The Impacts of Heatwaves on Population Distribution in the Subtropical City: A Case Study of Nanchang, China

Abstract

Global warming has intensified the frequency and intensity of heatwaves, particularly in urban areas, significantly affecting residents’ daily activities. Extant studies have mainly concentrated on the relationship between socio-economic attributes and the impacts of heatwaves on urban populations. However, the relationship between the built environment and the impacts of heatwaves on urban population distribution has not received much attention. Furthermore, most studies have overlooked the temporal heterogeneity in heatwave impacts on population activities and distribution. Therefore, taking the central urban area of Nanchang as the case, this study investigated the impacts of heatwaves on population distribution and their temporal heterogeneity. Moreover, it identified the nonlinear relationships between built environment factors and population changes during heatwaves by using the XGBoost model and SHAP method. The results revealed that heatwaves exerted the largest impacts on population distribution during weekend nights, followed by weekend daytime and weekday nighttime, with the least impacts observed during weekday daytime. Furthermore, location and transportation factors significantly affected population changes during heatwaves across most time periods, with their influences being associated with policy factors such as the high-temperature leave policy for workers in industrial zones located in urban fringe areas and the cooling zone establishment policy for citizens in subway stations. Moreover, land use and building form factors exhibited significant temporal heterogeneity in their impacts on population changes during heatwaves. This temporal heterogeneity was fundamentally driven by individuals’ heat adaptation behaviors, the spatiotemporal patterns of their daily activities, and the diurnal variations in the built environment’s influence on local thermal environment. These findings provide valuable insights to proactively alleviate the adverse impacts of heatwaves.

1. Introduction

Over the past few decades, global temperature has risen markedly, resulting in a substantial increase in the frequency and intensity of heatwaves worldwide, particularly in subtropical areas [1,2,3]. This trend is projected to persist in the future [4]. Meanwhile, rapid urbanization, characterized by population and industrial agglomeration as well as urban development and construction, has intensified high-temperature and heatwave phenomena in urban areas [5], adversely affecting various aspects such as residents’ activities [6,7,8], public health [9], labor productivity [10,11], and transportation [12,13]. Consequently, evaluating the impacts of heatwaves on urban population distribution and elucidating their underlying mechanisms are essential for enhancing urban resilience and mitigating heatwave impacts.

In order to identify the impacts of heatwaves on urban populations, researchers have conducted extensive studies from multiple perspectives, including population heat exposure risk assessment [14,15,16], mobility impacts [6,7], as well as changes in residential locations [17] and transportation modes [13,18,19]. Heat exposure risk assessment studies have constructed an evaluation framework by integrating three key factors—thermal hazards, exposure, and vulnerability—to assess thermal health risks at the urban or regional scale, and to help identify areas with elevated thermal health risks. These studies revealed that socioeconomic status and built environment characteristics significantly influenced residents’ thermal health risks [14,15]. Research on the effects of extreme high temperatures or heatwaves on population mobility [6,7], residential locations [17], and transportation modes [13,18,19] has demonstrated that the inherent patterns of residents’ daily activities undergo significant alterations during heatwaves. For instance, Gu et al. [6] analyzed changes in population mobility in Houston under extreme high temperatures and found that extreme high temperatures tended to inhibit movement toward the city center while promoting movement toward suburban areas. Wu et al. [18] examined the influence of weather conditions on respondents’ transportation mode choices and found that under extreme weather conditions, residents were more inclined to choose subways and private vehicles as their modes of transportation.

Although the extant studies provide valuable insights into the impact of heatwaves on urban populations, they have certain limitations. Most studies of population heat exposure rely on static census data [14,15], which lack timeliness and fail to capture short-term changes in urban population activities and distribution during heatwaves [16]. Some studies have indicated that in complex urban environments, dynamic variations in population distribution significantly influenced population exposure levels during heatwaves [20,21]. A critical aspect in effectively evaluating the impact of heatwaves on urban populations is obtaining detailed and accurate population distribution data. Mobile phone signaling data fulfill this need by providing high-resolution information on population distribution in both temporal and spatial scales. In recent years, several studies have utilized mobile phone signaling data to analyze how extreme high temperatures and heatwaves affect urban population mobility and distribution [6,7]. These studies have mainly concentrated on the correlation between socioeconomic attributes and the degree of heatwave impact, such as age, race, and wealth, extending their analyses to issues of population inequality and environmental justice in the impact of heatwaves [6,7,17,20]. However, the relationship between built environment and the influence of heatwaves on urban population distribution has been rarely touched. Numerous studies have demonstrated that built environment factors are associated with both the urban thermal environment [22,23,24,25,26,27] and population distribution [28,29,30,31]. In terms of the relationship between the built environment and the urban thermal environment, several studies show that residential and industrial areas exhibit relatively higher surface temperature compared to parks and riverside areas [23]. Furthermore, buildings are a critical driver of urban heat, and variations in building density and height significantly influence their impacts on surface temperature [24,25,26]. Regarding the relationship between the built environment and population distribution, extant studies have shown that factors such as the number of residential units, commercial facilities, and green spaces, as well as distance to subway stations and the city center, are significantly associated with urban population distribution patterns [28,29,30,31]. Moreover, the built environment plays a crucial role in influencing urban residents’ thermal health risks [14,15]. One study evaluating urban heat health risks in Hong Kong revealed that high building density was a common characteristic of communities with extremely high heat health risks [14]. Another study further demonstrated that different building types exerted distinct influences on human exposure to heat [15]. Consequently, there may be a correlation between the impacts of heatwaves on urban population distribution and the built environment, and the underlying mechanisms warrant further exploration. Additionally, the analysis of temporal heterogeneity in the impact of heatwaves on urban population distribution remains limited. Previous studies have demonstrated that urban population activities and distribution exhibit significant temporal heterogeneity across weekdays versus weekends, and different times of day [28,29,30,32,33]. Furthermore, the urban heat island and its relationship with the built environment also demonstrates temporal heterogeneity [26,34]. Therefore, investigation of the impacts of heatwaves on urban population distribution should take temporal variations into account.

Recently, nonlinear analysis methods, such as machine learning, have been extensively applied to the study of urban population activities and distributions, revealing nonlinear and threshold relationships between built environment factors and population [35,36,37,38,39]. Moreover, some studies have demonstrated that the urban thermal environment and the built environment are also nonlinearly related [22]. Consequently, the relationship between the impacts of heatwaves on urban population distribution and the built environment may also exhibit nonlinearity, necessitating the use of machine learning methods for a further in-depth exploration.

Therefore, this study takes the central urban area of Nanchang as the research area and utilizes hourly resolution population heat map data derived from mobile phone signaling to examine the impacts of heatwaves on urban population distribution. Specifically, we investigate the differences in these impacts between weekdays and weekends, as well as between daytime and nighttime. Furthermore, the XGBoost model combined with the SHAP method is employed to explore the nonlinear relationships between built environment factors and population changes during heatwaves at the grid scale. This research contributes to elucidating the detailed mechanisms underlying the influences of heatwaves on urban population distribution and their association with the built environment, and provides planning and management recommendations to proactively alleviate the adverse impacts of heatwaves.

The remainder of this paper is structured as follows: Section 2 elaborates the research materials and methods. Section 3 examines the general population distribution, provides a descriptive analysis of the impacts of heatwaves on urban population distribution across different time periods, and investigates the nonlinear relationships between built environment factors and population changes during heatwaves in these periods. Section 4 discusses the main findings of this study and proposes planning and policy recommendations to alleviate the adverse effects of heatwaves. Finally, Section 5 summarizes the key research findings and points out the limitations of the study along with directions for future research.

2. Materials and Methods

2.1. Study Area

Nanchang, located in southeastern China, is the capital of Jiangxi Province with an administrative area spanning 7195 km² and housing 6.57 million residents. Nanchang has a subtropical monsoon climate characterized by abundant sunshine and rainfall, with peak temperatures typically occurring in July and August. The region exhibits a hot and humid summer with the extreme maximum temperature of 40.9 °C. This climatic profile designates Nanchang among China’s four “Furnace” cities. Over the past 45 years, Nanchang has experienced rapid urbanization, with its population increasing sharply from 3.07 million in 1978 to 6.57 million in 2023. This population growth, coupled with urban expansion, has significantly intensified the urban heat island effect in Nanchang, resulting in a higher frequency of heatwaves than before. These heatwaves have posed substantial threats to both residential and environmental health [40,41,42]. This research selected the central urban area of Nanchang as the study area. According to the master plan, the study area consists of Donghu District, Xihu District, Qingshanhu District, Qingyunpu District, Honggutan District, as well as part of Xinjian District and Nanchang County, covering a total land area of 1198 km² (Figure 1). The Ganjiang River passes through the central urban area, dividing it into two halves. The eastern side is often referred to as the old urban area, while the western side is defined as the new town area.

2.2. Data

2.2.1. Weather Records

Hourly weather records from the main weather station in the central urban area of Nanchang during July to August 2024 were obtained from the China Meteorological Administration (CMA) to identify heatwaves and information about other weather conditions. The dataset comprised three key variables: air temperature, relative humidity, and precipitation. The selection of July to August was because this period represents the peak frequency of heatwaves in Nanchang.

2.2.2. Population Heat Map Data

The population heat map data used in this study were sourced from telecommunications providers, leveraging anonymized location visit records from mobile users. These datasets were spatially aggregated into standardized 800 m by 800 m grids based on cellular base station positioning. Each grid’s attributes include its center coordinates and the corresponding population density metrics. We obtained datasets with hourly temporal resolution for July and August 2024.

2.2.3. Data About the Built Environment

To investigate the built environment characteristics of each grid, the information for the heights and areas of buildings was sourced from the 3D-GloBFP building footprint dataset [43]. In addition, land use data were acquired from the Department of Natural Resources of Jiangxi province, and the acquisition process adhered to relevant confidentiality regulations. Further, the locations of subway stations were sourced from the 2024 Point of Interest (POI) dataset for Nanchang retrieved via Gaode Map. Finally, all these datasets were connected to each 800 m by 800 m grid section in the study area through ArcGIS 10.8.

2.3. Methods

2.3.1. Identifying Heatwave Events

The literature on heatwaves generally states the lack of a universal definition for measuring heatwave events, with significant variability in temperature metrics, duration criteria, and other indicators across different studies [44]. Considering the humid and hot climate characteristics of Nanchang during summer and drawing on definitions from the relevant literature [44,45], this study adopted the heat index (HI) as the temperature metric for defining heatwave events, with a threshold of 40.6 °C. The HI is a metric promoted by the National Weather Service (NWS) of the United States and is widely used to represent human-perceived equivalent temperature based on air temperature and relative humidity [46,47], making it particularly suitable for the hot and humid climate of Nanchang. The threshold of 40.6 °C represents the critical value at which the NWS issues a heat advisory in response to the elevated risk of severe health conditions, including heat cramps, heat exhaustion, and heat stroke [45,47]. Consequently, this study defined a heatwave event as at least five consecutive days with daily maximum HI exceeding 40.6 °C. The formula for calculating the HI is provided below.

$$\mathrm{H}\mathrm{I}={\mathrm{C}}_1+{\mathrm{C}}_2\mathrm{T}+{\mathrm{C}}_3\mathrm{R}+{\mathrm{C}}_4\mathrm{T}\mathrm{R}+{\mathrm{C}}_5{\mathrm{T}}^2+{\mathrm{C}}_6{\mathrm{R}}^2+{\mathrm{C}}_7{\mathrm{T}}^2\mathrm{R}+{\mathrm{C}}_8\mathrm{T}{\mathrm{R}}^2+{\mathrm{C}}_9{\mathrm{T}}^2{\mathrm{R}}^2$$

(1)

where T is air temperature (°C); R indicates relative humidity (%).

• ${\mathrm{C}}_1$ = −8.784695, ${\mathrm{C}}_2$ = 1.61139411, ${\mathrm{C}}_3$ = 2.338549, ${\mathrm{C}}_4$ = −0.14611605, ${\mathrm{C}}_5$ = −1.2308094 × 10⁻²,
• ${\mathrm{C}}_6$ = −1.6424828 × 10⁻², ${\mathrm{C}}_7$ = 2.211732 × 10⁻³, ${\mathrm{C}}_8$ = 7.2546 × 10⁻⁴, ${\mathrm{C}}_9$ = −3.582 × 10⁻⁶.

As shown in Figure 2, a heatwave event occurred in Nanchang in early August 2024. In late July, due to the influence of a typhoon, Nanchang experienced a cooling and precipitation process. In early August, under the influence of the subtropical anticyclone, Nanchang underwent a heatwave with a significant increase in the HI, which exceeded the threshold of 40.6 °C for most of that time. In mid-August, as the subtropical anticyclone weakened, the HI decreased, and the heatwave subsided. For comparative research, we selected two complete weeks—from 3–9 August 2024 as the heatwave week (highlighted in pink in the figure), and from 13–19 August 2024 as the normal weather week (highlighted in blue in the figure). The reasons for selecting these two weeks were as follows. Firstly, their proximity in time minimized the impact of other factors on population distribution. Secondly, there were no festivals, typhoons, or large-scale rainfall during these two weeks. Thirdly, the temperature difference between these two weeks was relatively significant.

2.3.2. Measuring Population Changes During Heatwaves (PCDH) at the Grid Scale

In order to quantify the impacts of heatwaves on population distribution, this study calculated the relative population changes during the heatwave week compared to the baseline week for each grid section due to spatial variations in population across grids within the study area. This approach had been widely adopted in studies examining the impacts of weather on public transport ridership [48]. We further extended this methodology to investigate the impacts of heatwaves on urban population distribution at the 800 m × 800 m grid scale. The normal weather week (13–19 August) was selected as the baseline week. The population changes during heatwaves (PCDH) for each grid were calculated using the following formula:

$$\mathrm{P}\mathrm{C}\mathrm{D}\mathrm{H}=({\mathrm{P}}_{\mathrm{h}}-{\mathrm{P}}_{\mathrm{b}})/{\mathrm{P}}_{\mathrm{b}}\ast 100\%$$

(2)

where ${\mathrm{P}}_{\mathrm{h}}$ is the population during the heatwave week (3–9 August), and ${\mathrm{P}}_{\mathrm{b}}$ is the population during the baseline week. When PCDH is greater than 0, it indicates an increase in population within the grid during heatwaves; conversely, a negative PCDH signifies a decrease in population. The PCDH is positively correlated with the grid’s attractiveness to people during heatwaves.

Given that people’s activity patterns and urban population distributions differed between weekdays and weekends, as well as between daytime and nighttime, we systematically analyzed four distinct time periods: weekday daytime, weekday nighttime, weekend daytime, and weekend nighttime. For each time period, we calculated the PCDH. Specifically, the sampling periods for the weekday daytime dataset were from 08:00 to 18:00 on weekdays (Monday to Friday). For the weekday nighttime dataset, the sampling periods were from 19:00 to 23:00 on weekdays. The weekend daytime dataset covered the period from 08:00 to 18:00 on weekends (Saturday and Sunday), while the weekend nighttime dataset included data from 19:00 to 23:00 on weekends. Data from 00:00 to 07:00 were excluded because of the low level of population activity during this period. Additionally, all data points with precipitation above zero were removed from each dataset to avoid potential biases introduced by precipitation events. For each of the eight datasets (comprising the heatwave week and the normal weather week across the four time periods), we calculated the average population count by summing the population counts at each sampling time point and dividing by the total number of sampling time points. This average served as the basic metric (${\mathrm{P}}_{\mathrm{h}}$ and ${\mathrm{P}}_{\mathrm{b}}$) for calculating the PCDH. Finally, grids with population below 150 were excluded in the final analysis to mitigate potential biases caused by anomalous fluctuations in low-population grids.

2.3.3. Selecting Built Environment Factors

To further investigate the relationships between PCDH for each grid and its built environment characteristics, we selected 10 indicators based on four dimensions: land use, location, transportation, and building form (

Table 1. Built environment factors selected in this study.

Dimensions	Indicators	Description	References
Land use	Proportion of residential land	Proportion of residential land in the grid section	Zhao et al., 2017 [23] Li et al., 2019 [29] Yang et al., 2023 [30] Chun et al., 2017 [49]
	Proportion of commercial and business land	Proportion of commercial and business land in the grid section
	Proportion of green space	Proportion of green space in the grid section
	Proportion of industrial land	Proportion of industrial land in the grid section
	Proportion of water body	Proportion of water body in the grid section
Location	Distance from the city center	Distance from the center of the grid section to the city center	Gu et al., 2024 [6] Yang et al., 2023 [30]
Transportation	Distance to the subway station	Distance from the center of the grid section to the nearest subway station	Wu et al., 2020 [18] Jiang et al., 2023 [19] Yang et al., 2023 [30] Long et al., 2023 [31]
Building form	Building density	The ratio of total building coverage area to the area of the grid section	Zhao et al., 2024 [25] Lan et al., 2017 [26] Yang et al., 2023 [30] Long et al., 2023 [31]
	Standard deviation of building height	Variation degree of the building height within the grid section	Lin et al., 2023 [24] Li et al., 2022 [51]
	Building height	Average building height within the grid section	Liu et al., 2024 [22] Lin et al., 2023 [24] Xu et al., 2024 [50] Li et al., 2022 [51]

), which are related to urban thermal environment and population distribution according to previous studies. For land use, we calculated the proportion of residential land, commercial and business land, green space, industrial land, and water body for each grid [23,29,30,49]. For location, we measured the distance from each grid’s center to the city center [6,30]. Additionally, the distance from each grid’s center to its nearest subway station was used as the indicator of transportation [18,19,30,31]. For building form, building density, standard deviation of building height, and building height were used to derive the characteristics of buildings in the grids [22,24,25,26,30,31,50,51].

2.3.4. Model for Investigating the Relationships Between PCDH and Built Environment Factors

(1)

Model selection

We developed four models to identify the association between PCDH and built environment factors for four distinct time periods: weekday daytime (Monday to Friday, 08:00–18:00), weekday nighttime (Monday to Friday, 19:00–23:00), weekend daytime (Saturday and Sunday, 08:00–18:00), and weekend nighttime (Saturday and Sunday, 19:00–23:00). In these models, PCDH served as the target variable, while built environment factors were used as the feature variables.

For all the four time periods listed above, we compared the performance of the linear OLS regression model and two machine learning models, including Random Forest (RF) and XGBoost. For the two machine learning models, the grids with attributes of PCDH and built environment factors were flattened and randomly divided into training and validation datasets, with 80% allocated for training and 20% for validation. The training and validation datasets were divided independently for each model corresponding to different time periods. The parameters of the two machine learning models were optimized by using a grid-search method with five-fold cross-validation. As shown in

Table 2. The evaluation metrics of the OLS, RF, and XGBoost models.

	Models	RMSE	MAE	R2
Weekday daytime	OLS	0.091	0.065	0.290
	RF	0.074	0.045	0.474
	XGBoost	0.055	0.035	0.729
Weekday nighttime	OLS	0.092	0.062	0.321
	RF	0.096	0.062	0.589
	XGBoost	0.048	0.032	0.845
Weekend daytime	OLS	0.082	0.060	0.248
	RF	0.085	0.062	0.448
	XGBoost	0.053	0.036	0.679
Weekend nighttime	OLS	0.091	0.063	0.305
	RF	0.086	0.055	0.522
	XGBoost	0.060	0.038	0.785

, across all four time periods, the XGBoost model exhibited the smallest values of Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), as well as the highest value of coefficient of determination (R²), indicating its superior performance compared to the other two models. Therefore, we employed the XGBoost model to identify the nonlinear relationships between PCDH and built environment factors.

(2)

SHapley Additive exPlanation (SHAP) method

SHAP is a technique based on Shapley values from cooperative game theory that interprets the outputs of machine learning models [52]. Although the XGBoost model is considered a black-box model, SHAP enables the visualization and analysis of its results, including the nonlinear effects of various features. In this study, SHAP was employed to interpret the prediction results of the XGBoost model, allowing for the identification of each built environment feature’s contribution to PCDH. It assigns a SHAP value to each feature of each predicted PCDH. The SHAP value reflects the strength and direction of influence of each sample’s attributes on the PCDH. The formula is as follows:

$${\varphi }_{\mathrm{i}}=\sum _{\mathrm{S}\in \mathrm{F}\backslash \left\{\mathrm{i}\right\}}{\displaystyle \frac{\left|\mathrm{S}\right|!\left(\left|\mathrm{F}\right|-\left|\mathrm{S}\right|-1\right)}{\left|\mathrm{F}\right|!}}\left[{\mathrm{f}}_{\mathrm{S}\cup \left\{\mathrm{i}\right\}}\left({\mathrm{x}}_{\mathrm{S}\cup \left\{\mathrm{i}\right\}}\right)-{\mathrm{f}}_{\mathrm{S}}\left({\mathrm{x}}_{\mathrm{S}}\right)\right]$$

(3)

where $\mathrm{i}$ is a feature, ${\varphi }_{\mathrm{i}}$ is the SHAP value of feature $\mathrm{i}$, $\mathrm{F}$ is the set of all features, $\mathrm{S}$ is the full subset of features without feature $\mathrm{i}$, $\left|\mathrm{S}\right|!$ is the factorial of the number of features contained in $\mathrm{S}$, ${\mathrm{x}}_{\mathrm{S}}$ is the input value in $\mathrm{S}$, ${\mathrm{f}}_{\mathrm{S}\cup \left\{\mathrm{i}\right\}}$ is the model trained with feature $\mathrm{i}$, and ${\mathrm{f}}_{\mathrm{S}}$ is the model trained without feature $\mathrm{i}$. If ${\varphi }_{\mathrm{i}}$ > 0, feature $\mathrm{i}$ is considered to be a positive factor, and vice versa, feature $\mathrm{i}$ is a negative factor.

3. Results

3.1. Population Distribution

Figure 3 and Figure 4 illustrated the general population distribution in the central urban area of Nanchang and its difference during various time periods. Overall, the population distribution exhibited a “single-center” pattern. The population was highly concentrated in the core area, which demonstrated the highest population density, while the urban fringe areas exhibited relatively lower population. Additionally, within these urban fringe areas, several small-scale population agglomerations emerged, including Wanli and Jiulonghu with concentrated residential functions, as well as the Economic Development Zone and High-tech Zone where manufacturing industrial functions were concentrated. Furthermore, as shown in Figure 4, there were considerable declines in population in several urban fringe areas during the nighttime period. These areas were mainly industrial zones and areas under construction.

3.2. Spatiotemporal Heterogeneity of PCDH

Figure 5 illustrated the spatial distribution of PCDH during daytime and nighttime periods on weekdays and weekends. Generally, the impacts of heatwaves on population distribution exhibited significant variation in both degree and spatial extent across different time periods. In terms of degree, heatwaves exerted the largest impacts on population distribution during weekend nights, followed by weekend daytime and weekday nighttime, with the least impacts observed during weekday daytime (

Table 3. Statistics of grid cell counts across different levels of PCDH.

	Weekday Daytime	Weekday Nighttime	Weekend Daytime	Weekend Nighttime
|PCDH| ≥ 10%	158	201	202	235
10% > |PCDH| ≥ 5%	188	215	258	304
|PCDH| < 5%	880	773	751	633
The total number of grid cells	1226	1189	1211	1172

Note: The count included incomplete grid cells at the boundaries. Due to the varying number of grid cells with a population greater than 150 at different time periods, the total number of grid cells differed among the four periods.

). Specifically, the number of grid cells with |PCDH| ≥ 10% during weekend nights was 235, markedly higher than 202 during weekend daytime and 201 during weekday nighttime. By contrast, only 158 grid cells exhibited |PCDH| ≥ 10% during weekday daytime, a number which was significantly lower compared to the other time periods.

Spatially, across all four time periods, areas with notably decreased population during heatwaves emerged in the eastern urban fringe. This area serves as a hub for manufacturing industries with its primary population consisting of manufacturing workers. During heatwaves, some factories implemented high-temperature leave policies for workers, leading to a notable reduction in the local population.

On weekdays, the impacts of heatwaves on population distribution exhibited significant diurnal differences. Concretely, in the southeastern and southwestern urban fringe areas, population decreased in daytime but increased at night during heatwaves. These areas were key zones for construction and expansion in the central urban area of Nanchang, featuring a high density of construction sites. According to relevant policies, these construction sites might suspend operations or shift work hours to nighttime to avoid daytime heat exposure during heatwaves [53]. These policies of adjusting work schedules during heatwaves contributed to significant diurnal differences in population changes within these areas. Additionally, in the core area, there was a notable population increase concentrated in the old urban areas with many old residential neighborhoods. These areas included Beijing Road, Erqi Road, Wenjiao Road, Ximazhuang, Chaoyangzhou, and Shizijie. However, in these areas, Beijing Road, Ximazhuang, and Chaoyangzhou experienced population declines at night. These old urban areas had a high resident density, with a relatively large proportion of elderly residents. During heatwaves, they tended to remain indoors during daytime to avoid intense sunlight and heat, while opting for appropriate outdoor activities at night [7,54].

On weekends, urban parks experienced significant population increases during heatwaves, with this trend being more pronounced at night than during daytime. This suggested that during heatwaves, people were more likely to choose urban parks as their weekend leisure destinations, particularly at night. Concretely, during daytime, population increases were primarily observed in urban parks such as Qianhu Park, Aixihu Wetland Park, and Xianghu-Fuhe Park. However, at night, the degree and spatial extent of population increases in urban parks further expanded, with additional increases noted in Ganjiang Citizen Park, Yaohu Wetland Park, Jiulonghu Park, and Rulehu Park. Notably, in these parks, some areas within Ganjiang Citizen Park and Yaohu Wetland Park experienced population decreases during daytime. Field surveys revealed that the lack of adequate shading facilities in these areas resulted in poor thermal comfort during daytime, leading to reduced visitor numbers. Additionally, many large commercial complexes, including the Honggutan Wanda Plaza trade area, the Mixc, Chaoyang trade area, and Jiulonghu Outlets Plaza, showed significant nighttime population increases. These large commercial complexes attracted residents for shopping and entertainment activities during heatwaves by providing comfortable indoor thermal environments through air conditioning. In summary, the population changes during heatwaves on weekends were highly related to recreation and entertainment activities.

3.3. Nonlinear Relationships Analysis Between PCDH and Built Environment Factors Based on the XGBoost Model and SHAP Method

3.3.1. Summary Plots of SHAP Values

Figure 6 presents the summary plots of SHAP values for each model, arranging factors from top to bottom based on their ranking of importance in impacting PCDH. The four decimal places on the left side of the scatter plots represent the average absolute SHAP value across samples, indicating the global importance of each factor. The results reveal that the impacts and importance of the factors manifested differently in varying time periods.

On weekdays, PCDH was mainly influenced by location, transportation, and building form factors during daytime, while the impacts of land use factors remained relatively minor. Specifically, during daytime on weekdays, location, transportation, and building form factors all ranked within the top five, while land use factors fell within the bottom five. Notably, during weekday nights, the influence of “distance to the subway station” and “proportion of residential land” significantly increased compared to the daytime. Concretely, the importance of “distance to the subway station” rose from fourth place during daytime to first place at night, indicating that nighttime leisure-related transportation modes were more vulnerable to heatwave impacts than daytime commuting transportation modes. Meanwhile, the importance of “proportion of residential land” increased from sixth place during daytime to fourth place at night. This shift aligned with the daily spatiotemporal activity patterns of most urban populations, who were primarily at workplaces during daytime and at home during nighttime [29]. The higher population density in residential areas during nighttime compared to the daytime enhanced the importance of “proportion of residential land” at night.

On weekends, the proportion of land use types related to residential and leisure activities exerted a significantly larger influence on PCDH compared to weekdays. To be specific, the importance of “proportion of residential land” increased from sixth (daytime) and fourth (nighttime) on weekdays to fourth (daytime) and first (nighttime) on weekends. Meanwhile, the importance of “proportion of green space” increased from ninth (daytime) and tenth (nighttime) on weekdays to fifth (daytime) and sixth (nighttime) on weekends. This aligns with the lifestyle patterns of urban residents, who predominantly stay at home and engage in leisure activities on weekends [28,33]. Additionally, the importance of “proportion of commercial and business land” was markedly higher at night than during daytime, showing a primarily positive correlation. This suggested that residents’ shopping and entertainment activities on weekends were more likely to occur at night during heatwaves.

In general, considering the results across all time periods, location and transportation factors exhibited high importance ranks in most periods, exerting significant influence on PCDH. In contrast, the importance of “proportion of water body” remained consistently low across all time periods. This should be attributed to the environmental characteristics of the central urban area. The limited water bodies in the central urban area yielded a relatively small sample size, resulting in the low importance of “proportion of water body”.

3.3.2. Dependency Plots and Spatial Distribution of SHAP Values

The dependence plots and spatial distribution of SHAP values for four time periods are presented in Figure 7, Figure 8, Figure 9 and Figure 10. Based on the SHAP dependence plots, we conducted a detailed analysis of the nonlinear relationships between built environment factors and PCDH. Furthermore, the spatial distribution map visualizes the SHAP values across each 800 m × 800 m grid section, revealing the varying contributions of built environment factors to PCDH across different spatial locations.

(1)

Land use

The influence of the proportion of residential land on PCDH exhibited a pronounced diurnal reversal characteristic (Figure 7a). Specifically, the proportion of residential land demonstrated predominantly positive correlations with PCDH during daytime and negative correlations at night. Furthermore, this diurnal disparity was significantly more pronounced on weekends compared to weekdays. Concretely, during weekday daytime, the influence range of the proportion of residential land was between −0.04 and 0.10, and the SHAP value exhibited an increasing trend when the proportion exceeded 20%. While during weekday nighttime, the influence range was from −0.06 to 0.08. The SHAP value dropped rapidly and turned from positive to negative when the proportion was less than 10% and remained predominantly negative in the subsequent interval. On weekends, the influencing range was notably larger than that on weekdays during corresponding time periods. During weekend daytime, the influence range expanded to a range of −0.05 to 0.125, with the SHAP value turning from negative to positive in the proportion of 15%. Conversely, during weekend evenings, the influence range widened further to −0.15 to 0.15, and the SHAP value turned from positive to negative in the proportion of 25%. These findings suggest that residents were more likely to stay indoors during hot and sunny daytime hours during heatwaves and go out at night, particularly for leisure activities on weekends.

The proportion of commercial and business land exhibited positive correlations with PCDH across all four time periods, but the spatial distribution pattern of its influence demonstrated temporal heterogeneity (Figure 7b). Specifically, during weekday daytime, its positive impacts were predominantly distributed in the core area, particularly in the old urban area. However, the clustering characteristics were relatively weaker compared to other time periods. In contrast, during weekday nighttime and weekend time periods, the spatial distribution of its positive impacts displayed pronounced clustering characteristics, primarily concentrating in large commercial complexes. Moreover, the local SHAP values were higher during these time periods than during weekday daytime, with the highest SHAP values observed during weekend nights. This suggested that large commercial complexes witnessed an increase in customer flow during heatwaves. Residents were more inclined to go to large commercial complexes with air-conditioned indoor thermal environments for cooling and engaging in shopping and entertainment activities, particularly during the peak leisure period of weekend nights [33].

The correlations between the proportion of green space and PCDH were relatively minor on weekdays but markedly intensified on weekends, with primarily positive correlations (Figure 7c). This suggests that urban green space served as critical venues for residents’ leisure activities during heatwaves on weekends. Previous studies have highlighted that urban green space, particularly wetland parks, effectively lowered temperature by plant evapotranspiration [55,56], creating a more comfortable thermal environment during heatwaves [57,58] and thus enhancing the appeal of such spaces to leisure visitors. To be specific, during weekend daytime, when the proportion of green space exceeded 40%, SHAP values were predominantly positive, demonstrating significant positive impacts on PCDH. These positive impacts were concentrated in areas such as Xianghu-Fuhe Park and Aixihu Wetland Park. Notably, during weekend nights, the positive impacts of the proportion of green space further expanded in both degree and spatial extent. The threshold for SHAP values shifting from negative to positive decreased from 40% during daytime to 15% during nighttime, with the spatial extent of positive impacts extending to areas such as Yaohu Wetland Park, Jiulonghu Park, Ganjiang Citizen Park, and Rulehu Park. This suggested that green space exhibited greater attractiveness to people at night than during daytime during heatwaves.

As shown in Figure 7d, it is noteworthy that several points with a high proportion of industrial land exhibited extremely low SHAP values during weekdays and weekend daytime, although the relationship between the proportion of industrial land and PCDH was not significant. Spatially, these points were concentrated in areas such as the Yaohu Aviation Manufacturing Industrial Zone and Fangda Steel Plant in the eastern fringe, suggesting a decline in workforce numbers at these manufacturing factories during heatwaves. This aligns with reports indicating that some factories implemented high-temperature leave policies for workers during heatwaves. Additionally, previous studies have demonstrated that heatwaves decreased working hours for employees in high-temperature exposure industries [59] and led to reduced output in manufacturing [10].

The proportion of water body showed no significant correlation with PCDH across the four time periods and exhibited extremely low importance (Figure 7e). Spatially, due to the absence of recreational functions or adequate shading facilities around certain water bodies, the thermal comfort provided by them was compromised, leading to a mosaic-like distribution of both positive and negative impacts in water bodies and their surrounding areas.

(2)

Location

As shown in Figure 8, the distance from the city center exhibited significantly negative correlations with PCDH when exceeding the threshold of 23 km across all four time periods. This suggested a marked reduction in population in urban fringe areas during heatwaves. This phenomenon was closely associated with the circular structure of the industrial zones of Nanchang. The significant negative impacts observed beyond 23 km corresponded to the manufacturing industrial zone circle, which is concentrated in the industrial zones on the eastern fringe area. The primary contributors to this population decline during heatwaves were manufacturing workers. The thermal comfort in these zones was relatively poor [23,50], and during heatwaves, some factories adopted high-temperature leave policies, leading to a decrease in the number of workers in this area.

(3)

Transportation

As shown in Figure 9, the distance to the subway station exhibited a “U”-shaped nonlinear relationship with PCDH. Notably, during weekday nights and weekends, the distance to the subway station demonstrated significant positive impacts within the range of less than 750 m, indicating marked population increases in subway stations and their surroundings. The air-conditioned and cool environment provided by subways exerted a strong appeal to people during heatwaves. Furthermore, the Nanchang Metro Group implemented summer people-oriented policies by establishing “Citizens’ Cooling Zones” in the spacious and cool underground areas of subway stations, which further attracted more citizens to seek refuge from the heat in subway stations.

(4)

Building form

The influence of building density on PCDH exhibited a significant diurnal reversal characteristic (Figure 10a). During daytime, it was predominantly positively correlated with PCDH, whereas at night, it demonstrated predominantly negative correlations. Specifically, during weekday daytime, building density was primarily positively correlated with PCDH, with positive impacts concentrated in the core area, particularly in the old urban area characterized by extremely high building density. At night on weekdays, its influence demonstrated a pronounced threshold effect, as the SHAP value shifted from positive to negative when building density reached 0.05. During weekend daytime, building density showed a positive correlation with PCDH. Conversely, during weekend nights, building density exhibited a significant negative correlation with PCDH, with its positive impacts mainly distributed in areas with lower building density in the urban fringe areas and blue–green space. This diurnal difference in the impacts on PCDH could be attributed to two factors: firstly, the positive effect of building density on local air temperature was significantly stronger at night than during daytime [25,26], reducing the attractiveness of high-density areas to people during heatwaves at night compared to the daytime; secondly, during heatwaves, people were more likely to remain indoors during daytime to avoid intense sunlight and extreme heat but go out at night [7], leading to the reversed diurnal influence of building density on PCDH.

The standard deviation of building height was predominantly positively correlated with PCDH during daytime (Figure 10b). Concretely, it demonstrated a threshold effect in both daytime periods, with the SHAP value shifting from negative to positive when the standard deviation of building height reached 7.5. This could be attributed to the improvement effect of the higher standard deviation of building height on the local thermal environment in summer [24,51]. Spatially, areas exhibiting positive influence were primarily distributed in the core area and Wanli, where the standard deviation of building height was relatively high. Negative impacts were predominantly concentrated in urban fringe areas characterized by a lower standard deviation of building height, particularly in the eastern fringe dominated by industrial zones and areas under construction. Additionally, the importance of standard deviation of building height at night decreased compared to the daytime, and its correlation with PCDH lacked significant characteristics. This was consistent with the finding in previous studies that the influence of standard deviation of building height on the thermal environment weakened at night [60].

As shown in Figure 10c, building height exhibited a predominantly positive correlation with PCDH during daytime on weekdays. Spatially, the positive impacts covered broader areas, whereas the negative impacts were predominantly concentrated in the industrial zones and areas under construction in the eastern fringe areas characterized by relatively low building height. In the other three time periods, the importance of building height markedly decreased and showed negative correlations.

4. Discussion

4.1. The Impacts of Heatwaves on Population Distribution and Their Relationships with Built Environment Factors

Our research quantitatively investigated the impacts of heatwaves on urban population distribution by measuring population changes during heatwaves (PCDH) at the grid scale. The results indicated that the heatwave impacts on population distribution differed significantly across various time periods. Heatwaves exerted the largest impacts on population distribution during weekend nights, followed by weekend daytime and weekday nighttime, with the least impacts observed during weekday daytime. This finding suggested that the temporal heterogeneity in heatwave impacts on urban population distribution aligned with the theory of spatiotemporal constraints in time geography [61]. During weekday daytime, most individuals were engaged in work-related activities, with the temporal and spatial scope of their activities constrained by work schedules. Consequently, the influence of heatwaves on their activities was relatively limited. In contrast, during the other three periods, people primarily engaged in leisure activities, which were relatively flexible and less constrained. Therefore, heatwaves significantly influenced the temporal and spatial decisions regarding their leisure activities [28,29,32,33]. The largest impact observed during weekend nights might be attributed to residents’ increased tendency to choose nighttime for leisure activities during heatwaves [7].

The analysis of the nonlinear relationships between PCDH and built environment revealed that the location and transportation factors demonstrated high importance across most time periods, serving as critical indicators influencing population changes during heatwaves. Moreover, their influences were associated with external factors such as policies. Specifically, in terms of location, the urban fringe areas, especially the eastern fringe areas, exhibited a significant negative impact on PCDH, indicating a marked population decrease in these areas during heatwaves. This finding contrasted with previous studies, which suggested that extreme high temperatures inhibited movement to the city center while promoting suburban activities, leading people to relocate from urban cores to cooler suburban areas [6]. The discrepancy could be attributed to differences in functional layout characteristics among cities. Nanchang, as one of China’s important manufacturing industrial bases, features an industrial zoning arrangement in a circular pattern. The urban fringe is predominantly characterized by the manufacturing zone, especially the eastern fringe, where numerous factories are concentrated. Workers in industrial zones face a significantly high risk of heat exposure [62,63,64] and have low productivity [10,65] during heatwaves, particularly in subtropical areas [66]. Consequently, some factories implemented high-temperature leave policies for their workers during heatwaves. This resulted in a decrease in the workforce within these factories, thereby contributing to the observed significant negative impact on PCDH in these areas.

In terms of transportation, subway stations and their surrounding areas exhibited significant positive influences on PCDH during the three non-working periods (weekday nights and weekends), indicating population increases in these areas during heatwaves. This phenomenon could be attributed to two factors: changes in transportation modes due to heatwaves and policy-related influences. On the one hand, heatwaves alter people’s transportation preferences. A previous study conducted in Beijing, Shanghai, and Shenzhen demonstrated a positive correlation between temperature and subway usage [19]. During heatwaves, the air-conditioned environment inside subways provided a cooler transportation option compared to walking or cycling, thereby enhancing their attractiveness to travelers. On the other hand, the Nanchang Metro Group implemented a people-oriented policy by establishing “Citizens’ Cooling Zones” in the spacious and cool underground areas of subway stations in summer. This policy effectively increased the attractiveness of subway stations to the general population, irrespective of transport needs. Therefore, the increase in both traveling and non-traveling populations led to positive influences on PCDH in subway stations and their surrounding areas during heatwaves.

Our research further revealed that the influences of land use and building form factors on population changes during heatwaves demonstrated significant temporal heterogeneity between daytime and nighttime, as well as between weekdays and weekends. This heterogeneity was fundamentally driven by individuals’ heat adaptation behaviors, the spatiotemporal patterns of their daily activities, and the diurnal variations in the built environment’s influence on the local thermal environment. For instance, influenced by individuals’ heat adaptation behaviors, the impacts of “proportion of residential land” and “building density” on PCDH exhibited a pronounced diurnal reversal in their correlations. These factors were positively correlated with PCDH during daytime but negatively correlated at night. This indicated that during heatwaves, residents were more likely to stay indoors or at home during daytime to avoid outdoor exposure to intense sunlight and extreme temperatures, while opting to go out at night. This finding aligned with a previous study on human mobility during extreme heat events in Houston, which reported a decrease in daytime activity levels but an increase in nighttime activity levels in extreme hot weather [7].

Additionally, the temporal heterogeneity in the impacts of building form on PCDH was influenced by the diurnal variations in their effects on local thermal environment. Specifically, the standard deviation of building height exhibited a positive correlation with PCDH during daytime, whereas the correlation became insignificant at night. This temporal heterogeneity might result from the diurnal differences in how the standard deviation of building height regulated the local thermal environment. During daytime, a higher standard deviation of building height promoted wind circulation for heat dissipation and increased the reflection of solar radiation, effectively improving the thermal environment [24,51]. Consequently, areas with a higher standard deviation of building height demonstrated strong attractiveness to people during heatwaves. At night, however, the heat radiated from the ground was blocked by buildings, and a higher standard deviation of building height acted as insulation for the ground, weakening its ability to improve the thermal environment [60,67], which led to an insignificant correlation between the standard deviation of building height and PCDH at night. Additionally, the diurnal reversal in the impacts of building density on PCDH might also stem from its influence on environmental thermal comfort besides individuals’ heat adaptation behaviors. Buildings possess high heat capacity, absorbing heat during daytime and releasing it at night [68]. This causes their warming effect on the local thermal environment to be more pronounced during nighttime than daytime [26]. Consequently, the thermal comfort in high building-density areas is reduced at night during heatwaves. This diminished the attractiveness to people of high building-density areas at night compared to daytime, thereby contributing to the diurnal variation in the impacts of building density on PCDH.

Under the influence of the spatiotemporal patterns of residents’ daily activities, the proportion of land-use types associated with recreation and entertainment functions exhibited significant spatiotemporal heterogeneity between weekdays and weekends in their impacts on PCDH. Specifically, “proportion of green space” showed a more pronounced effect on weekends than weekdays, with a positive correlation with PCDH. Furthermore, the positive impact of “proportion of commercial and business land” exhibited a pronounced clustering characteristic in the large commercial complexes during weekday nights and weekends, while this clustering characteristic was weaker during weekday daytime. These spatiotemporal heterogeneities aligned with the temporal patterns of residents’ leisure activities, which predominantly occurred on weekends, as well as weekday nights [28,29]. Previous studies have also demonstrated that urban green space and shopping centers exhibited a strong attractiveness to people during heatwaves, resulting in population increases in these areas [8,69,70]. Our study extended these findings by emphasizing the significant temporal heterogeneity between weekdays and weekends in the attractiveness to people of these areas during heatwaves, driven by the spatiotemporal patterns of residents’ daily activities. Additionally, this study also revealed that green space was more attractive to people at night than during daytime in heatwaves. This diurnal variation aligned with a previous study suggesting that some people preferred nighttime green space visits during heatwaves [8], reflecting the role of individuals’ heat adaptation behaviors in the relationship between “proportion of green space” and PCDH. Another prior study about the diurnal variation of green space cooling effects indicated that in humid–hot climate conditions such as Nanchang, green space exhibited stronger cooling effects at night than during daytime [71], enhancing nighttime thermal comfort in green spaces. This finding underscored the role of diurnal variations in green space’s thermal environment for their attractiveness during heatwaves. Overall, the spatiotemporal patterns of individuals’ daily activities, individuals’ heat adaptation behaviors, and the diurnal variations in green space’s influence on local thermal environment collectively shaped the relationship between “proportion of green space” and PCDH.

4.2. Implications for Urban Planning and Management

Based on our findings, several planning and management approaches can be proposed to proactively alleviate the adverse impacts of heatwaves. Firstly, it is recommended to promote flexible working arrangements during heatwaves, such as remote online work, to alleviate the spatiotemporal limitations of most residents during weekday daytime. This measure can reduce their exposure to extreme heat and associated health risks. Secondly, governments can develop nighttime markets and tours during heatwaves to expand nighttime leisure and consumption opportunities, aligning with the tendency of residents to engage in outdoor leisure activities at night during heatwaves. Thirdly, the research found that large commercial complexes served as critical cooling refuges for citizens during heatwaves. In urban planning and management, these areas could be designated as “urban cooling centers”, with the government providing additional public cooling facilities to mitigate health risks associated with outdoor activities. Fourthly, urban green spaces emerged as vital leisure destinations for residents on weekends during heatwaves. Therefore, in the (re)development of urban green space, more natural and artificial shading features, such as trees, lakes, ponds, and pavilions, should be incorporated to enhance their cooling capabilities. Furthermore, since green space and parks attracted more visitors at night than in daytime during heatwaves, those parks with restricted operating hours should extend their nighttime opening hours to meet citizens’ cooling and leisure needs. Fifthly, in urban fringe areas characterized by manufacturing industrial zones, comprehensive planning should prioritize the integration of residential complexes, recreational facilities, and green leisure spaces to establish industrial community hubs. This approach not only encourages workers to access diverse after-work services locally but also improves the thermal environment within these urban fringe areas. Finally, considering the increased population around subway stations during heatwaves, when selecting the locations of new subway stations, urban planners should give more consideration to the urban functions and environmental characteristics surrounding the sites, reduce the walking distance for citizens to reach the stations during heatwaves, and improve their walking thermal comfort.

5. Conclusions

Taking the central urban area of Nanchang as the case, this study investigated the impacts of heatwaves on population distribution from the perspective of temporal heterogeneity. Using the XGBoost model combined with the SHAP method, it further explored the nonlinear relationships between built environment factors and population changes during heatwaves at the grid scale. The main findings are summarized as follows:

(1) Heatwaves exerted the largest impacts on population distribution during weekend nights, followed by weekend daytime and weekday nighttime, with the least impacts observed during weekday daytime. This finding suggested that the temporal heterogeneity in heatwave impacts on urban population distribution aligned with the theory of spatiotemporal constraints in time geography.

(2) Location and transportation factors demonstrated high importance across most time periods, serving as critical indicators influencing population changes during heatwaves. Moreover, their influence was associated with external factors such as policies. For instance, in the location dimension, urban fringe areas dominated by industrial clusters experienced a notable reduction in population during heatwaves due to the policy of granting workers high-temperature leave in some factories. Furthermore, in the transportation dimension, the people-oriented policy of the Nanchang Metro Group to establish “Citizens’ Cooling Zones” in the cool underground space of air-conditioned subway stations contributed to an increase in population during heatwaves in subway station areas.

(3) The land use and building form factors, including the proportions of residential land, green space, and commercial and business land, as well as the building density and the standard deviation of building height, exhibited significant temporal heterogeneity in their influence on population changes during heatwaves. This temporal heterogeneity was fundamentally driven by individuals’ heat adaptation behaviors, the spatiotemporal patterns of their daily activities, and the diurnal variations in the built environment’s influence on the local thermal environment.

These findings are important for assessing the impacts of heatwaves on urban populations. They facilitate an in-depth analysis of how heatwaves influence population distribution in various time periods, as well as the relationships between their impacts on population distribution and the urban built environment. The findings can provide useful planning and management recommendations to proactively alleviate the adverse impacts of heatwaves.

However, this study has several limitations. Firstly, the sample size was limited to a single heatwave event in Nanchang in August 2024. Future research should broaden its scope by incorporating data from multiple cities and various years for comparative analyses, thereby enhancing the generalizability of the findings. Secondly, the spatial resolution of the population heat map data utilized in this study was relatively coarse, which might restrict the accurate detection of small-scale population movements and potentially influence the precision of the statistical analysis of population changes during heatwaves. Thirdly, factors regarding “mobility” were little touched in this study. Future research should leverage population data with higher spatiotemporal resolution and richer information to identify the specific direction and magnitude of population mobility changes during heatwaves. This would further illuminate the more detailed mechanisms underlying the impacts of heatwaves on urban population distribution. Fourthly, this study did not investigate the relationships between population changes during heatwaves and socio-economic attributes. Future research could leverage multi-source data to further examine these relationships.