Abstract

Background

The city built environment plays a crucial role in influencing population vulnerability to temperature extremes, yet population-based evidence has been limited.

Methods

We included 21,494 urban residents from a nationally representative cohort study. Temperature extremes were defined using residential address-specific thresholds lasting for ≥ 3 days. Street view images within participants’ residences (500-m radius) were evaluated using semantic segmentation by DeepLabV3 Plus-ResNet101 pretrained by the Cityscapes dataset. Cox proportional hazard models and interaction models were applied to explore the moderating effects of street view-derived built environments on the effects of temperature extremes on mortality.

Results

Each additional day of heatwave (95th) and coldspell (5th) duration per year was associated with a 6% (HR = 1.06, 1.03–1.08) and 4% (HR = 1.04, 1.01–1.08) increase in all-cause mortality risk, respectively. Lower sky view factor (SVF) and openness, and higher grayness, building coverage, and interface enclosure (IE) were associated with increased risks of mortality. Street view-derived built environment factors could modify the associations between temperature extremes and mortality. Specifically, high levels of IE, along with low levels of SVF and openness, across different road levels intensified the effects of heatwaves. Conversely, low levels of IE, along with high levels of SVF and openness, in residential roads amplified the effects of coldspells.

Conclusions

Our findings provide insights for evidence-based urban planning and public health strategies, emphasizing the need for adaptive, context-specific urban design that balances the potentially competing demands of population heat resilience and cold adaptation.

Similar content being viewed by others

Background

The increasing frequency and severity of temperature extremes, particularly heatwaves and coldspells, have become significant public health concerns worldwide [1, 2]. Extreme heat, in particular, is the leading cause of weather-related deaths, with an estimation of 236 deaths per 10 million residents, which has risen by over 50% in the past two decades [1]. Heatwaves and coldspells have also been linked to higher risks of all-cause and cardiovascular disease (CVD) events and mortality [3]. Moreover, the disease burden attributed to temperature extremes is expected to worsen in the global context of climate change [4].

Efforts to mitigate the health impacts of temperature extremes have traditionally focused on interventions such as green spaces, green roofs, and cooling centers designed to alleviate the urban heat island (UHI) effect [2, 5,6,7,8,9,10]. Meanwhile, early warning systems and local vulnerability maps have also been used in reducing temperature-related mortality [11, 12]. However, these measures often face challenges in scalability and accessibility, limiting their real-world applications and effectiveness [12, 13].

Emerging evidence highlights the critical role of the built environment (i.e., the human-made surrounding environments) in shaping the resilience of populations to temperature extremes [14,15,16]. Built environment factors such as urban greenness, building density, openness, and street layout have been found to influence microclimatic conditions and thermal comfort, which may modify the health effects of temperature extremes [17,18,19]. Nevertheless, prior evidence has primarily focused on region-level (e.g., county or tract census) isolated metrics (e.g., green spaces) [5, 9, 13], which fail to capture the surrounding built environments exposed by individuals in the real-world context. Recent advancements in artificial intelligence and geospatial technologies provide opportunities to characterize street-level built environments through accurate quantification of street view imagery [14, 20,21,22]. Combined with population-based cohort data, this approach can provide deeper insights into how specific built environment characteristics might modify individual vulnerability to temperature extremes.

In this study, we aim to fill this gap by investigating the impact of temperature extremes on mortality, particularly examining how built environment factors derived from street view imagery around participants’ residences may modify the associations. Our findings would provide evidence-based insights for urban planning and public health interventions aimed at mitigating the adverse health effects of temperature extremes.

Methods

Study design and cohort profile

The China National Survey of Chronic Kidney Disease (CNSCKD) is a nationwide study that employed a multistage stratified sampling design to ensure broad geographical coverage and representativeness across China [23]. As shown in Fig. 1, the first step in the selection process involved choosing 13 provinces (Additional file 1: Fig. S1), autonomous regions, and municipalities. The selected regions covered four major climate zones of China, including the Temperate Monsoon (Beijing and Shandong), Subtropical Monsoon (Sichuan, Jiangsu, Zhejiang, Hubei, Hunan, and Shanghai), Tropical Monsoon (Guangdong and Guangxi), and Temperate Continental (Inner Mongolia, Xinjiang, and Ningxia) climates. Within each of these regions, one district from both urban and rural regions, three subdistricts from each district, and five communities from each subdistrict were subsequently selected using simple random sampling. The participants were further chosen using simple random sampling within each community.

Fig. 1

Flow chart for the participants included in the final analyses

The inclusion criteria for baseline enrollment required individuals to be residents of the selected communities, aged 18 years or older, and to provide consent to participate in the study. This process resulted in the enrollment of 50,550 participants, whose demographic, health, and lifestyle baseline data were collected between 2007 and 2010. Trained healthcare professionals carried out comprehensive assessments, which included interviews, anthropometric measurements (e.g., weight, height, waist circumference), and biological sample collection. A detailed questionnaire was used to gather sociodemographic information, medical history, and lifestyle factors. Blood pressure was measured three times, and blood and urine samples were taken after a 10-h fast. These samples were then analyzed in laboratories following standardized protocols.

Follow-up continued through December 31, 2017, during which mortality data were tracked. The study linked participants to the China Cause of Death Reporting System to ascertain both all-cause and cause-specific mortality. Death causes were coded according to ICD-10 guidelines, which classify the primary causes into categories comprising CVDs (I00–I99), respiratory diseases (J00–J99), cancers (C00–C99), and others [24]. To ensure accurate data matching, each participant’s death was identified using their administrative personal identification, with additional verification against demographic details (name, birth, sex, and family address).

As shown in Fig. 1, participants with incorrect address information (N = 3), without kidney function data (N = 3343), and those without detailed address information or with inconsistent addresses between baseline and death records (N = 118) were excluded. Considering the limited availability of road and street view data in rural regions, we narrowed the cohort to urban-based settings. Participants without available street view images (N = 3844) were excluded, and a total of 21,494 participants were included in the final analyses.

The research protocols were approved by the Biomedical Ethics Review Committee of Peking University (approval code: IRB00001052-20030). Signed informed consent was collected before the study began. More details are provided in Additional file 1: Supplementary Methods.

Temperature extremes

We utilized the China Daily Air Temperature (CDAT) dataset [25], which provides long-term time-series daily near-surface air temperature (Ta) data at 0.1° spatial resolution across China, comprising daily maximum (Tmax), minimum (Tmin), and 24-h average (Tavg) temperatures. CDAT integrates multi-source datasets (remote sensing products and in situ measurements) and applies corrections for regional climatic discrepancies, enabling robust Ta estimation under both clear-sky and non-clear-sky conditions. Validation against ground stations demonstrated high accuracy (R2 = 0.95–1.00) [25, 26]. In line with previous studies, heatwaves and coldspells were defined using residence-specific (0.1° grid) thresholds to account for regional climate variability [27,28,29]. Heatwaves were defined as ≥ 3 consecutive days during April ~ September with Tmax exceeding the residence-specific warm-season percentile thresholds (80th, 85th, 90th, or 95th) [27]. Coldspells were defined as ≥ 3 consecutive days during October ~ March with Tmin below the residence-specific cold-season percentile thresholds (20th, 15th, 10th, or 5th) [28]. To align temperature threshold definitions with the cohort period, residence-specific percentile thresholds for heatwaves and coldspells were computed using the 2007–2017 (cohort baseline and follow-up window) subset of CDAT. Durations of heatwave and coldspell events were calculated by summing these days.

Street view-based built environment factors

The schematic diagram of street view image processing is shown in Fig. 2. All road networks (primary, secondary, tertiary, residential, and others) within a 500-m radius of participants’ residences were extracted from OpenStreetMap [30, 31]. This distance has been widely adopted in previous environmental epidemiology and urban health studies as it approximates a reasonable walking distance and represents the scale of individual exposure [32, 33]. Secondary, tertiary, and residential roads were also separately analyzed, as these road levels may be associated with daily activities and environmental exposures of the participants. Sampling points were generated at 100-m intervals along selected streets using geographic coordinates (latitude/longitude). This ensured systematic coverage of streetscapes relevant to participants’ residences. At each sampling point, four-directional (0°, 90°, 180°, 270°) street view images were retrieved via Baidu Map API (the street view images were not available before 2013, and thus we collected historical images from 2013 to 2017). A total of 45,804 images were processed using a semantic segmentation model DeepLabV3 Plus-ResNet101, pretrained on the Cityscapes dataset (19-class pixel annotations, mean intersection over union [mIoU] = 0.762) [34, 35]. Six pixel-proportion-based indicators were further calculated for all (comprising primary, secondary, tertiary, residential, and others), secondary, tertiary, and residential roads [19, 36,37,38], comprising:

1. (a)

Greenness: Defined as the proportion of pixels classified as vegetation and terrain. This metric represents the perceived density of natural elements. This metric can complement satellite-based vegetation indices by capturing greenness directly perceivable at the pedestrian level [32].

1. (b)

Sky view factor (SVF): Defined as the proportion of pixels classified as sky. This metric indicates the degree of vertical openness. The image-derived SVF can capture the actual visible sky from a specific viewpoint compared to the theoretical geometric SVF calculated by the 3D model [19].

1. (c)

Grayness: Defined as the proportion of pixels classified as built structures, including buildings, walls, fences, poles, traffic lights, and traffic signs. This metric indicates urbanization intensity and human-made dominance over the landscape [18, 36].

1. (d)

Building coverage: Defined as the proportion of pixels classified as buildings and walls. This metric indicates the spatial compactness of building footprints and vertical structures. Compared to Geographic Information System (GIS)-based metrics, the grayness and building coverage can capture the vertical presence of a wider range of structural elements [36].

1. (e)

Openness: Defined as the proportion of pixels classified as terrain, sky, roads, and sidewalks. The openness indicates unobstructed horizontal areas that facilitate physical or environmental circulation at the pedestrian level.

1. (f)

Interface enclosure (IE): Defined as the proportion of pixels classified as buildings, walls, fences, poles, and vegetation that enclose the space creating physical boundaries. High IE values suggest strong boundary enclosure and environmental blockage, which may influence microclimatic conditions like wind flow [19, 38]. More details are provided in Additional file 1: Supplementary Methods.

Fig. 2

Schematic diagram of street view image processing. A Streets of various levels within 500 m of the residential address of the participants; B latitude and longitude sampling of the roads at 100-m intervals; C four-direction street view image recognition using the DeepLabV3 Plus-ResNet101 pretrained on the Cityscapes dataset; D typical street views for each level of road; E typical street views with high levels of built environment factors of interests

Statistical analyses

Cox proportional hazards regression models were employed to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for mortality risks associated with temperature extremes and built environment factors based on the time-to-event data [39]. The Cox proportional hazards model is given by:

$$h(t|X)={h}_{0}(t)\cdot \text{exp}({\beta }_{1}{X}_{1}+{\beta }_{2}{X}_{2}+\cdots +{\beta }_{k}{X}_{k}+{\beta }_{X}X)$$

where (h(t|X)) is the hazard of mortality at time (t) for an individual with covariate values. The covariates included demographic and socio-economic variables, lifestyle factors, health-related conditions, and annual average temperature and relative humidity, with detailed settings described in Additional file 1: Supplementary Methods. Missing data in categorical covariates were handled by creating a separate missing category, which was included in the models as a dummy variable. No numerical variables had missing data. ({h}_{0}(t)) is the baseline hazard at time (t), and ({\beta }_{1},{\beta }_{2},...,{\beta }_{k}) are the regression coefficients associated with the covariates ({X}_{1},{X}_{2},...,{X}_{k}). The key exposure variables of interest ((X)) are annual numbers of heatwave/coldspell days and street view-derived built environment factors. To account for the dynamic nature of environmental changes over the study period (2007–2017), temperature extremes were treated as a time-varying covariate in the Cox models [39, 40]. No violation of the proportional hazard (PH) assumption was detected, indicating the absence of time-dependent effects on hazard ratio estimates. Multicollinearity diagnostics showed that all variance inflation factors (VIFs) were below 5 across models, suggesting no multicollinearity among the covariates.

To examine the moderating effects of street view-derived built environment factors on associations between temperature extremes and mortality, their main effect terms and multiplicative interaction terms were included in the models [41]. Built environment factors were divided into two groups (Lower and Higher) based on the median levels. The interaction model is specified as:

$$h(t|X,B)={h}_{0}(t)\cdot \text{exp}({\beta }_{1}{X}_{1}+\cdots +{\beta }_{k}{X}_{k}+{\beta }_{X}X+{\beta }_{B}B+{\delta }_{XB}(X\cdot B))$$

where (X\cdot B) is the interaction term, ({\delta }_{XB}) is the coefficient of the interaction term, and (B) is one of Lower or Higher. The subgroup coefficients of (X) were calculated to be ({\beta }_{X}) ((B=\text{Lower})) and ({\beta }_{X}+{\delta }_{XB}) ((B=\text{Higher})). The statistical significance of the ({\delta }_{XB}) was examined [41].

Sensitivity analyses were conducted to assess robustness. First, non-linear exposure–response associations between temperature extremes and mortality were explored using restricted cubic splines with three knots [42]. Second, two-pollutant models adjusted for annual averages of fine particulate matter (PM2.5), ozone (O3), and nitrogen dioxide (NO2) to address residual air pollution confounding [43,44,45]. We also conducted a three-pollutant model including PM2.5 and NO2 to assess co-pollutant confounding common in urban contexts [46]. Third, directed acyclic graphs (DAGs) were used to identify a minimally sufficient adjustment set, ensuring covariates were selected based on causal pathways [47]. Fourth, built environment factors were reclassified into three categories (Low, Medium, and High, based on tertiles) to test the stability of the moderating effect examinations. Fifth, to account for potential regional variations, we further included a random intercept for province or climate zone in the models. Sixth, considering the opposite roles of greenness and built structures in cooling effects [38], we also calculated IE_no_vegetation by excluding vegetation and repeated the main and interaction analyses.

The data analyses were performed using Python (version 3.12), ArcGIS (version 10.8), and R (version 4.3.2). The statistical significance level was set as two-sided α = 0.05. Anonymous data were analyzed strictly following the individual privacy and ethical guidelines.

Role of the funding source

The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

Results

Baseline characteristics

Among 25,338 urban participants, street view images were available for 21,494 individuals (84.8%). Compared with those with images, participants without images were slightly younger, more likely to be females, and had lower income and education levels (Additional file 1: Table S1). Nevertheless, the absolute differences between the original cohort participants and those with images were below 2% for all subgroups. These results suggest that although selection bias due to missing street view data cannot be excluded, the final sample was generally representative of the original cohort. A total of 21,494 urban residents were included in the final analyses, with 20,699 in the non-death group and 795 in the death group (Table 1). The leading cause of death was cancer (N = 328, 41.3%), followed by CVDs (N = 267, 33.6%) and respiratory diseases (N = 57, 7.2%). The average age of participants was 52.2 ± 14.4 years, with the death group being significantly older (66.0 ± 11.8 years vs 51.7 ± 14.2 years, P < 0.001). Participants in the death group also had lower income and education levels and higher rates of current smoking and alcohol consumption ≥ 3 times per week. Chronic disease conditions comprising HTN, diabetes, and CKD were more prevalent in the death group.

Exposure levels

Additional file 1: Table S2 presents the distribution of exposure levels of temperature extremes and built environment factors among participants. In general, the death group was exposed to longer durations of heatwaves and coldspells at various thresholds. Meanwhile, the death group was exposed to lower levels of SVF and openness, but higher levels of grayness, building coverage, and IE across all roads. Similar patterns in SVF, grayness, building coverage, openness, and IE were observed for secondary, tertiary, and residential roads. Greenness in secondary roads was higher in the death group, whereas no significant differences in greenness were found between the non-death and death groups for all roads, or for tertiary and residential roads. Spearman correlations between built environment factors are shown in Additional file 1: Table S3. Greenness was consistently negatively correlated with grayness and building coverage (ρ: − 0.32 to − 0.48) and positively correlated with IE (ρ: 0.10 to 0.71). SVF and openness showed strong positive associations (ρ: 0.92 to 0.96) and were strongly negatively correlated with IE (ρ: − 0.96 to − 0.93). Grayness and building coverage were strongly positively correlated (ρ: 0.97 to 0.99), and they were also positively correlated with IE (ρ: 0.32 to 0.76).

Associations between temperature extremes and mortality

Heatwaves and coldspells exhibited threshold-dependent associations with increased risks of all-cause and cause-specific mortality, with stronger associations for more extreme thresholds (Fig. 3). For instance, each additional day of heatwave duration at the 80th, 85th, 90th, and 95th percentile threshold was associated with a 2% (HR: 1.02, 1.01–1.03), 3% (HR: 1.03, 1.02–1.05), 4% (HR: 1.04, 1.02–1.06), and 6% (HR: 1.06, 1.03–1.08) higher risk of all-cause mortality, respectively. Each additional day of coldspell duration at 20th, 15th, 10th, and 5th percentile threshold was associated with a 2% (HR: 1.02, 1.00–1.03), 2% (HR: 1.02, 1.01–1.04), 2% (HR: 1.02, 1.00–1.04), and 4% (HR: 1.04, 1.01–1.08) higher risk of all-cause mortality, respectively. Heatwaves at the 95th percentile threshold were associated with increased risks of respiratory disease (HR: 1.12, 1.02–1.24) and cancer (HR: 1.05, 1.01–1.09) mortality. In contrast, coldspells at the 5th percentile threshold were associated with increased risks of CVD (HR: 1.11, 1.04–1.18), but not respiratory disease or cancer mortality.

Fig. 3

Association of temperature extremes with all-cause and cause-specific mortality. A Heatwaves; B coldspells. Note: Adjusted hazard ratio with 95% confidence intervals associated with 1 day increase in durations of temperature extremes per year. Models were adjusted for age, body mass index, sex, income, education, diet, physical activity, alcohol consumption, tobacco use, history of hypertension, diabetes mellitus, chronic kidney disease, annual average temperature, and relative humidity

Associations between street view-derived built environment factors and mortality

Street view-derived greenness showed distinct patterns across different road levels (Fig. 4). Per standard deviation (SD) increase in greenness was associated with reduced all-cause mortality in all roads (HR: 0.89, 0.83–0.95) and tertiary roads (HR: 0.90, 0.83–0.98), but no significant associations were found for secondary or residential roads. In contrast, other built environment factors generally showed similar patterns. For instance, per SD increment in SVF, grayness, building coverage, openness, and IE in secondary roads was associated with an 18% decrease (HR: 0.82, 0.74–0.91), 14% increase (HR: 1.14, 1.06–1.24), 15% increase (HR: 1.15, 1.07–1.25), 18% decrease (HR: 0.82, 0.74–0.91), and 21% increase (HR: 1.21, 1.10–1.34) in all-cause mortality, respectively.

Fig. 4

Association of level-specific street view-based built environment factors with all-cause and cause-specific mortality. Note: Adjusted hazard ratio with 95% confidence intervals associated with one standard deviation (SD) increase in factors. Models were adjusted for age, body mass index, sex, income, education, diet, physical activity, alcohol consumption, tobacco use, history of hypertension, diabetes mellitus, and chronic kidney disease

Street view-derived built environment factors and vulnerability to temperature extremes

As shown in Fig. 5A, high building coverage and IE levels in tertiary and residential roads consistently intensified the effects of heatwaves on all-cause mortality, while high SVF and openness levels across different road levels consistently mitigated the effects. For instance, in residential roads, high SVF (P for interaction < 0.001) and openness (P for interaction < 0.001) could mitigate the effects of heatwaves on all-cause mortality, while high building coverage (P for interaction = 0.002) and IE (P for interaction < 0.001) showed opposite directions.

Fig. 5

Moderating effects of level-specific street view-based built environment factors on associations between temperature extremes and all-cause mortality. A Heatwaves; B coldspells. Note: Adjusted hazard ratio with 95% confidence intervals associated with 1 day increase in durations of temperature extremes per year in groups with lower and higher (based on the median level) built environment factors. In addition to covariates, models included the interaction terms of the temperature extremes and built environment factors. P inter, the statistical significance of the interaction terms (P for interaction)

In contrast to the heatwave findings, the moderating effects of built environment factors on coldspells showed opposite directions, especially at the residential road level (Fig. 5B). Specifically, participants in the high IE group had a mitigated effect of coldspells on mortality compared to those in the low IE group (P for interaction < 0.001), while high SVF (P for interaction = 0.012) and openness (P for interaction < 0.001) intensified the effects.

For cause-specific mortality, the results for CVD and cancer mortality were generally consistent with those for all-cause mortality, whereas built environment factors did not significantly modify the effects of temperature extremes on respiratory disease mortality (Additional file 1: Figs. S2–S7).

Sensitivity analyses

Exposure–response curves showed consistently similar associations between temperature extremes and mortality, as shown in Additional file 1: Fig. S8. Two- and multiple-pollutant sensitivity analyses reported robust findings when additionally adjusted for PM2.5, O3, and NO2 (Additional file 1: Table S4). The DAG identified age, sex, income, education, temperature, and relative humidity as the minimally sufficient adjustment set (Additional file 1: Fig. S9), and the result remained robust when including these variables as covariates only (Additional file 1: Table S5). Moderating effects of street view-based built environment factors based on three-category-sensitivity analyses also highlighted similar moderating effects of SVF, openness, and IE, as shown in Additional file 1: Fig. S10. After including province or climate zone as a random effect, the associations were slightly amplified for province and similar for climate zone, with statistical significance remaining robust in both cases (Additional file 1: Table S6). After recalculating IE with vegetation pixels excluded, effect estimates varied slightly, and the overall patterns of the main associations and the effect modification remained consistent (Additional file 1: Tables S7 and S8).

Discussion

Real-world evidence on how micro-scale urban design interacts with temperature extremes can offer actionable insights for mitigating the health impacts of temperature extremes in the global context of climate change. Our population-based study found that heatwaves and coldspells are associated with increased risks of all-cause, CVD, and cancer mortality. Beyond the direct effects of temperature extremes, we also observed that street view-derived built environment characteristics modify these associations. Specifically, high levels of IE, along with low levels of SVF and openness, across different levels of roads consistently intensified the effects of heatwaves on mortality. Conversely, for coldspells, low levels of IE, along with high levels of SVF and openness, in residential roads amplified the mortality risks.

Our study reveals threshold-dependent associations between temperature extremes and mortality, with stronger effects observed at more extreme temperature thresholds, which are consistent with previous evidence across countries [28, 48,49,50]. A multi-country, multi-community study across 400 communities in 18 countries reported stronger associations between heatwaves and mortality when more extreme temperature thresholds were applied [27]. A study among deaths in 31 Chinese capital cities also reported higher risks of all-cause and cause-specific mortality associated with coldspells defined by extreme percentile thresholds compared to fixed thresholds [28]. A study among 151,001 deaths aged ≥ 65 years in São Paulo also found stronger effects of temperature extremes on CVD mortality at more extreme temperature thresholds with longe