Dynamic Assessment of Population Exposure to Urban Flooding Considering Building Characteristics

Abstract

Under intensifying climate change impacts, accurate quantification of population exposure to urban flooding has become an imperative component of risk mitigation strategies, particularly when considering the dynamic nature of human mobility patterns. Previous assessments relying on neighborhood block-scale population estimates derived from conventional census data have been constrained by significant spatial aggregation errors. This study presents methodological advancements through the integration of social sensing data analytics, enabling unprecedented spatial resolution at the building scale while capturing real-time population dynamics. We developed an agent-based simulation framework that incorporates (1) building-based urban environment, (2) hydrodynamic flood modeling outputs, and (3) empirically grounded human mobility patterns derived from multi-source geospatial big data. The implemented model systematically evaluates transient population exposure through spatiotemporal superposition analysis of flood characteristics and human occupancy patterns across different urban functional zones in Lishui City, China. Firstly, multi-source points of interest (POIs) data are aggregated to acquire activated time of buildings, and an urban environment system at the building scale is constructed. Then, with population, buildings, and roads as the agents, and population behavior rules, activity time of buildings, and road accessibility as constraints, an agent-based model in an urban flood scenario is designed to dynamically simulate the distribution of population. Finally, the population dynamics of urban flood exposure under a flood scenario with a 50-year return is simulated. We found that the traditional exposure assessment method at the block scale significantly overestimated the exposure, which is four times of our results based on building scale. The proposed method enables a clearer portrayal of the disaster occurrence process at the urban local level. This work, for the first time, incorporates multi-source social sensing data and the triadic relationship between human activities, time, and space in the disaster process into flood exposure assessment. The outcomes of this study can contribute to estimate the susceptibility to urban flooding and formulate emergency response plans.

1. Introduction

Over the last two decades, flooding has been one of the most far-reaching and severe disasters globally [1]. Urban flooding, currently a common natural disaster, poses a serious threat to sustainable urban development and people’s lives, exacerbated by global climate change and ever-increasing urbanization [2,3]. Furthermore, projections indicate that by 2030, the urban population will constitute 60% of the global total [4]. The significant transformation of urban land use patterns, chiefly led by the rapid expansion of construction land, has escalated the stress on urban drainage systems during extreme rainfall, thereby intensifying the adverse socio-economic consequences of urban flooding [5]. Therefore, for populations most impacted by these disasters, precise risk assessments or predictions are vital to implement timely measures and alleviate their effects.

Modeling the collective consequences of hazards within the urban system and human behavior within the hazard environment is critical to assess the risks of natural hazards. Flood risk emerges from the dynamic interaction between natural hazards and human vulnerability [6]. Populations, as one of the most dynamic entities within geographical environments, exhibit spatial heterogeneity and temporal dynamics in distribution [7]. The interactions between flooding and dynamic population distribution is critical for the exposure assessment of urban flood disasters. This necessitates the integration of the social drivers of the human system and the natural drivers of the hydrological system as endogenous variables within the urban ‘human-water’ nexus [8]. Consequently, given the ‘natural-social’ nature of urban flooding, conducting a risk assessment requires a comprehensive understanding of interaction between nature and social behavior. Due to this complexity, studies in this field are categorized under complex adaptive systems.

Modeling and simulating problematic occurrences frequently utilizes research methodologies such as Bayesian Networks, Microsimulation, Cellular Automata, Artificial Intelligence, and agent-based modeling (ABM) [9]. Amongst all these methods, ABM is a model that is driven by physical rules and inputs to simulate the dynamic distribution. ABM composed of agents as proxy objects equips researchers with the capability to construct, scrutinize, and experiment with models comprising these interactive agents within a given environment. Agents operate according to specific rules, facilitating complex and heterogeneous interactions. Once aligned with a specific geographical locale, these agents can be incorporated into a simulation environment using geographic techniques [10,11]. This bottom–up modeling approach for intricate systems is widely applied in urban flood disaster risk assessment [12,13]. Dawson et al. proposed an agent-based flood risk management model that integrates multi-agent simulation with hydrodynamic modeling to assess flood vulnerability under various scenarios [9]. Zhu et al. employed ABM to simulate urban residents’ behaviors during flood events and introduced a high-resolution dynamic exposure modeling approach [14]. O’Shea et al. utilized a hydrodynamic agent-based model (HABM) to investigate the influence of direct and indirect warnings on flood response mechanisms [15].

Current research using ABM on dynamic simulation methods for flood exposure primarily adopts urban environments as the hazard-prone environments. These studies simulate the interactions among hazard-prone environments, hazardous factors, and exposed elements, with a particular focus on modeling population response behaviors in disaster scenarios. Within the urban environmental system, human activities is another fundamental component in ABM for simulating urban flood disasters. Spatially, urban functions are typically delineated based on land use, with the community or block level commonly used as a detailed, fine-scale unit [14]. Temporally, human activities are driven by behavior rules that guide agents, enabling the exploration of spatial interactions during flood events. Cities are complex systems. On one hand, the complexity of urban environmental systems results in localized variations in flood disasters, and the evaluation of spatial resolution at the geographic unit level significantly influences simulation outcomes [16,17]. Several studies have refined flood risk assessments to the building level, thereby significantly enhancing the accuracy of these evaluations [18,19,20,21]. On the other hand, geographic entities within the urban environment impose spatiotemporal constraints on human behavior. For instance, store opening hours and school schedules determine the scope of human activities. While current research has focused on modeling behavioral rules in response to disasters, less attention has been paid to the spatiotemporal constraints imposed by geographic environmental factors on human behavior. Thus, we draw attention to buildings as specific geographic entities within urban environments. As key places for human activities, buildings bear the socio-economic attributes and characteristics of activities, interactions, and emotions [22]. Compared to land use or blocks, buildings offer a finer granularity within the urban environmental system.

Traditionally, obtaining detailed information about urban geographical features, such as land and buildings, has often relied on surveys. However, the continuous influx of social media and VGI data have paved the way for researchers to garner more intricate urban semantic information [23] and provided new data sources for assessing urban flood simulations [24,25,26,27,28]. The regular updates and extensive data from points of interests (POIs) and social reviews, supported by GIS spatialization technology, offer a novel perspective for geospatial analysis and expression [29]. For example, the category of a POI directly reflects the function of the associated building, while business hours from review websites define the active periods of individuals within these spaces. By effectively integrating and utilizing such data, it is possible to enhance the constraints within agent-based modeling (ABM) simulations, significantly improving the accuracy and reliability of urban flood risk assessments.

This study aims to propose a novel method for dynamic population exposure to urban flooding using agent-based modeling. A building-level urban environmental model is developed and an agent-based model is designed that integrates the semantic attributes of buildings. New datasets from social sensing data are employed for dynamic population simulation. This model examines the spatiotemporal dynamics of urban populations during extreme rainfall-induced flooding events, enabling the quantification of dynamic population exposure in such extreme flood scenarios.

2. Study Area and Data

2.1. Study Area

The primary urban region of Lishui City, located in the southwestern part of Zhejiang Province, China, serves as the focus of this study (Figure 1). Lishui lies within a subtropical monsoon climate zone, characterized by frequent extreme weather events, such as intense rainfall and localized torrential downpours. Flooding, often triggered by heavy rainfall, remains the region’s most significant disaster. Notably, in 2014, Lishui experienced a catastrophic torrential rain event, classified as a 50-year return period event, which resulted in severe economic losses due to widespread flooding of roadways, traffic disruptions, and extensive damage to residential areas.

2.2. Data Sources

The data used in this study primarily consist of basic geographic data, open-source social sensing data, and population demographic data (http://tjj.lishui.gov.cn (accessed on 10 December 2024)).

1. Basic geographic data:

These are provided by the local government, including DEM data, land use, drainage facilities data, road data, and hydrological data, which are used for urban environmental modeling and urban flooding simulation. Additionally, building footprints and demographic statistics are included.

2. Open-source social sensing data:

POIs data, comprising category, location, and other attribute information, have emerged as a vital data source for urban geographical research within the domain of open-source social sensing. The acquisition of POI data was conducted from three sources: Baidu Maps (https://map.baidu.com (accessed on 10 December 2024)), Ctrip (https://www.ctrip.com (accessed on 10 December 2024)), and Meituan (https://lishui.meituan.com (accessed on 10 December 2024)). Baidu Maps is a prominent online electronic mapping service in China, while Ctrip serves as the country’s largest travel platform. Meituan, serving as a popular review website in China, has distinct characteristics in comparison to electronic map POI data, as it includes comprehensive information regarding business hours and addresses for POIs. To achieve a comprehensive dataset, a merging process was undertaken through toponym matching, resulting in the aggregation of 11,103 POIs.

3. Methodology

In this paper, the assessment of population exposure to urban flooding is based on the high-resolution exposure assessment method proposed by Zhu et al. and Dai et al. [14,30]. Zhu et al. [14] first proposed a high-resolution dynamic exposure modeling method by using ABM to simulate residents’ activities during flooding. The method consists of three steps: first, urban flood simulation using the LISFLOOD-FP model to predict the spatial and temporal distribution of floods; second, using an agent-based model to simulate the movement of residents during urban floods; and finally, exposure to urban floods was simulated for populations, roads, and buildings. Building on this, Dai et al. expanded the framework by designing a computational approach for evaluating urban rainfall-induced disaster exposure, which includes four main components:

1. Hazard-related urban environment modeling: The urban area is divided into irregular blocks based on road and utility networks. Blocks are classified by land use, mainly residential, commercial, entertainment, or educational, representing key locations for daily activities.

2. Agent-based modeling of human activity: Population is modeled using an agent-based approach to simulate the spatiotemporal distribution and dynamics of human activities. Agents start in buildings and move along the road network based on behavioral rules, with scenarios for both normal and disaster conditions.

3. Human–hazard coupled method: This method combines disaster processes with agent-based simulations, coupling hazard evolution with human activities to capture their interaction.

4. Dynamic exposure assessment: Focusing on buildings, roads, and populations, this approach overlays exposure data (e.g., building locations) with hazard scenarios (e.g., flood extent) to calculate dynamic exposure values.

In summary, the calculation of population exposure to urban flooding in this study includes three main components: urban environment modeling, human activity simulation, and dynamic exposure assessment (Figure 2). The first step involves acquiring rich building attribute information from multi-source geographic data to construct the urban environment at the building scale. The second step involves designing an agent-based model that considers both urban flooding and building characteristics to simulate the spatiotemporal distribution of the population. Finally, an extreme rainfall scenario over a 50-year return period is simulated using LISFLOOD-FP to quantify dynamic population exposure.

3.1. Urban Environment Modeling at Building Scale

The assessment of urban flood exposure based on ABM explores the interactions and influences among various elements such as the urban environment, population, and disaster factors across both temporal and spatial dimensions. From the perspective of the urban environment, the key variables of interest in the ABM simulation include the functional characteristics of urban areas, which serve as the initial basis for population distribution, as well as the temporal constraints imposed by urban geographical features that affect daily population movements. Consequently, this study constructs a building-scale urban environmental model, focusing primarily on two characteristics: the functionality of buildings and their activity times. The POI data provide rich semantic information regarding the locations and names of various places of interest, reflecting the distribution of different elements within the urban context [31,32], and have been widely applied in the study of urban functional zoning [33,34]. Furthermore, in recent years, researchers have gradually employed relevant algorithms to achieve functional recognition at the building scale [35]. On the other hand, buildings, as carriers of human activities, often have temporal attributes that define when these activities occur. Most buildings exhibit regular activity hours, with schools and government offices having specified opening times. In contrast, commercial buildings may have more variable hours of operation. Residential buildings are generally considered to have human activities occurring 24 h a day. Additionally, the increasingly rich information from review websites allows for the quantification of operating hours. Therefore, this study extracts the functional characteristics and temporal attributes (activity time) of buildings from multi-source data, including online maps and review websites, ultimately achieving urban environmental modeling at the building level. The processing flow is illustrated in Figure 3.

1. The integration of POIs from Baidu Maps, Meituan, and Ctrip consists of three steps. The first step is data cleansing, which primarily involves removing duplicates and handling missing values to ensure the consistency and quality of the input data. Next, the data integration phase involves matching POIs from different sources based on their names, merging field information to create a unified dataset. Finally, in the reclassification step, criteria are defined for each category, allowing the POIs to be reclassified into four types: residential, commercial, educational, and recreational.

2. Since building footprints contain only geometric information without any attribute data, a common approach is to establish a buffer zone around the buildings to capture surrounding POIs [36,37]. However, due to positional inaccuracies of the POIs and the potential for classification errors in densely built environments, this method can be problematic. Therefore, this study utilizes the address information of POIs to identify those with similar addresses within the buffer zones of buildings. By employing Natural Language Processing (NLP) tools, the addresses of the POIs are parsed into address components. Utilizing the Baidu Maps API, the addresses are standardized, and any missing address components are filled in. Finally, the Jaro–Winkler distance formula is utilized to calculate the degree of similarity between address strings with the following formula:

$$d_{jaro}\left(S_1,S_2\right)={\displaystyle \frac{1}{3}}\left({\displaystyle \frac{m}{\left|S_1\right|}}+{\displaystyle \frac{m}{\left|S_2\right|}}+{\displaystyle \frac{m-t}{m}}\right)$$

(1)

$$d_{jaro-winkler}\left(S_1,S_2\right)=d_{jaro}\left(S_1,S_2\right)+\left(lp\left(1-d_{jaro}\left(S_1,S_2\right)\right)\right)$$

(2)

where $\left|S_i\right|$ is the length at $S_i$, $m$ is the number of matched characters, $t$ is the number of character transitions, $d_{jaro}\left(S_1,S_2\right)$ is the Jaro distance between $S_i\mathrm{a}\mathrm{n}\mathrm{d}$ $S_2$, $l$ is the prefix match length, $p$ is the prefix match weight. Higher values of Jaro–Winkler computation indicate higher similarity.

If a POI falls within the buffer zones of multiple buildings, its associated building is determined based on the degree of address similarity. On this basis, the type and activity time of the POIs are assigned to the corresponding building footprints.

3. For buildings that cannot be identified through spatial similarity, this study employs Kernel Density Estimation (KDE) method to determine their functions [37]. The spatial distribution differences of POIs reflect, to some extent, the distribution characteristics of building functions. Therefore, the KDE method is employed to convert POIs into a continuous density surface that represents these distribution characteristics, which can be utilized to infer building functions.

The KDE calculates the density contribution of each sample point to the central point within a specified range, generating a smooth surface through a kernel function to represent the density distribution. Taking into account the variation in the quantity of POI data across different categories and their uneven spatial distribution, this study utilizes normalized KDE for density calculations, which can be expressed with the following formula:

$$f\left(x\right)={\displaystyle \frac{1}{N{h}^d}}\sum _{i=1}^{N}K\left({\displaystyle \frac{x-x_i}{h}}\right)$$

(3)

$f\left(x\right)$ is the KDE computation function at position $x$, $d$ is the spatial dimension, $h$ is the bandwidth, $N$ is the number of points at distance $x\le h$ from position $x$, and $K$ is the spatial weight function.

3.2. Agent-Based Modeling of Human Activity

ABM uses a type of computational model that can simulate the actions and interactions of autonomous agents in order to assess their effects on the system [38]. Human movement within a city follows certain patterns, and studies have shown that 93% of human behavior can be predicted [7], indicating a regular system of time utilization in society [39]. In this paper, we develop a multi-agent model of human activities that incorporates interaction rules among agents. Three types of intelligence are defined in this study: human agents, road intelligence, and building intelligence. The human agents are further categorized into home population, pendulum population, and random population based on their behavioral characteristics.

In this paper, the behaviors, states, and activity locations of human agents are defined by behavioral rules, as illustrated in Figure 4. The state of human agents is linked to their behaviors and activity locations, with roads serving as the medium through which their mobility is realized. The study simulates the behavior of human agents in three stages. In the initialization phase, human agents are positioned in residential buildings, with their state set to “rest”. The commuting phase, initiated by the commuting start time, prompts human agents to transition from their residential buildings to their target buildings, following the respective behavioral rules. This movement persists until the commuting end time, which marks the onset of the subsequent phase. Concurrently, the state of the human agents undergoes a transition, and at the conclusion of the commuting period, they initiate their return journey from the target building to their residential or other buildings. In the risk response phase, human agents detect environmental risks and respond accordingly. These environmental risks are defined by the reduced or impeded mobility of roadways due to inundation, resulting in the obstruction of human ingress and egress to and from buildings. In the event of road closures due to flooding, human agents proactively circumvent potentially hazardous routes, opting for less hazardous alternatives to reach their desired destinations. In instances of building inundation, human agents are expected to avoid entering or remaining within affected structures.

The GAMA Platform has been selected as the primary ABM simulation platform. In the context of integrated modeling with GIS, the GAMA Platform demonstrates considerable advantages in data integration, spatial analysis, planning, decision-making, and the modeling of complex systems [40]. It can also provide valuable feedback based on population activity patterns, environmental conditions, and external disturbances.

3.3. Dynamic Exposure Assessment Based on Scenarios

Urban floods result from the interaction of various factors, including rainfall, surface runoff, surface flow convergence, and drainage network convergence, and represent a dynamic process. To accurately simulate this process, this study follows the methods outlined in the relevant literature [14]. The main input for urban flood modeling considers rainfall, with a 50-year return period extreme rainfall event designed using the Chicago Method (CHM).

$$I={\displaystyle \frac{1265.3\times \left(1+0.578\times lgp\right)}{167\times {\left(T+5.919\right)}^{0.611}}}$$

(4)

where I is the average storm intensity in mm/min, T is the rainfall duration in minutes, and P is the rainfall recurrence period. The relevant parameters were obtained from official sources. A 2 h rainfall duration was selected to model and simulate the process in which rainfall intensity decreases over time within a specified recurrence period, with the average intensity gradually diminishing as the duration increases.

The two-dimensional hydrodynamic model LISFLOOD-FP is coupled with the urban stormwater management model SWMM to simulate the hydrological processes of urban surfaces and the hydrodynamic behavior of underground pipeline networks, enabling the exchange of water between surface runoff and the underground drainage system. LISFLOOD-FP is based on a grid model that integrates 1D and 2D hydraulic models. Since its release, it has been widely applied in various research scenarios, including global river flooding and urban flooding, yielding promising results. The model describes diffusive flow using discrete continuity and momentum equations on a grid, allowing it to represent two-dimensional dynamic flow fields.

The main influencing factor of flood hazard in Lishui City is specified as water depth in the existing studies [41]. Therefore, in this paper, only the simulated water depth is considered as the basis of judgment [14,20]. Exposure is categorized into four classes of low, medium, high and extreme according to different inundation water depths as in Equation (5).

$$S_p=f_p\left(H\right)=\left\{\begin{array}{c}1 0.25<H\le 0.75\\ 2 0.75<H\le 1.50\\ 3 1.50<H\le 2.50\\ 4 H>2.50\end{array}\right.$$

(5)

Water depth exceeding 0.25 m but less than 0.75 m is classified as low risk. Water depth ranging from 0.75 m to 1.5 m is categorized as medium risk. When the water depth falls within the interval of 1.5 m to 2.5 m, it is deemed high risk. Furthermore, any water depth greater than 2.5 m is considered to represent extreme flood exposure.

4. Results

4.1. Urban Environment Modeling Based on Multi-Source Data

The study area yielded a total of 12,045 POIs from Baidu Maps, 1680 review-based stores from Meituan, and 789 stores from Ctrip. After a comprehensive data cleaning and integration process, the final dataset comprised 11,103 POIs, including 1163 residential-related POIs, 358 school-related POIs, 3797 company-related POIs, and 5785 recreation-related POIs. The spatial distribution of these points is shown in Figure 5. The analysis reveals a notable concentration of residential and commercial areas within the study region. At the block level, commercial and residential buildings are interspersed. Additionally, a temporal analysis of building attributes identifies a clear tidal phenomenon, with urban populations aggregating during specific periods, as depicted in Figure 6.

4.2. Dynamic Simulation of Population Distribution

We utilized the ABM platform to initialize 14,580 agents and simulate population movement under a 50-year return period rainfall scenario. The program completed in 4 h and 23 min on a machine with the following specifications: CPU—AMD Ryzen 7 3700×, RAM—32 GB. As illustrated in Figure 7, the simulated human activity density exhibits a marked increase around 8:00 AM, coinciding with the transition of individuals from residential areas to non-residential spaces, such as commuting to work or school. The number of people engaged in outdoor activities undergoes a significant surge, with the majority of individuals departing from their homes and entering urban roadways. By 10:00 AM, a significant proportion of the working population has arrived at their respective workplaces, while students have commenced their journey to educational institutions. This results in a decline in the number of commuters on the roads. During the day, the population activity is predominantly concentrated in the southwestern part of the study area, which corresponds to the city center. This spatial distribution aligns with the locations of non-residential buildings. Conversely, by approximately 10:00 PM, individuals typically return to their residences, leading to a reallocation of the population towards the residential zones.

In order to validate the accuracy of human activity simulation, two methods were employed. First, a comparison was made between the population distribution data from Baidu Huiyan and the population movement simulated in this study. Baidu Huiyan data are based on the location data of over one billion monthly active users on Baidu, and it provides a reliable representation of human activity in urban areas. This dataset has been widely used in various research applications [42]. As demonstrated in Figure 8a,b, the simulation results were validated by comparing the population flow peaks at 10:00 and 18:00. The statistical outcomes of Baidu Huiyan data, based on a 500 m resolution grid within the study area, served as the ground truth. These were then compared to the simulated population distribution at the same scale. The results indicate a strong correlation between the simulated population distribution and the Baidu Huiyan data. Additionally, to further validate the plausibility of population movement, we referenced the navigation distances provided by the Baidu Maps Navigation API, as used in related studies [43]. From 8:00 to 10:00 and 18:00 to 20:00, during these key periods, the travel distance of the simulated agents was compared with the travel distance derived from Baidu Maps’ path retrieval, using the starting and ending points of the paths as the departure and destination locations, respectively. The results are shown in Figure 8c,d, demonstrating a high correlation between the two, with an R² value of 0.92. In summary, the population activity simulation results in this study accurately reflect the actual conditions in the study area.

4.3. Results of Dynamic Exposure Assessment

In the context of a 50-year return period rainfall disaster scenario, the temporal and spatial parameters are specified as follows: rainfall was set to begin at 7:00 a.m. and persist for two hours. The rainfall intensity was calculated using the Lishui City formula in conjunction with the Chicago rainfall pattern. The water depth in the study area was simulated. The parameters for this model were derived from prior studies [14,20,30], and the flood simulation results were applied in this paper. Previous research has extensively explored the methods for flood simulation within the study area and provided measured data for validation. We will not elaborate on these details here. The simulation of population activity assumes that 90% of the population follows a pendulum-like movement pattern. The results show that the flooded area reaches its peak after approximately 40 min, as illustrated in Figure 9. A total of 2352 buildings are affected, representing 19.52% of the total number of buildings in the study area. When integrated with the population distribution, the impacts of the disaster at four distinct levels are as follows: the low level affects 1920 individuals, the medium level affects 450 individuals, the high level impacts 80 individuals, and the extreme level affects 10 individuals.

Figure 10 illustrates the impacted population throughout the flooding process under the aforementioned scenario. As the rainfall intensity reaches its peak at 40 min, the global exposed population concurrently attains its maximum at 50 min. Following a decrease in rainfall intensity, the behavioral patterns of human agents exhibit a normalization, leading to slight fluctuations in the exposed population. With the cessation of rainfall, the high-exposure population dissipates, resulting in a further reduction in the total global exposed population.

4.4. Comparison with Block-Scale Simulations

The above flood simulation results are overlaid with blocks, as shown in Figure 11a. Taking the red polygon as an example, the entire population of the area is considered as exposed population at the block scale, as illustrated in Figure 11b. Conversely, at the building scale, only the population associated with the relevant buildings is counted as the exposed population, as shown in Figure 11c. When the maximum inundation occurs, the research area experiences four levels of disaster impact: mild level with 4000 people, moderate level with 18,910 people, severe level with 9850 people, and critical level with 2690 people. This result is significantly larger than the simulation based on buildings. The total number of people affected is four times greater than that of the simulation based on buildings.

This study compares and analyzes the dynamic changes in the number of exposed individuals within a block under different modeling approaches, using a 50-year return period storm event with a rainfall duration of 2 h as the disaster scenario. The block shown in Figure 11 was selected as a typical case study, primarily consisting of commercial buildings. By simulating two approaches—building-scale and block-scale modeling—the variation in the number of exposed individuals during flooding is presented in Figure 12. The simulation results indicate that human agents exhibit distinct risk-responsive behaviors under the flood disaster scenario. As urban flooding progresses, the number of exposed individuals increases in both block-scale and building-scale simulations. Notably, the number of exposed individuals estimated through the building-scale simulation is significantly lower than that derived from the block-scale simulation.

5. Conclusions

This study investigates the behavior of urban populations during urban flooding events. It introduces a novel social sensing data source that extracts building characteristics, enhancing the agent-based modeling approach used for assessing urban flood exposure. The empirical research was conducted in the Lishui city district of China, simulating population dynamics under a heavy rainfall scenario with a 50-year return period. The findings provide insights into the affected population during maximum inundation scenarios and the varying degrees of damage experienced.

1. ABM at building scale provide higher accuracy. In the experiments presented in this paper, under identical conditions, the block-based assessment of the affected population is found to be four times higher than the building-based assessment.

2. Multi-source social sensing data enhance the spatiotemporal constraints of ABM. POIs, as typical social sensing data, can extract the functional and human activities of buildings, thereby establishing a ternary relationship of “people–spacetime–disaster”.

In this paper, social sensing data are integrated with the simulation of population dynamics in urban flooding disasters. Further exploration in this area is possible. First, additional data sources could be introduced to enrich building height information and incorporate psychological and behavioral factors to examine population movement under different scenarios. Second, advanced computational methods, such as reinforcement learning, could be employed to address the limitations of rule-based systems and more accurately simulate real-world population distributions.