In 2018, there were 315 natural disaster events recorded with 11,804 deaths, over 68 million people affected, and US$
131.7 billion in economic losses around the world. Earthquakes were the deadliest type of disaster, accounting for 45% of deaths, followed by flooding at 24% [1
]. Efforts have aimed at preventing disasters and improving resilience, and the construction of emergency shelters is an important component of these efforts [2
]. Evidence from disasters indicates that emergency shelters can reduce casualties to a certain extent in the event of a disaster [3
]. An emergency shelter is a facility where government agencies or pre-established voluntary organizations conduct assessments and provide disaster services for evacuees who do not have destinations. These facilities are able to accommodate people and provide food and water, as well as basic first aid, pet shelter (as appropriate), health support, and basic disaster services [5
]. Numerous studies have made progress in areas such as shelter planning support methods, space suitability assessment, locations optimization, emergency materials and facilities allocation, and sheltering behavior and psychology [6
]. Since the users of the shelters are citizens, a key issue during possible evacuation activity is determining the spatial distribution of potential evacuees for emergency shelter demand, which is directly related to the rationality of shelter locations and evacuation efficiency, that is, whether all the shelters in the city can provide refuge for people at a reasonable evacuation distance and time during a disaster [8
Mapping the regional distribution of the population is a basis for analyzing the shelter demand [13
]. Assessing the population distribution accurately is difficult because the population is mobile, with differences in both size and distribution at different times during day and night [15
]. Traditional geospatial data, such as land cover data, have been widely used to describe population distribution [16
]. With the development of technology and the richness of data, some scholars have incorporated big data or multi-source data as the data foundation to map fine-scale population distribution. Mossoux et al. [18
] proposed an inverse calibration method to reduce the errors in roof type classification for high-resolution remotely sensed data processing, and the application of this method to population distribution assessment showed its potential for producing systematic population maps, especially in areas where regular census data are unavailable. Ma et al. [19
] used subway smart card data to assess hourly dynamic changes in community population distribution, which can be applied in estimating the potential number of evacuees under different disaster scenarios and to support future urban planning. To produce population distribution map at the building scale, Yao et al. [20
] introduced a down-scale approach to refine spatial distribution accuracy from street level to grid level based on points of interest (POIs) and real-time social platform user densities data, and the results can guide numerous urban construction practices such as disaster prevention, policies making, and resources allocation optimization.
The estimation of shelter demand is generally obtained through a superposition analysis of potential areas affected by disasters and population distribution [7
]. Using census data, Chen et al. [8
] predicted the population density distribution of future urban planning at the street level based on the current population distribution, and then identified the areas that may be affected by the disaster according to a risk analysis. Finally, shelter demand is estimated by overlapping the results of the above two processes in a geographic information system (GIS). Sugimoto et al. [21
] selected a tsunami caused by an earthquake as a study scenario and calculated the scale and spatial distribution of possible casualties considering the maximum inundation areas combined with factors such as the time required to seek refuge after an earthquake, the inundation depth, flow rate, and the evacuation speed. Vecere et al. [22
] used ERGO-EQ and HAZUS-MH tools and set the hazard, vulnerability, and exposure of the earthquake as input parameters to estimate the number of the displaced population and shelter demand, accompanied by a comparative analysis of results and recommendations for improvement based on local conditions. Chou et al. [23
] developed a disaster loss estimation system based on HAZUS (a nationally applicable standardized methodology that contains models for estimating potential losses from earthquakes, floods, and hurricanes in America) to analyze the number of potential buildings damaged and the number of homeless under different intensities of earthquake and counted by district. Due to the uncertainty of disasters, the shelter construction must meet the needs of both day and night disaster response. Based on high-resolution aerial imagery, census data, and high-precision land use data, Yu and Wen [9
] used a GIS-based analytic hierarchy process to assess the need for shelters in daytime and nighttime disaster scenarios. Chen et al. [13
] reported that in addition to considering the time of disaster occurrence, the demand should be differentiated according to different evacuation situations. In the case of an emergency evacuation, the demand should be the maximum number of people present during the day and night; in the case of short-term and long-term refugees, the demand is equal to the number of night residents due to having sufficient notice for organizational actions. To improve the response capability in complex disaster environments, some scholars assessed the shelter demand from the perspective of different disaster types and intensities [8
To understand whether the city can provide adequate shelters and enable evacuees to take refuge nearby, an evacuation simulation is necessary [21
]. Various areas, such as evacuation schemes [25
], decision-making methods [26
], transport modes [27
], influencing factors [28
], evacuation behavior [29
], differences in evacuees [30
], and routes selection [31
] have aroused the research interest. In terms of spatial scale, some researchers selected the entire city as the object, and some focused on specific areas within the city [32
]. In general, few evacuation simulation studies are directly related to the emergency shelter. The key problem with simulation experiments is establishing the relationship between evacuation activities and urban environment to process complex spatial, facility, and road network data as the main parameters [34
]. Yamada [35
] theoretically proposed two methods of network traffic to optimize urban emergency evacuation plans when assigning each resident to a nearby shelter. The first method involves modeling the city as an undirected graph, and the evacuation plan is obtained by solving the shortest path problem on the graph; the second method considers the shelter capacity as a limitation and the problem is transformed into a minimum cost stream goal-oriented solution. Filipe and Kacprzyk [36
] developed an evolutionary learning algorithm for the assignment of paths and shelters based on the assumptions of fairness and global optimality, supplemented with heuristic methods to solve the evacuation problem with limited shelter capacity. This method considers the pedestrian simulation scenario on the road network, the intersections correspond to nodes, and the street segments connecting the intersections are modeled by links. The simulation results are closely related to assumptions such as various traffic scenarios and speeds. Lee and Hong [37
] chose the evacuation of different slope areas under a flood scenario as a case to calculate the possible speed and evacuation distance of evacuees when selecting the shelter locations. The results showed that the evacuation distance of five minutes on flat ground is about 120 m different from that of 15° slope areas. For a dynamic spatial allocation of urban shelters, Yu et al. [38
] developed an agent-based simulation method that can estimate the evacuation time of residents from their locations to shelters and detect the congestion of evacuation routes. Yuan et al. [39
] proposed a traffic evacuation simulation system based on an integrated multi-level drive decision model that generates agent behavior in a unified framework, and the agents’ activities are determined by various existing behavioral models widely used in different driver simulation models. The system can support emergency managers when designing and evaluating more realistic traffic evacuation plans to produce the best evacuation scheme. With the development of emerging technologies, some scholars have begun to focus on intelligent evacuation [40
Geospatial data plays a crucial role in disaster relief. Now, volunteered geographic information (VGI) and web-based crowdsourcing data are being used in the field of disaster management, and GIS have evolved from tools that were limited to situational awareness to fundamental tools supporting. The development and widespread use of online mapping, remote sensing and VGI are providing decision makers with necessary information in emergency response [44
]. Traditionally, emergency managers identified the disaster response period as a blind period, where victims needed to ensure their own safety and security [45
]. Due to volunteer crowdsourcing, people share real-time, time critical and location specific information whenever a disaster occurs, which can help the emergency managers to take necessary actions to reduce the disaster risk, increase the real-time awareness and predict the direction [46
]. In a project of crisis mapping, the researchers used crowdsourced event data from Tweets and Facebook posts to produce the crisis map, and the practice showed that people used the map not only for situational awareness reporting including multimedia such as photos and short videos, but also for offering both help with materials and personal assistance. However, VGI and crowdsourcing are bringing dramatic changes in emergency management [47
This study originated from the practice of emergency shelter planning in China. In the planning, we found it difficult to analyze the high-precision spatial distribution of shelter demand. In addition, the distribution of shelters was usually based on the service radius, which leads to the scientific doubt of the results. Therefore, we try to increase the quantitative basis for shelter allocation through evacuation simulation. At present, few studies focused on the above two issues at the urban scale. Using POI data, and Python programming language, we explored the distribution of shelter demand and shelter status after an evacuation simulation in the downtown areas of Guangzhou. The rest of this paper is organized as follows: Section 2
first presents an overview of the study area and a description of types of data and the data processing, and then introduces the methods of shelter demand assessment and evacuation simulation. Section 3
provides an analysis of the size and distribution of daytime and nighttime shelter demand at the plot scale in study area, along with the distribution of evacuees and the residual capacity of shelter. Section 4
discusses the deficiencies and limitations of this study. Conclusions and directions of future work are presented in Section 5
. We hope that this work can provide a direct quantitative basis for adding new resources for shelters in the process of urban renewal activities, and form a reference for land reuse and disaster prevention space organization in future urban planning.
In the field of disaster relief, accurate determination of population spatial distribution is challenging. In shelter demand analysis, we assessed the distribution of daytime and nighttime population using POI and residential building data with household information. Residents compose the majority of the night population, and the mobile population is so small that we evaluated the nighttime population distribution using only the residential building data with a relatively acceptable error. The working people are the main proportion of daytime population, but the mobile population and other unemployed people also represent a large proportion. In the calculation, combined with the official statistics of practitioners, a mean value method was used to estimate the population size of each POI, and then we totaled the population size in each plot according to the POIs contained in the plot. For the same type of POI, the scale differences between individual POIs may be large, so the population size they represent should differ, resulting in large errors in some plots. However, for the downtown areas of a mega city such as Guangzhou, with a huge and mobile population, the data and methods that can be used to map the population distribution with high precision are limited. In terms of the difficulty of data acquisition, POIs are relatively easy to obtain at the technical level. Although certain errors exist in the data, POIs provide meaningful data to analyze the distribution of the urban daytime population. In the future, to improve the reliability of the results, high-precision spatial-temporal analysis of shelter demand should be conducted with the help of mobile phone signal data or social platform data.
In the evacuation simulation in this study, we assumed that the shelter demand in each plot was concentrated in the geometric center of the plot and shelters were set as points. Therefore, the simulation became a simplified point-to-point multi-objective allocation problem. In reality, a certain distance exists between the two types of points and the boundary of the plot. Because the spatial scale cannot contain roads within the plot, the perpendicular line from the point to the boundary is set as this distance. These perpendiculars also play a role in evacuation routes; the simulation did not consider the micro traffic environment of the plot. If attempting to perfect an evacuation plan, people should take refuge from their locations along the real routes, and the results will provide accurate evacuation distance and required time information, but the data used in this study cannot accurately identify the location of people. In addition, we assumed that the people in each plot are located in the geometric center and were ready for an evacuation, without considering distance and time factors for the evacuation from the interior of the building to the outdoor site, which may result in an underestimation of the evacuation time to varying degrees. The aim of this study was to investigate whether the supply and demand of shelters on the spatial scale of the main urban area can maintain a balance within a reasonable evacuation distance, so addressing the problem of building evacuation at the microscopic level was beyond the scope of this study.
In the setting of the simulation environment, establishing an urban environment model would be ideal if the variables of the evacuation process are as close as possible to the real disaster situation, especially for the traffic environment. The evacuation simulation requires route information. In data processing, to conduct the simulation, we transformed the original complex double-line road network into a single-line network without considering the signal lights. Due to the large spatial scale of the case area, the processed road network neither contained all the branches and lanes nor the roads inside the plots, which decreased the precision of the length of the evacuation route, thereby decreasing the precision of the evacuation time results. A hypothetical condition for the evacuation route network is that the post-disaster roads are intact and the evacuation network is still stable and reliable. Due to insufficient data, we did not assess the reliability of the road network. Therefore, the simulation is an idealized evacuation scheme. In the next step, the unexpected interruption of routes, the reliability of the route network, and the difference in route capacity should evaluated using traffic environment modeling. A feasible solution would be to use traffic big data and consider environmental variables to formulate evacuation plans under different traffic scenarios.
The diverse behaviors and psychology of people experiencing disaster may directly impact the evacuation process and results. In the event of a disaster, based on individual preference or actual conditions, staying at home, taking refuge in situ, traveling to a further shelter, and staying with family or friends are alternative options. In terms of transportation mode, people may choose to evacuate on foot or by public transportation. We aimed to judge the supply and demand for urban shelters through a quantitative evacuation simulation, and to provide a basis for ensuring shelter provide adequate coverage to the surrounding population via the gradual and continuous pre-disaster preparation. This preparation process does not need to distinguish complex behaviors and psychology, so the simulation did not consider them as influencing factors as a walking scenario evacuation. As a possible consequence, in actual evacuation activities, people may tend to concentrate in some of the shelters because of certain behavioral preferences, so an organized evacuation and temporary emergency plans are particularly important. In addition, we did not differentiate evacuation groups such as children, the young and middle-aged, and the elderly. These groups have different needs and abilities during disasters, which directly affect evacuation behavior, the transportation model, and evacuation time, thereby impacting the simulation results as well. Studying small- or medium-scale regions based on more data types and evacuation behavior experiments is necessary.
In any case, the assumptions we set in the simulation become uncertain factors that affect the reliability of the results. The number of evacuees and evacuation time in the results cannot represent the actual situation in disasters, resulting in limited applications in practice. In pre-disaster preparations, we can tell urban planning professionals where there are gaps in shelters, so that we can increase the available shelter resources in future planning, but we cannot give an accurate area. In a possible emergency evacuation, the emergency managers can judge where evacuation risk may exist based on the results of evacuation simulation. In this way, emergency resources can be targeted in the decision-making process. However, the results cannot exactly tell the managers how long the evacuation will be completed and how many people need to be evacuated, because the real evacuation depends on many factors, as assumed and simplified in this study.
Using data of POIs, residential buildings, and industry employee statistics, we assessed the shelter demand and post-evacuation shelter status in the downtown areas of Guangzhou, including Liwan, Yuexiu, Tianhe, and Haizhu. In the analysis of shelter demand and based on the daytime and nighttime disaster response needs, the population distributed in each plot between day and night was defined as the number of potential evacuees. We found significant differences in the size and spatial distribution of shelter demand in daytime and nighttime, and the demand is unevenly distributed within the area. Haizhu, Liwan, and Tianhe had dramatic day–night demand changes. During the day, the demand is mainly distributed in Yuexiu and Tianhe, which have mature urban functions and numerous enterprises. At night, the demand is mainly concentrated in the plots with residential buildings, with the demand in other plots being obviously less. The total shelter demand for day and night refuging was about 7.929 million people, which is 2.454 million more than the resident population. The evacuation simulation showed that the average evacuation time of all 16,883 routes was 12.6 min, and some were more than three hours due to the huge number of people located at the demand points. After the evacuation, 558 of 888 shelters exceeded their capacity to varying degrees. In terms of shelter capacity, Haizhu had the largest over-capacity of 1.62 million people. Conversely, for the residual capacity, shelters that did not exceed the capacity in Tianhe can accommodate another 890,000 people, which was the highest amount. The results indicated that the supply and demand of shelters in the study area was unbalanced and the distribution was uneven, preventing the shelters from completely covering the potential evacuees. Our findings provide a direct quantitative basis to guide the amount and size of new shelter resources during urban renewal activities, and being a reference for land reuse and disaster prevention space organization in future urban planning.
In future research, improving the accuracy of shelter demand assessment based on more data types is necessary. During disasters, numerous complex variables affect evacuation, such as unexpected interruption of roads, physical obstacles, secondary disasters, and human behavior and psychology, which have been difficult to quantify in related practices and research fields. The next step in evacuation simulation is to set environmental variables as close as possible to the post-disaster conditions for small- or medium-scale areas. In addition, we established a database containing demand points, emergency shelters, and evacuation routes information. Therefore, building a decision-making system and management platform for shelter planning will be another research direction.