1 Introduction

Informal settlements and settlements of displaced communities (e.g., humanitarian settlements), provide crucial shelter for people, including those negatively impacted by natural hazards and human-caused crises, such as wildfires, floods, tsunamis, earthquakes, hurricanes, conflicts, and industrial accidents [1]. Informal settlements are defined as: “residential areas where inhabitants lack secure tenure, access to basic services, where housing may not comply with planning and building regulations, and which are often situated in geographically and environmentally hazardous areas” [2]. Apart from the reduced facilities and construction standards in these settlement types, it is often the case that the construction materials offer less protection to emergency situations and such settlements have access to fewer fire protection measures than might be the case in traditional settlements. In addition, because of the displacement, there might be an increased tendency for the misuse of electricity and the use of open flames for lighting, cooking, and heating (likely out of necessity). As a result, informal and humanitarian settlements are prone to fire as secondary incidents (e.g., the 2019 fire in Al-Hol camp in Syria, the 2020 fire in Lesvos, Greece, and the 2021 fires in Cox’s Bazar, Bangladesh and Makassar, Indonesia). Other factors, such as the congested dense nature of these settlements, limited resources, training, and access for firefighting, and inadequate emergency planning, contribute to the difficulty in conducting an efficient and safe evacuation in an emergency.

Reports of major fires, some with multiple fatalities and the displacement of thousands of people, are common in informal and humanitarian settings [3,4,5]. However, without a global fire reporting system, it is difficult to calculate the total cost of these fires to residents, the environment, budgets, and programming—making it difficult to make a case for budget allocation in areas already affected by disaster. Three examples indicate how significantly fire can undermine humanitarian activities. Between January and April 2020, 15 fire incidents which affected 15,000 people broke out in Borno state, Nigeria. One of these fires killed 15 refugees, injured 15 others, destroyed 1250 shelters and several community buildings, and displaced 8000 people [3]. In September 2020, a fire at Moria Reception and Identification Center in Lesvos, Greece displaced almost 13,000 refugees—who were already displaced [4]. In March 2021, a fire in Cox’s Bazar killed at least 11 people, injured another 560 and displaced over 45,000 people [5].

There is evidence of deaths occurring during these informal settlement fires due to people becoming trapped and other issues affecting evacuation. For example, several fatalities from the March fire in Cox’s Bazar have been attributed to security fencing trapping refugees [6]; i.e., that attempt to address security concerns affected safety provisions. In response to these incidents, humanitarian practitioners and agencies are increasingly recognizing the importance of addressing fire hazards in humanitarian settlements, as well as the lack of contextually appropriate, institutionalized fire safety provisions.

Drawing on studies of human responses during informal settlement fires in Costa Rica and South Africa, as well as expert-led fire risk assessments of humanitarian settlements in Kenya, Thailand, South Sudan, Bangladesh, and Lebanon [7,8,9,10], this work evaluates the feasibility and value of applying pedestrian, agent-based fire evacuation modelling to study evacuation from humanitarian settlements. To the authors’ knowledge, few studies have been conducted applying evacuation models to understand the consequences of humanitarian settlement fires due to the challenge in collecting necessary data and building such a large and complex simulation model [11]. This work is a first step in exploring the relationship between the physical environment and people’s movement during fire emergencies in these settings. This work involves the use of an existing model to demonstrate the potential of such a tool in exploring the conditions that might be produced in such emergencies and the factors underlying them. The ultimate goal is to obtain a better understanding of evacuation dynamics through the use of viable tools and to identify opportunities to improve settlement design and emergency planning through a case study approach. The application of Pathfinder evacuation software to one of the camps inside Kutupalong refugee camp in Cox’s Bazar, Bangladesh, is used to demonstrate the potential of such models and provide a foundation for more formal and rigorous analysis.

This study does not aim to replicate an actual fire event or to generate actionable evacuation times. Instead, it aims to evaluate the relevance of existing fire engineering tools and methods to humanitarian contexts and begin to explore how the environment of humanitarian settlements may restrict evacuation. This is an initial study aiming to highlight the key challenges and opportunities with such an approach and to demonstrate opportunities for further research that may support humanitarian aid in such situations. Of course, many other locations might reasonably have been selected, other scenarios considered or other tools applied. In an engineering application, one such tool might be applied to a range of scenarios which is explored using a Monte Carlo approach—to investigate specific sensitivities and drill down into the consequences of the underlying dynamics. However, it is premature to do such formal analysis. Prior to this, the potential of applying such tools to humanitarian settlements needs to be demonstrated—allowing interested parties to better understand the insights provided and also demand the formal outline of a method that has been subject to rigorous testing. As such, this work is a first step. Conducting a full Monte Carlo approach at this stage might actually mask the underlying limitations of the modelling approach– clouding the evacuation processes captured and the narrative produced. This more qualitative approach provides a basis for subsequent quantitative assessment.

This article provides a new contribution to the field as we develop a method to build the physical environment of a large humanitarian settlement and explore the possibility of using simulation models to study evacuation outcomes in settlement fires. This paper shares insights from relevant literature on evacuation processes and related behavior during informal and humanitarian settlements fires, wildfires, and other incident types. The methods and findings from the application of Pathfinder to the case study settlement are discussed. Limitations and potential next steps are discussed to contextualize the work completed and to set out research questions to be explored in the future, as well as the practical limitations and questions regarding the application of evacuation tools and methods to inform humanitarian action.

2 Literature Review

The evacuation process (at the individual level) is often simplified during engineering analysis and can be divided into three stages: (1) the time between awareness that something might be wrong until a decision on whether or not to evacuate is made (pre-decision), (2) the time between the evacuation decision and when purposive evacuation movement to safety begins (protective action), and (3) the time period when movement to safety begins and ends (travel) [12]. Stages 1 and 2 are often combined and referred to as the pre-travel phase. Different actions are often performed during each stage, including information seeking during Stage 1; gathering personal items, warning and helping others, and firefighting during Stage 2; and movement, assisting others, resting, and changing routes can occur during Stage 3. Various individual, social and environmental factors can influence the timing associated with these stages (e.g., shortening or lengthening the time spent in a particular stage). Factors that lengthen the time spent in each stage can, in turn, increase the likelihood that individuals are exposed to harmful conditions from the fire and the length of this exposure, leading to injuries or deaths.

With this understanding of the evacuation process in mind, we examined the literature of human behavior and evacuation from different types of incidents, including settlement fires, building fires, and larger outdoor wildfire events. Studies of behavior during these types of events can provide insights on the relevant behaviors to simulate in our evacuation case study.

2.1 Fires and Evacuations from Informal and Humanitarian Settlements

Over the past five years, there has been an emerging field of research into fires in informal settlements. While fire in informal settlements is not a new problem, resources to support research in this area have only become available in recent years. Most studies focus on relatively few locations, primarily Cape Town, South Africa and Nairobi, Kenya. Here, researchers have sought to understand fire risk, behavior, practices and interventions at the household and settlement levels. Additionally, there is growing recognition that these fires are not just technical and physical challenges but have complex social, political, and economic dimensions, and responses need to take these considerations into account [13]. This reflects comparable complexity seen in more traditional building fires—although the emphasis on certain actions varies.

However, in comparison with fire behavior and risk studies, much less is known about human responses during actual fire incidents within settlement contexts. While an understanding of residents’ behaviors and decisions during a fire is critical to develop fire safety measures, there is little attention in the literature to human behavior during fire incidents in informal settlements [14]. Two studies have investigated the types of actions residents were likely to perform during fire incidents, one in Costa Rica [10] and the other in South Africa [15], based on acquired video footage of each event. This might be due to the challenges in capturing such data, but also to the prioritization of addressing other (very genuine) concerns regarding the physical environment and their resilience to fire events.

Guevara Arce et al. [10] reviewed 70 min of video footage recorded by the Costa Rican Fire Corps during their response to a 2019 fire in El Pochote, an informal settlement located in San Jose, Costa Rica. The video showed human behavior of residents and firefighters, including fire response activities. While the study found that human behavior responses were highly coupled with the local context, the authors observed behaviors similar to those seen in other fire incidents, including information seeking, non-evacuation behavior (i.e., children playing), preparedness behavior (i.e., gathering personal items), warning or instructing other evacuees, re-entry into the fire affected area, instructing emergency officials, helping others, and firefighting behavior. Quiroz et al. [15] analyzed fire spread, human behavior and firefighter’s response and operations during a fire incident in South Africa by reviewing 33 min of video footage recorded by a transit CCTV camera (entire fire incident), the Fire & Rescue Service Incident Report, and interview transcripts with firefighters who responded to the incident. The study identified some of the actions that residents undertook during the fire incident, including information seeking, firefighting attempts, valuables gathering and re-entry behavior; however, the authors stated that future work is needed to identify the factors that influenced these behaviors. In both cases, familiar actions and decisions were observed—however, their frequency and timing may have diverged from other fire incidents. This is useful as it indicates engineering tools used in other fire incidents might provide insights here.

Additionally, expert-led fire risk assessments of humanitarian settlements also offer valuable insights[16, 17]. While these reports focus on identifying fire hazards and vulnerabilities and making recommendations for fire risk reduction, they do mention evacuation challenges, including delayed or lack of warnings, nighttime fire risks, higher densities of people, lack of familiarity of some residents with the camp’s layout, and restricted or challenging evacuation routes (e.g., the use of metal fencing, longer travel distances, and/or poor road conditions). Again, such challenges exist in other residences, albeit to varying degrees.

2.2 Literature on Fire Evacuation in Building Fires and Disaster Events

Since evacuation data are lacking from settlement fires, we can investigate findings of evacuation studies of buildings and other outdoor fires (e.g., wildfires). These studies can identify if the same challenges (identified from settlement fires) also exist in other hazards, as well as the impact of these challenges on evacuation process. They also provide insights into the absence of key fire protection and communication systems from settlements and their potential impact on human performance.

First, we can look to building fires to understand the impact of delayed warnings on occupant safety. Insufficient alarm or notification systems in building fires, e.g., hotels [18, 19], office buildings [20], and residential buildings [21], have been linked to longer delays or pre-travel time periods. Without ‘official’ notification, occupants in these fires had to rely on other cues from within or outside of the building to become aware that something was wrong. Studies have shown that occupants who are located farther away from the fire event, those who are secluded or isolated from others, and those with sensory disabilities, as examples, were less likely to witness physical and social cues from the fire event, and in turn, were more likely than others to encounter delays in beginning their evacuation process (i.e., entering stage 1) [12].

Research on building fires [22] and community-scale disasters [23] have demonstrated the dangers associated with night-time hazards. In residential buildings, especially, hazards that occur at night are known to be more deadly than those that occur in daytime because the affected population is more likely to be asleep. Sleeping residents require a cue, e.g., a fire alarm [24] to wake them up first, and then once awake, often require additional time to prepare for evacuation; e.g., to dress themselves and others before evacuating. This process is further complicated if evacuees have sensory issues and/or are incapacitated due to drugs or alcohol [25].

On a similar scale to settlement fires, we can also look to studies of wildfires for insights into evacuation challenges. First, research has shown that a lack of official warnings from trusted sources (or a reliance on informal communication) is more likely to delay or inhibit evacuation [26, 27]. Additionally, people in wildfires and other hazards are more likely to ‘wait and see’ or delay their evacuation decision and movement until they witness cues from the fire, delaying their response—sometimes preventing them from evacuating safely [28]. Studies of the recent 2018 Camp fire in Paradise California, US, also highlight the links between vulnerable, densely-populated communities and the impact of blocked evacuation routes increasing exposure and the likelihood of death and injury [29]. The residents of Paradise were disproportionately older and mobility impaired, making it difficult for them to physically evacuate in time, and fire blockages of two of the main evacuation routes caused heavy congestion and delays on the other available routes [30].

Finally, research into building and wildfire emergencies has examined the impact of familiarity with the social and physical environment on evacuation outcomes. For example, wildfire research has shown that evacuees engage in ‘intermediate’ trips where they travel to other locations within the affected area (e.g., family or friend’s houses) before moving to safety [31]. This implies a familiarity with both the area and the population. Additionally, wildfires and other hazards have shown that if familiar with the location, evacuees are more likely to take evacuation routes that they typically use [32, 33]. Familiar evacuation routes may be longer in distance or take longer to travel than others; and engagement in additional and/or longer trips can increase the time spent in stage 3 (evacuation) before reaching safety. They may also not be optimal in terms of the overall performance of the community.

The research from building and wildfire evacuation studies provides support for the types of evacuation challenges likely to be present in settlement fires. Missing; however, is a better understanding of the outcomes or consequences associated with these likely trends within informal or humanitarian settlement contexts. In lieu of performing a series of evacuation drills or post-fire behavioral studies, simulation models can be used to understand the impact of these evacuation challenges and trends on overall evacuation outcomes (e.g., times, congestion points, etc.), and in turn, the overall life safety of the settlement community under various fire scenarios.

3 Methods

Evacuation modelling is used to represent evacuation scenarios from a case study community (Camp 9 within Cox’s Bazar) given assumed fire conditions. Based upon the findings from evacuation literature, simulation results will be used to analyze the impact of four relevant evacuation-related elements and their influence on performance: i.e., density (or number of residents within the camp), pre-travel delays, evacuation route choice (of familiar route), and restricted routes. This article explores the ways in which one evacuation model can be used to estimate evacuation outcomes during settlement fires and the feasibility of such methods in understanding the life safety offered by settlement designs under various fire and evacuation conditions.

In this section, an overview of the case study location is shared along with information about past fire incidents and fire safety efforts. The research methodology is discussed including the choice of using the Pathfinder software for the project.

3.1 Case Study Community—Cox’s Bazar, Bangladesh

Since August 2017, over 756,000 Rohingya fled discrimination and violence in Myanmar and crossed the border to Bangladesh [34]. As of a December 2021 report, over 900,000 Rohingya refugees reside in 34 refugee camps in Cox’s Bazar district, and nearly 19,000 Rohingya refugees reside on Basan Char, a Bangladesh island [35]. Of the Rohingya population in Cox’s Bazar, Bangladesh, across all camps, 52% are children, 44% adult (18–59 years), and 4% older adults (60 + years), with the average family size of the total registered population as 4.8 people [36]. Additionally, 53% of the population is female.

There have been at least 260 fires in the Rohingya refugee camps in Cox’s Bazar since April 2019 [37]. Larger fires have affected thousands of people, some with multiple fatalities and rapid fire spread over large areas. According to an IOM quarterly report (April 2021), there were significant increases in the number of fire incidences in March 2021 [38]. Community fire prevention and preparedness measures had begun implementation to deal with the increase in fires, including voluntary night watch committees, removal of flammable materials around shelters, the creation of household ‘go bags’, distribution of fire extinguishers and water buckets, and the burial or sale of important items [39].

One particular fire which began on 22 March 2021 in the Rohingya refugee camps affected an area of approximately 0.64 km2, killed at least eleven people, injured 560 others, and displaced over 48,300 people [40]. The fire destroyed over 1600 facilities including hospitals, distribution points, and learning centers. The fire appears to have started inside or near an individual shelter and subsequently spread laterally to adjacent dwellings and eventually to shelters and other buildings/facilities.

Evidence from the March 2021 fire offers insights to critical fire response challenges, including evacuation efforts. Of particular relevance is the fencing with barbed wires that surrounded the camps, limiting egress of the camps [41] and potentially placing lives at risk.Footnote 1 Fire risk assessors who worked in Cox’s Bazar, Bangladesh in 2019 also noted that “uneven pathway distribution inside camps makes the calculation of safe egress routes difficult” [8].

3.2 Evacuation Modeling Approach

A variety of modelling approaches can be employed to simulate evacuation performance from different settings [42]. This variation is produced through models representing key aspects of evacuation scenarios in different ways, in addition to different user interactions with the model (i.e., inputs and outputs). Models adopt different approaches in the way that they simulate or represent the physical space (e.g., via a coarse network vs. continuous spatial representation), population characteristics (e.g., individual agents vs. representing the population as a flow), population behaviors, agent/population movement, and incident development. They also differ in their input requirements and the output produced.

This work requires that evacuees are represented as individuals (to better capture detailed movement around the streetscape), can respond according to local conditions (affecting the speed of their response, the routes used and their targets) and evacuee capabilities, and are sensitive to the design of the spatial setting. A sub-set of current evacuation models can represent these factors beyond small-scale/residential properties. Several model reviews are available [42,43,44], outlining the capabilities of the models and their capacity to represent large populations (> 500 evacuees). After examining these documents and the modelling requirement, the Pathfinder simulation tool (2022 version) was selected to simulate evacuation from a location within the Rohingya refugee camps in Cox’s Bazar (Camp 9).Footnote 2

The Pathfinder model is a widely used agent-based simulation tool that provides users with the ability to estimate evacuation outcomes from a variety of building settings, including residential and office buildings, stadia, hospitals, and aircraft, among others [45, 46]. Pathfinder supports the import of geometry in a variety of formats and uses a 3D “triangulated mesh” to represent the structure, allowing for the refined representation of complex designs. The triangulation also allows for the movement of agents within the space on a continuous plane [42]. Additional features include group movement, assisted evacuation, and customizable populations—to reflect the specifics of the evacuation scenario of interest. The simulation tool also presents its output in both numerical and 3D visualization formats, allowing for quantitative analysis of evacuation performance and the animation of evacuation outcomes from scenarios of interest for visual inspection. In a simple demonstration case, Pathfinder was able to simulate an area with 50,000 people [45], more than the maximum number of people represented in this case study. It should be noted that Pathfinder is not unique in these capabilities. It was not intended to assess evacuation models for their suitability through this study. We only selected a readily available pedestrian evacuation model that meets all the specific requirements for modelling humanitarian settlement fires to develop the methodology for this study and identify some of the challenges faced. Other models that meet the same set of requirements can also be used with the methodology.

Due to the size of the settlement and the exploratory nature of this work, we focused on simulating evacuation from one section of the Rohingya refugee camps in Cox’s Bazar, Camp 9, to reduce the demand for computational power to run the simulation. This camp was chosen since it was one of the camps affected by previous fire events and contains a portion of the border fencing, which could slow movement along or even hinder usage of evacuation routes in some cases. It is not felt that representing this section of the Cox’ Bazar settlement diminishes the value of the work presented here; however, it is certainly the case that examining the entire settlement would have presented a more complex technical challenge and have been considerably more computationally expensive.

The complex process used to import the details of the camp’s layout, distribution of buildings and road networks into the Pathfinder model is discussed below. This is instructive as it differs from the normal process of creating a building geometry—given the scale and wider set of actions/technologies involved. It is expected that the approach employed will also be instructive to simulating evacuation from other outdoor environments (e.g., wildfire events) and/or reproducing or expanding the results presented in this paper.

The following section describes building the geometry and configuring the evacuation scenarios in detail—both the representation of the initial conditions and the evacuee response across the different scenarios. The application is beyond the intended application area of the Pathfinder model (e.g., building structures) and so the effort is documented to demonstrate the tasks conducted and the results produced, and provide shortcuts for future researchers and practitioners.

3.3 Building the Geometry

As mentioned in the previous section, Camp 9 of Cox’s Bazar rather than the entire refugee camp was chosen for modelling to reduce the demand for computational power to run the simulation. If we were attempting to obtain representative results of the evacuation process of Camp 9, this would have required closely mapping the actual perimeter of the camp. As we only intend to use this region to build a large-scale evacuation to demonstrate the method developed, and also considering that an evacuation of the camp may not necessarily follow the boundary of Camp 9, a simplified approach was adopted to select a rectangular area of dimensions 1070 m × 764 m that contains a large portion within and around Camp 9 to build the geometry for simulation (see Fig. 1). This was deemed sufficient to capture key evacuation dynamics, allowing us to demonstrate the method while being less computationally expensive.

Figure 1
figure 1

A map of Cox’s Bazar (courtesy of OSM) that shows the boundary of Camp 9 (in orange) (https://www.openstreetmap.org/relation/8009019) and the rectangular boundary of the area chosen for simulation (red box). The bottom left corner shows the original map scale of 300 m/1000 ft

Full size image

The creation of Building Camp 9’s geometry within Pathfinder was executed in two stages. The first stage involved importing satellite data into Blender that represented the geometry employed. Blender is a free/open-source, python-based 3D computer graphics software used for a variety of applications (ranging from GIS visualization to movie effects) [47]. Figure 2 represents the key steps required to go from model inception to running the model. It should be noted that these steps are specific to the community/settlement scale (given the need to represent the external landscape) and would not be required when looking at buildings. The steps involved (1) selecting the region of study, (2) locating the selected region in BlenderGIS (which is an add-on to Blender), (3) importing the data from the selected provider (OpenStreetMap (OSM), Google Maps, etc.), (4) creating the evacuation paths by extruding them to a specified size, (5) exporting the 3D model to Pathfinder and then (6) finalizing the model in Pathfinder by creating the road network, rooms, access doors and exits (required to represent the movement of people within this space).

Figure 2
figure 2

Process associated with building the geometry for use within Pathfinder representing the six tasks required for settlement applications

Full size image

Using BlenderGIS, satellite data was obtained via publicly available data—from Google Maps, OSM, Bing and Esri [48]. For model building in this case, OSM was chosen as it provides reliable data on roads, paths and building footprints for the area examined. Using the “base map” function, the OSM map data was exported to Blender. OSM provided the model with information on the roads and buildings within the area of interest, including geospatial data for roads and buildings and supporting road details (size, type, speed limits, etc.). The road network was imported from OSM into Blender in a “Curve” format, which had to be converted to a polygon format to be successfully imported into Pathfinder (see Fig. 3).

Figure 3
figure 3

Curve format imported from OSM onto Blender

Full size image

The second stage involved constructing the 3D environment for modelling within Pathfinder. From BlenderGIS, the model was exported as a “Colladae” (DAE) file and imported into Pathfinder. The model contained three geometry layers: the Basemap, Buildings, and Roads. The imported Colladae representation contained gaps in the connecting roads due to how the geometry was bevelled, which required manual correction within Pathfinder.Footnote 3

To refine the model, the roads were moved to another Z-axis for “room” generation to ensure the path and the land descriptions did not interfere with each other—effectively, representing where space might be occupied on the road network. The entire road layer was converted into one “room”, which represented a pseudo road networkFootnote 4 within the selected region. This process was repeated with the Basemap and buildings together, with each division of terrain surrounded by roads together with the buildings/houses within the division being defined as one “room” to accommodate the simulated population.Footnote 5 The boundary of the rooms is mainly determined according to the layout of the road network obtained from the OSM data. Therefore, the size of the rooms vary. In general, a “room” contains from a few up to tens of buildings/houses, representing a confined neighbourhood. The road layer was then set to the same Z-axis as the Basemap and building geometry, yielding a completed geometry of the selected region.

It should be noted that although the gaps in buildings could form part of the walkable space, they are generally not free space or not suitable for many people to walk through because of uneven surfaces, the presence of plants and obstacles, and difficulty in navigation. It is understood that only people living locally within the residential divisions would be able to efficiently navigate through these gaps. Therefore, it is assumed that during an evacuation, residents will emerge out of the local residential divisions and walk to their nearest laid roads and naturally formed paths, and then move towards the exits via the road network. This assumption could be changed to reflect different behavioral assumptions in additional simulations.

To allow evacuees’ travel across the geometry, access “doors” were used to connect the rooms (i.e., the residential divisions) with the road network—effectively forming paths and providing target points at the end of the paths. This required the manual addition of doors into the geometry at locations that might be traversable based on inspecting the satellite data.Footnote 6 See Fig. 4 that shows an example of laid road and a piece of road network in the same small region in the camp (c, b) shown in different platforms; where Fig. 4 shows examples of door placements within an excerpt of the larger simulated geometry.

Figure 4
figure 4

An example of laid road and a piece of road network in the same small region in the camp (c, b) shown in different platforms and (d) modelled in Pathfinder. (d) Specifically shows door placements, displayed in yellow-orange, connecting buildings to the road network lines along evacuation routes

Full size image

The primary focus of Pathfinder’s modelling capabilities targets indoor building evacuation. However, it has also been used to examine large-scale people movement at events, stadia, and outdoor crowds [44]. To support these applications and further demonstrate the credibility of its use here, additional work was performed to implicitly implement the varying terrain within the geometry; i.e. represent the impact of different real-world terrain on evacuating residents. Ideally, topography can be exported into the model via satellite data and applied to the building and floor (i.e., base) layers; however, since this is a demonstration case, we opted to implicitly represent topography by sectioning off selected areas of the camp, within which we reduced the movement speed multiplier to lower than average. In an engineering analysis, topographic information should be included to develop an accurate representation of the simulated space and these data, particularly pertaining to Bangladesh can be obtained from multiple sources including: Natural Earth [49], the United States Geological Survey [50], and the Humanitarian Data Exchange [51]. Adding expertise from the field of geospatial sciences to the engineering team when accessing and using these data is also encouraged. Note that the resolution of the OSM map is 0.149 m/pixel at zoom level 20 on the Equator.Footnote 7 The latitude of Camp 9 is about 21.18 degrees; thus, the resolution of the Camp 9 map is 0.149 * cos (21.18) = 0.139 m/pixel, which is assumed to provide a reasonable representation of the distribution of mid-sized buildings, the gaps between buildings and the road network in the geometry modelled for this study.

The simulated geometry included a total of 66 exits. In this case, ‘exit’ refers to the locations where the exit pathways cross the edge of the simulated geometry—enabling evacuees to be recorded as having reached safety by moving along a road/pathway to a place remote from the assumed incident location. Table 1 shows the distribution of exits by direction on the geometry and the size of the exits (i.e., the width of the pathway at the location of intersection with the geometry’s edge). Due to the resolution of the map data, it was necessary to estimate the widths along the paths in order to produce a representative network of pathways. In turn, the exit width of each pathway is rounded to its nearest integer value with increments of 0.5 m where necessary between 1 m and 2 m. It is important to note that while these estimates may in fact reduce the accuracy of outcomes in an engineering application, they do not detract from the findings of this demonstration case. The measurements might also be refined should an engineering application rather than a demonstration case be conducted.

Table 1 The Distribution of the Simulated Exits by Width and Location on the Geometry
Full size table

3.4 Scenario Development—Configuring the Model

Scenarios have been designed to (1) represent initial conditions and evacuee responses, and (2) examine the performance of the model in an application area beyond the norm. A total of seven scenarios have been simulated as part of this analysis. These scenarios are produced by varying five key parameters in evacuation analysis: population numbers, pre-evacuation times, evacuation routing preferences, route access and resident characteristics. These were then used to configure the Pathfinder model to simulate the scenario initial conditions and response assumptions. A summary of these scenarios and the factors examined for each evacuation scenario are outlined in Table 2.

Table 2 Seven Scenarios Simulated within Pathfinder
Full size table

Pathfinder’s Steering Mode is used in the simulations to better represent agent movement and interaction within the irregular spaces of the simulated camp area as well as possible congestion due to merging flows. This mode represents local interactions between agents requiring more refined navigation around other agents and objects. It should be noted that the default settings of Pathfinder are intended to represent a limited set of building evacuation scenarios and are not necessarily intended to simulate an evacuation from Camp 9 within Cox’s Bazar. In a more complete analysis, observations might be sought to more credibly represent settlement evacuation scenarios. However, (a) it unlikely that such data is currently available, and (b) the purpose here is primarily to demonstrate the types of insights that might be captured from model application.

Scenarios 1–3 explore the impact of varying the total resident population size on the evacuation conditions produced and the overall evacuation time. The Humanitarian Data Exchange [52] was used to generate estimates of the number of individuals located in Camp 9 (approximately 33,400). A total population of 35,000 was therefore assumed in Scenario 2, with minimum and maximum estimates assumed as ± 10,000 people (25,000 in Scenario 1 and 45,000 in Scenario 3). This population was randomly distributed throughout the occupiable spaces of the geometry; i.e., people were not initially located on roads or unoccupied land types (e.g., farmlands). This approach of randomly distributing the population was adopted for all of the scenarios examined. The populations in Scenarios 1–3 adopted the default occupant profile in Pathfinder which assigns an unimpaired movement speed of 1.19 m/s to each agent. This could easily have been varied or derived from a distribution; however, this functionality is not being examined here.

Scenarios 1–3 also assumed default inputs for route preference (i.e., choosing the nearest exit) and access (i.e., all exits were available), with a 0 s pre-travel (or delay) time to examine the sole influence of increasing the number of residents on the simulated evacuation time and maximizing the demand on the routes used. It was also assumed in Scenarios 1–3 that the evacuating population employed the default resident characteristics distribution (e.g., relating to resident size) provided in Pathfinder with a parameter selected allowing evacuees to resolve high density conditions. This selection was recommended by the model developers to avoid agents getting stuck within the simulation, given the unusual variety in route width and route complexity.Footnote 8 Overall, Scenarios 1–3 focused on examining the impact of changing the size of the population on evacuation performance given that all other elements remained constant.

Scenario 4 was included to examine the impact of introducing an initial delay time given an assumed population size of 35,000. All other scenario assumptions were left as before. The results of this scenario could therefore be directly compared against the results from Scenario 2 as the only change was the introduction of the delay time distribution. It was also assumed that the populations’ delay times followed a normal distributionFootnote 9 to represent the potential response of evacuees during a daytime fire scenario. While no pre-travel delay data exist for humanitarian or informal settlement fires, pre-travel delays are a common occurrence in both building fire and wildland fire settings, and thus was assumed to occur during a fire in Cox’s Bazar (especially as people located further away from the fire may need time to assess what is going on). A likely pre-travel delay distribution (i.e., normal distribution, mean of 15 min and standard deviation of 5 min) from the buildings’ literature was used in these scenarios [53]. In reality, this delay may be considerably longer and distributed differently—more closely approximating a larger-scale event, such as a wildfire evacuation; however, we are primarily demonstrating that such distributions can in principle be represented and that their impact understood by comparing results from different scenarios.

Scenarios 5–7 explored the impact of varying resident characteristics, access and exit preferences respectively. This allowed for direct comparison with Scenario 4—to assess the impact of individually modifying these three factors, while keeping other factors the same.

Scenario 5 assumed that 35,000 people were located with the geometry (as in Scenario 2 and 4). It also assumes the modified resident profile shown in Table 3, rather than employing the default population profile in Pathfinder. This profile was used to better represent the expected population within the geometry by accounting for walking speeds based on more representative demographics (e.g., wider age distribution than might be expected in non-residential buildings). Existing literature was used to provide broadly indicative (but simplified) pedestrian walking speeds [54] and a settlement profiling assessment report from 2019 was used to approximate the percentage of residents in our study area within particular age brackets [36].

Table 3 Resident Profile used in Scenarios 5–7
Full size table

In Scenario 6, evacuation preference was the only variable altered, when compared with Scenario 5. Since previous literature identified evacuee tendencies toward familiar routes, it was of interest to test the influence of route choice in a humanitarian settlement context [33]. It was assumed that 40% of the population used a ‘main exit’ to evacuate. This assumption was implemented by assigning the residents initially located adjacent to any paths leading to a main road (40% of the total population) to evacuate via the main roads and exits, while all other residents sought the exits closest to them. The four exits that each measured 5 m in width were identified as ‘main exits’ in this scenario: one on the northwest side of geometry, one on the east (center) side of the geometry, and two on the southern side of the geometry (one due south and the other southeast). The purpose of this scenario was to observe if congestion was more prevalent in situations where evacuees predominately used the main routes—and the impact that this might have on the overall evacuation times produced. It should be noted that this 40% could easily have been modified to represent different patterns of familiarity.

Finally, in Scenario 7, evacuation routes were restricted. Since barbed wire fencing may have hindered evacuation outcomes in previous fires within Cox’s Bazar, all the doors/exits located in the South-Eastern portion of the model’s geometry were disabled. Such a loss might also have been due to the location of a fire incident. In this case, none of the evacuees could choose these exits. Figure 5 shows the geometry of the model and the exits blocked (shown as red lines). Incremental changes were made between Scenario 5, 6 and 7 allowing the impact of the change in single parameters on the results produced. Again, in more complete analyses additional parameters would be modified. However, the purpose here is to demonstrate the potential of such limited changes to maintain control over the conditions examined.

Figure 5
figure 5

Geometry of the model and exits blocked (blocked exits are indicated by the red line along the southern and eastern edges of the camp’s geometry and any exit that crosses that red line was blocked)

Full size image

In all scenarios, the following additional assumptions were made, in that all simulated agents:

  • Evacuated on foot only,

  • Evacuated and began moving directly to an exit once their delay time had ended,

  • Were aware of at least one way out of the affected area,

  • Were without physical impairments that might have hindered evacuation movement,

  • Acted individually (i.e., not part of a group),

  • Would not change their choice of exit once the simulation began.Footnote 10

It should be noted that these assumptions could be modified to reflect other scenarios of interest. Originally, it was of additional interest to implement family units within the simulated population, in that individuals located within a family unit would move together in a group, simultaneously. The “movement group” function in Pathfinder was explored to account for families of varying sizes (e.g., from 1 to 5 people distributed throughout the geometry). Even though population distribution data from online resources were available, the implementation of family units within the Pathfinder simulations was impractical due to its impact on simulation run time. When implemented, the inclusion of ~ 4000 movement groups caused the simulation to run approximately 8 to 12 times longer than the simulations without the family group feature. Therefore, the family units feature was not used in this initial study and will be explored in future research, allowing the set of scenarios described here to be examined.

3.5 Running the Model

Once the model was configured, Scenarios 1–7 were simulated once each. Pathfinder is a deterministic model and would not have produced different outcomes as the inputs were not perturbed for each scenario. In a full engineering analysis, initial parameters would have been perturbed to generate a distribution of results for each scenario. However, here the focus was instead on exploring different scenarios than the robustness of the outcomes produced. The outcomes reported in each case were (1) overall evacuation times, (2) graphs showing number of people evacuating over time, (3) the behavioral components affecting arrival times, (4) the number of agents using specific exits and (5) how long these exit points took to clear for each of the seven scenarios. Example approaches to graphing and tabulating these results are shared.

The seven scenarios and their results are not meant to provide definitive evacuation clearance times for our selected geometry of Cox’s Bazar, but instead are meant to explore how such a tool might be used to quantify evacuation times from such a settlement, and to demonstrate how these times may change as different scenarios are examined by modifying parameters; e.g., resident delays in response, preferences for less efficient routes (e.g., main roads), restricted access/egress points, and varying resident characteristics.

4 Results

Table 4 and Fig. 6 show the results from all seven scenarios. The impact of varying population size on evacuation performance is represented by comparing the simulation results of Scenarios 1–3: as the number of residents increased from 25,000 (Scenario 1) to 45,000 (Scenario 3), the overall evacuation times increase from 25 min to almost 44 min. Given that all other parameters were kept constant, this result is likely due to the increased stress placed on the path capacity given the increase in the number of evacuees simultaneously using these paths.

Table 4 Overall Evacuation Times for Scenarios 1–7
Full size table
Figure 6
figure 6

Number of simulated residents exited over time for Scenarios 1–7

Full size image

Figure 6 indicates that in Scenarios 1–3 (where the population size was varied) over 80% of the camp evacuated between 500 and 750 s—reflected in the steep gradient as multiple paths were fully used allowing comparable arrival rates. This continued until the exits become congested and the rest of the population (20%) took longer to evacuate as they waited to flow through the exit pathways. Similar resident movement characteristics and no pre-travel delays were assumed in Scenarios 1–3; therefore, the different demand levels in Scenario 1–3 and the subsequent flow of simultaneous evacuees on the paths (producing the congested conditions) likely produced this outcome.

Additionally, the simulation results show that with a base population of 35,000 people, total evacuation (or clearance) times of the simulated geometry can range from 34.8 to 72.6 min, when comparing Scenario 2 with Scenarios 4–7 in which the underlying physical or behavioral conditions were modified. The impact of the pre-travel delay time distribution is apparent in Scenarios 4–7 in that evacuees in these four scenarios did not evacuate until after 500 s had elapsed, and most of them did not do so until after around 1000 s have elapsed—reflecting the initial delays experienced. This shows the importance of this model functionality and its impact on certain scenario conditions. These curves appear shifted along the x-axis. The curves for Scenarios 4–7 are similar in shape, with the scenario restricting access to the South-Eastern exits of the camp (Scenario 7) causing the longest evacuation times given evacuees had to travel longer distances to reach an exit and also placed more demand on these exits than was previously the case. In Scenarios 4–7, approximately 80% of camp residents evacuated between 1500 and 2000s, with the remaining 20% requiring an additional 1500 to 4500 s to clear the simulated area.

The number of evacuees using each of the exits (grouped by compass direction) is shown in Table 5. These results are influenced by both the number of evacuees present (i.e., comparing the total population sizes in Scenarios 1–3) and exit accessibility (i.e., see changes in balance of exit use in Scenario 7).

Table 5 The Numbers of People Who Used Exits in Each of the Four Directions (West, North, East and South) for all Seven Scenarios
Full size table

The impact of exit usage (and other scenario changes) on evacuation times through these exits is shown in Table 6. Scenario 6 placed an undue burden on the capacity of the main exits (in the North of the area examined) producing an overall evacuation time of over 65 min, while the loss of exits (and subsequent complexity and overloading of local exits in the East) in Scenario 7 produced evacuation times of over 72 min. This indicates the importance of diagnostically examining the underlying conditions—and the capacity of such simple modelling to provide insights into the sensitivity of the evacuation performance to specific changes in the routes used and/or behavioral response.

Table 6 Time(s) of the Last Person to Exit the Pathway Exits in Each of the Four Directions for all Seven Scenarios
Full size table

In addition to identifying the factors that underpin changes in performance, it is also possible to identify the state of the evacuating community over time. We use Scenario 4 here to demonstrate this capability (see Fig. 7), although it could be applied throughout. Figure 7 shows the proportion of the overall population engaged in particular activities—still at home having not yet responded (in blue), in the process of walking to a place of safety (in orange), and having evacuated to a place of safety (in grey). This breakdown would allow a practitioner to determine the status of the community at any point in time and might also inform intervention strategies. It is apparent that 80% of the population had yet to respond by approximately 7–8 min (perhaps warranting more communication intervention) and for the next 10 min approximately 20% were walking—i.e., on the road and exposed (perhaps warranting observation of key routes). In such a way the deployment of responders might be more effectively informed.

Figure 7
figure 7

Breakdown of population response over time for Scenario 4

Full size image

By observing the visual output, it is apparent that there was a significant decrease in congestion in Scenarios 4 and 5 compared with Scenario 2—with congestion primarily produced at three pinch points in the path network. This is mainly due to the prolonged response phase and varied travel speeds represented in Scenarios 4 and 5, although the evacuation times of Scenarios 4 and 5 are 25% and 34% longer than that in Scenario 2 respectively.

In Scenario 6, 40% of evacuees were assigned a ‘main exit’. This produced congestion along these main routes when compared with Scenario 2 where residents used their nearest path to an exit, creating a more distributed response across the path network. Figure 8 shows the differences produced by different route selection. In Fig. 8a, the response during Scenario 2 is shown—evacuees using their nearest route. The conditions from Scenario 6 are shown in Fig. 8b, c where congestion forms at an intersection resulting from some evacuees using their nearest exit while others selected the main path.

Figure 8
figure 8

Examples from selected locations within the simulated geometry that show the onset of a bottleneck

Full size image

Finally, Scenario 7 represented the impact of fencing around southern and eastern portions of the camp (see Fig. 5) resulting in the loss of routes out of the area. Scenario 7 produced the longest overall evacuation time of approximately 73 min—a 14%–109% increase in the evacuation times over the other four scenarios that also involved 35,000 residents. The additional congestion produced by this loss of egress routes is visible in Fig. 9, where evacuees who had used the lost routes in other scenarios now overloading remaining pathways.

Figure 9
figure 9

Snapshot of bottlenecking in Scenario 7

Full size image

5 Discussion and Concluding Remarks

This study demonstrated the process of building the settlement geometry (and the additional steps required beyond those typically involved in generating an occupied building) and the use of evacuation modelling to explore evacuation processes and outcomes in large outdoor humanitarian settlements during fires. Through a case study approach, the Pathfinder model was used to demonstrate the impact of four key factors on evacuation outcomes in one of the camps inside Kutupalong refugee camp in Cox’s Bazar, Bangladesh. These factors included density, pre-travel delays, evacuation route choice, and restricted routes.

The results show that Pathfinder, along with the method developed to build a large-scale geometry based on third-party GIS data, can be used to estimate evacuation outcomes during settlement fires, i.e., that it is possible for such an application to be conducted. This article outlines the process allowing others to use this method to better investigate the life safety offered by other settlement designs under various fire and evacuation conditions. Model output and analysis demonstrated that the evacuation times produced (even from this relatively simple area) were sensitive to population density, route availability, route choice, initiation times, and resident characteristics, as we would have expected from simulations of other environments (e.g., building fires). All factors tested resulted in increased evacuation times (when compared to Scenario 2); however, more importantly, clearance times significantly increased (e.g., almost doubled) in scenarios that incorporated main route usage (Scenario 6) and exit restrictions (Scenario 7). The model was sensitive to varying these key parameters. Next steps might then include a more rigorous analysis of these sensitivities to provide confidence in the veracity of the model and the value of the results produced.

These prolonged evacuation times during fires in settlement areas can produce serious consequences, especially in cases with fast moving fires. While fire was not simulated in this case study, we can look to data collected from actual and experimental fires in informal settlements to better understand potential consequences of these delays [9]. Studies of fires in informal settlements in South Africa found that the linear fire spread rates could be as high as 2.3 m/min [15, 55] and large-scale laboratory experiments found areal spread rates as high as 115 m2/min. Assuming worst case conditions and, in turn, a maximum areal spread rate, the fire could have burned an area equivalent to a soccer or US football field (i.e., ~ 8400 m2) in the time that it took residents to fully evacuate the modelled area for Scenarios 6 and 7. Fires initiating or spreading to densely populated residential areas (e.g., 10 m2/p) [56] or congested pathways (e.g., 0.19 m2/p for Level of Service F, queuing) [53] within the camp could threaten the lives of thousands of residents. Such a comparison of fire spread with evacuation time follows the ASET/RSETFootnote 11 approach commonly used in building fire safety engineering. The utility of such an approach has been suggested for bushfires planning and preparation [57] and as demonstrated in this study, can be used in large-scale settlement fires as well.

The goal here has been to demonstrate the potential for such an evacuation tool to be applied and that it might reasonably be used diagnostically to explore the impact of different scenarios on overall evacuation performance and establish the underlying factors that drive this performance. Of course, the absolute accuracy of such estimates will be enormously reliant on the data available; however, the potential has at least been demonstrated. The next step is to test the scope and refinement of the model’s functionality, its capacity to perform as expected and capture real-world conditions—through verification and validation testing. This work has shown that such tests are at least warranted. A suitably tested and configured simulation tool has the potential to aid in settlement design, mitigation efforts and in emergency planning—potentially making better use of the (possibly limited) resources available in safeguarding the settlement residents.

5.1 Limitations

The work has a number of limitations. First, it is important to recognize that the results, as with any pedestrian modelling tool, do not reflect the realities of any specific evacuation or fire incident, and are only indicative of the types of scenarios that might be of interest and of the conditions that might be produced. The goal here has been to demonstrate the potential of such a tool rather than draw detailed insights from the application described. Such a limitation does not directly affect our demonstration, but will influence a future engineering application of this tool which might be explored through detailed testing.

Given that it was larger and more complex than the buildings normally represented by Pathfinder, additional steps and assumptions were required to reproduce the landscape involved. The simulated area within Cox’s Bazar was built using BlenderGIS and then incorporated into Pathfinder, using the publicly available data from Google Maps, OpenStreetMap (OSM), Bing and Esri, as explained in the methods section of this paper. Given the incomplete maps and data available for simulated area, and the difficulties incorporating certain elements (e.g., terrain), individual building locations are approximated. Such limitations primarily relate to the overall process rather than the application of a simulation tool. For instance, the spatial representation of such systems is rapidly improving, enabling more detailed representation. As such, this limitation is diminishing—likely reducing the future assumptions needed.

Additionally, for the path network (and occupiable spaces) to be created in a functional manner, a number of user assumptions were required. The mesh developed to enable evacuee movement was larger and more complicated than normally generated by Pathfinder. In some instances, the paths produced were counter-intuitive with unexpected queues developing. To combat these issues, some of the advanced behavioral features within Pathfinder where adjusted (e.g., current room travel time, current room distance penalty, etc.). Modifying these parameters improved the outcomes; however, these changes were technical in nature rather than modelling part of the underlying scenario being examined.

This process was also prone to error, with pedestrians occasionally located in spaces from which egress paths could not be generated and where overlapping objects interrupted path generation. This was solved by modifying the objects (e.g., doors) and/or allowing the model to resolve high density situations in a more granular manner. However, this still led in some instances to residents getting stuck (i.e., a similar and small proportion in each scenario), in which case they were simply excluded from the analysis. These assumptions might evolve as geospatial representation improves and also as our expertise in model application in this area improves—reducing error through increased experience.

Assumptions were also made regarding the behavior of the agents within the simulated area. The study was prepared with the limited inputs and assumptions received from publicly available sources and personal communications with humanitarians working in Cox’s Bazar. These insights do not capture key demographic information and expected behaviors—primarily due to the lack of supporting data. One such key parameter is pre-travel times, and since data reflecting pre-travel delays from outdoor settlements are not yet available, assumptions were therefore made to estimate delay distributions. Additional assumptions were made when simulating other parameters (evacuation route preference and resident characteristics) to demonstrate that these factors could be examined, with more representative investigation left for when such data are available. Future analyses would benefit from collecting behavioral and movement data from evacuations during settlement fires to assist with model inputs and a reduction in the number of assumptions made.

Finally, it is acknowledged that we used Pathfinder outside of the conditions for which it has been verified and validated. Due to the lack of evacuation data from humanitarian or informal settlement studies, Pathfinder or any other simulation tools (to the authors’ knowledge), have not been verified or validated for these settings. However, the goal of this work was to demonstrate the potential application to an adjacent field—settlement evacuation as opposed to building evacuation. The potential was demonstrated; however, the quality and scope of such application should be explored in subsequent work.

5.2 Future Work and Next Steps

This initial application of Pathfinder for use in humanitarian settlement environments lays the foundation for additional research opportunities in evacuation and modelling studies. Pathfinder is one of a number of pedestrian modelling software that are available for use in similar settings, including Legion, STEPS, EXODUS and Mass Motion, to name a few. Future work may compare the use of additional tools to simulate the same scenarios to identify differences in the software’s capabilities and movement algorithms. Also of interest in future work is the collection of data from a specific evacuation event with a particular settlement for validation of these and other model results.

Future studies should also consider the collection of data on other evacuation challenges including affiliative behavior (e.g., collecting family members in another area of the camp before evacuating), choosing not to evacuate, and firefighting behavior, among others [33].

Additional parameters can also be incorporated within future evacuation scenarios, including considering fire conditions, the actions of first responders, evacuation from other camps within Cox’s Bazar, the impact of evacuee behavior (e.g., intermediate trips and/or firefighting), and/or evacuation outcomes from different settlements within Bangladesh and internationally.

Given the complexity and scale of the evacuation problem in humanitarian settlement environments, it is critical that evacuation models are rigorously tested to determine their capacity to estimate credible outcomes from a range of representative scenarios. It would also be worthwhile to conduct further sensitivity analysis on the factors involved by assessing the evacuation performance when varying the upper and lower limits of each variable. This would move the current demonstration of potential to the investigation into viable use.

A goal of this work is to encourage other researchers to explore evacuation topics within informal or humanitarian settlements to answer some of the many questions posed by this case study—especially to formally examine the capacity of simulation tools to be applied, create a method by which such a tool might be consistently applied and the quantify the credibility of their performance through formal testing. Too many lives have been lost already in fires in these high density and poorly fire-protected settings. Researchers should aim to apply their knowledge of building evacuations and evacuation from outdoor settings (e.g., wildfires) to fill current gaps in understanding and inform the protection of settlement residents exposed to fire risks.