants aimed to absorb environmental information, while during evacuation, the focus was on escaping quickly, leading to consistent visual search behavior.

Evacuation strategy and evacuation distance
The participants' evacuation strategies directly affected the evacuation distance.In terms of moving distance, the final evacuation route of each group is shown in Figure 12.
The figural group had more paths than the other groups, particularly near the wrong exit.Meanwhile, the environmental group favored Route 1 to reach the exit, consistent with Figure 9.They used Routes 2 and 3 less, indicating a clearer destination and shorter travel distance.Figure 13 displays the travel distance and speed of all three groups.
The figural group had the longest average moving distance (108.57± 57.81 m), followed by the vista group (80.59 ± 42.17 m) and the environmental group (67.83 ± 25.74 m), with a significant difference between the figural and environmental groups (F (1, 52) ¼ 4.56, p ¼ 0.047).Moving distance correlated with moving duration, resulting in similar moving times across groups.Surprisingly, the figural group's moving speed (1.22 ± 0.23 m/s) was significantly different from the environmental  group's (1.47 ± 0.26 m/s), with F (1, 52) ¼ 6.43 and p < 0.05.Three groups' moving speeds increased, with small differences in standard deviations (SDs), indicating group differences rather than individual factors or random errors.Video interviews and route analysis revealed that 36% of environmental group participants chose Route 1, compared to only 11% in the figural group.The incorrect path selection increased moving distance.The figural group's probability of staying put was 332% higher than the environmental group's, contributing to the significant difference in moving speed.
In terms of evacuation strategy, since all participants searched for exits according to their own spatial layout knowledge at the beginning of evacuation, we excluded these invalid data and adopted only the option after the first change in evacuation strategy.The results showed that the figural group changed the evacuation strategy 3.18 ± 0.96 times, the vista group changed it 1.8 ± 0.82 times, and the environmental group changed it 1.3 ± 0.77 times.Significant differences were observed between the figural group and the other two groups (vista: F (1, 52) ¼ 93.24, p < 0.01; environmental: F (1, 52) ¼ 269.35, p < 0.01).The most ideal evacuation behavior is to choose the shortest path to reach the exit according to the spatial layout memory, although only 7% of the figural group did so; however, more than 55% of the environmental group found their way from the beginning to the end based on spatial layout memory without referring to other elements.The vista group's performance was between those of the other two groups, being better than that of the figural group.Twenty-six of the vista group completed the wayfinding completely based on spatial layout memory.After getting lost, each group followed others or moved according to the evacuation sign, but these two behaviors took more time.There was no significant difference among the three groups.

Evacuation experience
The evacuation experience helps reflect whether the experiment was reliable and whether the analysis results were ecologically valid (Schmuckler 2001).We tested the reliability and validity of the participant questionnaire, and the results are shown in Table 3. Table 4 shows the results of the evacuation experience questionnaire.In Q1, no significant differences were observed among the groups.Most participants felt nervous and exhibited a desire to escape quickly, potentially due to the presence of flashing lights and alarms.Q2 revealed no significant differences among the groups, and the average score of 2.99 suggested that the scenario was perceived as authentic and resembled a real evacuation.In Q3, despite no significant differences, the average score was 1.29, indicating MVR's advantage over PCVR in terms of reduced dizziness and better synchronization of visual and body movements for spatial cognition.Q4 revealed significant differences between the environmental group and the other two groups, with the environmental group adapting faster during evacuation due to consistent learning materials and forms.The vista group faced challenges in corresponding memory landmarks with the evacuation scenario, while the figural group struggled to match the learning materials' layout with the evacuation scenario, resulting in longer evacuation times and increased difficulty.

Space layout learning and evacuation performance
The spatial scale is a key concept in both geography and psychology, influencing learning effects of spatial layouts at different scales (Brassel and Weibel 1988).During the learning stage, the figural group performed best in sketching, but this did not guarantee better evacuation performance.Maps provide intuitive spatial layouts, but large-scale learning loses details, hindering a quick connection between cognitive maps and real space during evacuation.Landmarks are crucial for cognitive map building and wayfinding, while small-scale learning demands greater mental rotation ability (Wraga et al. 2000).The vista and environmental groups observed more details, requiring observation, coding, memory, and connection to learn spatial layouts.Moving in  space aids spatial knowledge acquisition and cognitive map building (Wiener et al. 2009, Weisberg and Newcombe 2016, Li et al. 2021).
During evacuation, people quickly locate memorized landmarks for positioning and orientation.The environmental group's continuous spatial movement allowed for more complete cognitive maps, while the vista group relied on discrete observation points and landmarks to connect local cognitive maps, limiting the number of connected maps.The questionnaire showed that the vista group's early evacuation performance was equivalent to that of the environmental group.However, 55% of participants changed their wayfinding strategy after two observation points, and the vista group struggled to navigate correctly.In the figural group, the first change in wayfinding strategy location was more dispersed, not concentrated near specific observation points.The vista group, with local cognitive maps, attempted to connect discrete spatial layouts during evacuation, requiring simultaneous spatial layout learning and wayfinding.
The user experience questionnaire highlighted their difficulties, as the vista group differed from the environmental group only in their restricted movement and higher visual search frequency during evacuation, suggesting continuous cognitive map improvement.The environmental group had the lowest visual search frequency, reflecting their clear knowledge of the destination.Although the figural group observed the overall situation from a single perspective, applying this knowledge in practice was challenging.The environmental group followed this approach, gaining equal or greater spatial knowledge than the figural group, indicating that direct experience benefits evacuation.Though unable to move freely, the vista group used anchor point theory (Evans 1980) by locating unique objects and obtaining overall environmental information through multiple perspectives.Thanks to memorizing more details, the vista and environmental groups achieved better evacuation outcomes, demonstrating the importance of detail for positioning and orientation during fire evacuation, particularly when smoke obscures vision and reduces available anchor points.

Evacuation strategy
Interestingly, a notable proportion of participants across the three groups (figural: 25.92%, vista: 32%, environmental: 22.22%) selected Route 2 as the initial moving path, despite it being seemingly the worst choice given it leads away from Exit 2. While this was understandable for the vista group, given their difficulty in establishing a cognitive map, it was perplexing for the figural and environmental groups, which were more likely to perceive the exit direction.The participants who chose Route 2 were divided into two groups: group A (those who turned left) and group B (those who turned right).Analysis of the MRT scores showed no significant difference between groups A and B in the figural group, although group B's MRT score (1.82 ± 1.07) was 16.7% higher than group A's (1.56 ± 0.64).A one-way ANOVA on SBSOD scores revealed no difference between the groups.However, Pearson correlation analysis in the figural group indicated a negative correlation between group A and preference for using maps (R ¼ −0.603, p < 0.05), reflecting differences in map navigation abilities.This may be related to mental rotation ability, unmeasurable by sketching tasks.No significant difference in SBSOD or MRT between the two groups might be due to the sample size, suggesting that increasing the samples might uncover significant differences.
We divided the environmental group into two: group A (those who turned left) and group B (those who turned right), conducting a one-way ANOVA on SBSOD, MRT, and OPT scores.The results showed no significant difference in MRT between the groups, but significant differences were found in OPT (F (1, 25) ¼ 7.74, p < 0.05) and specific items like difficulty in understanding directions (F (1, 25) ¼ 12.38, p < 0.05), and ability at reading maps (F (1, 25) ¼ 24.96, p < 0.01).The OPT score of group A was 346.76 ± 142.82, and the SBSOD score of group B was 145.28 ± 83.35.Additionally, significant differences were observed in perceiving a 'mental map' of the environmental group (F (1, 25) ¼ 10.728, p < 0.05), aligning with our hypothesis.These results indicate difficulties in group A regarding direction determination, consistent with OPT performance, but the result on map reading was puzzling as it did not correlate with orientation skills.Dong et al. (2022) emphasized measures like direction and distance estimation for spatial knowledge, but in this study, there was no apparent connection between spatial knowledge acquisition and evacuation results.Traditional methods were found inapplicable to wayfinding during fire evacuation.The relationship between spatial capacity and evacuation is nonlinear.Wolbers and Hegarty (2010) asserted that individual differences in acquiring spatial knowledge are significant, whether through direct experience, virtual environments, or maps, and these differences in spatial scales also impact behaviors during fire evacuation.

Conclusion and future work
This study explored the relationship between spatial layout learning at different scales and fire evacuation behavior.The results demonstrate that the learning effects of spatial layout exert a significant influence on individual evacuation behavior.Specifically, learning within a smaller spatial scale facilitates the acquisition of more nuanced details.Conversely, in the context of large-scale spatial learning, individuals may more readily grasp a holistic understanding of spatial knowledge, yet encounter challenges in translating this comprehension into actionable evacuation knowledge.Furthermore, the study reveals a correlation between enhanced spatial ability and more rational evacuation behavior, although this relationship is characterized by nonlinearity.Such a finding suggests that proficiency in certain wayfinding tasks does not necessarily translate into equivalent performance during fire evacuation scenarios.The intricate relationship between spatial knowledge acquisition and evacuation effectiveness necessitates the development of more sophisticated models and methodologies for comprehensive explanation and prediction.Our research contributes to the field by: (1) comprehensively comparing the impact of different scale spatial learning materials on fire evacuation outcomes using advanced simulation methods, providing valuable data and evidence; and (2) investigating the relationship between spatial cognition, spatial ability, and fire evacuation behavior, revealing the process from spatial cognition to evacuation behavior, which is influenced by individual orientation and habit preference.
This study presents several limitations.First, as an empirical investigation, the reliability of the conclusions largely depends on the experimental design and the conditions of the participants.In this study, the classification of participants is based on three spatial ability tests (SBSOD, OPT, MRT), and there are other reliable testing methods that could be employed.A more comprehensive spatial testing approach would contribute to a more rational grouping of participants.Second, our recruitment primarily focused on current students.Although these students come from various regions and utilized spatial ability tests as a benchmark, there still exist issues such as similarities in participants' ages.A more diversified sample would aid in the generalization of the conclusions drawn in this study.Factors such as age, educational background, and health status could be considered, and the methods and procedures of this study are equally applicable to a broader sample.
Jianlin Wu obtained his B.S. from Lanzhou Jiaotong University in 2020.He is currently a doctoral student at Southwest Jiaotong University, researching 3D GIS and holographic visualization.He provided key feedback and revision suggestions during the paper writing process, enhancing the logical flow and readability of the paper.
Weilian Li obtained his Ph.D. from Southwest Jiaotong University in 2021.He is currently conducting research on 3D GIS and spatial cognition at the University of Bonn in Germany and in Shenzhen.He provided valuable field investigation data and observations during the research process, offering empirical support for the study.
Ya Hu obtained his M.S. degree from Southwest Jiaotong University in 2005 and his Ph.D. from the same university in 2019.His research focuses on geographic visualization.He played a role in project management throughout the research process, ensuring the progress and quality of the research, and continuously provided encouragement and support to the team.
Jigang You obtained his B.S. degree from Southwest Jiaotong University in 2019.He is currently pursuing his Ph.D. at the same university, with a research focus on geographic augmented visualization.He actively collaborated in the data collection phase, utilizing her expertise in fieldwork to gather high-quality data.Her insights into the research topic also enriched the discussions and contributed to refining the research questions.
The participants are required to indicate the direction by drawing arrows.The scores of the participants are calculated by Equation (1).

Figure 1 .
Figure 1.Overall framework of the experiment.

Figure 2 .
Figure 2. Overview of evacuation scenario and setting of fire related objects.

Figure 3 .
Figure 3. Learning materials for figural group and vista group.

Figure 5 .
Figure 5.The measurement method of MVR fixation point.

Figure 6 .
Figure 6.Calculation method of angle change.

Figure 7 .
Figure 7. Duration of observation of objects in the scene by the vista and environmental groups.

Figure 8 .
Figure 8.(a) Eye-tracking density map of the figural group and (b) sketch results.

Figure 9 .
Figure 9. Visual cone center points of each group in different zones.

Figure 10 .
Figure 10.The proportion of visual attention for different objects.

Figure 11 .
Figure 11.Visual search efficiency of each group.

Figure 12 .
Figure 12.Evacuation routes for each group.

Figure 13 .
Figure 13.The moving distance, time, and speed of each group during evacuation.

Table 1 .
Mean and standard deviation.

Table 2 .
Sketch drawing evaluation.Draw the Euclidean distance of the deviation between the room center point and the real center point.Room sizeSketch the absolute value of the room size deviation from the real room size.Room nameWhether the room is named correctly, represented by the correct rate.

Table 3 .
Reliability and validity test results.