Introduction

In the past decades, epidemic diseases have become the most serious threat to global urban sustainability by affecting physical and mental health, and other domains of life (Mouratidis 2022). The prevention and control measures of these diseases, such as lockdowns, quarantines, social isolation and the closure of businesses and public facilities also caused an indirect impact on public health and human well-being in cities. Dengue fever (DF) is a mosquito-borne disease caused by the dengue virus, which threatens more than half of the world’s population (World Health Organization 2023a). From 2000 to 2019, the World Health Organization documented a ten-fold surge in reported DF cases worldwide increasing from 500, 000 to 5.2 million spreading across 129 countries (World Health Organization 2023b). Female mosquitoes, the viral vector, carried the DF virus when they biting an infected person in the viremia phase (Soghaier et al. 2015). The virus typically infects mosquito salivary glands after 5–12 days, after which viral transmission to another health person can occur during blood feeding. Currently, no vaccines or antiviral drugs are available to prevent DF infection (Wilder-Smith 2020); the main strategies for prevention are epidemic risk assessment, early public prevention and control actions, and personal protection measures that rely on knowledge of the driving process of DF transmission. Due to the immense global devastation caused by DF, understanding the drivers of DF epidemics and effective prevention and control measures could contribute to working towards the good health and well-being for the Sustainable Development Goals (Boischio et al. 2009; Yoshikawa et al. 2020).

Urban socioeconomic conditions play an important role in mosquito-borne epidemics. Urbanization in the Peruvian Amazon drove the DF spreading by human transportation networks along highways and rivers (Guagliardo et al. 2014). Variations in wealth and poverty, usually assessed in terms of Gross Domestic Production (GDP) and property prices were closely related to the degree of public health (Degarege et al. 2019) and linked to the risk of infectious diseases (Chen et al. 2019). In cities, higher population density (Lin and Wen 2011) and ease of transportation (Cao et al. 2017; Lefebvre et al. 2022) could also increase the potential risk of infectious diseases when compared to rural areas.

Pavlovsky (1966) introduced the concept of "landscape epidemiology" and proposed that microscopic pathogens are determined by landscape patterns and dynamics (Ig 1975). In the context of this conceptual framework, Reisen (2010) proposed that landscape patterns could have a significant impact on vector-borne disease emergence. In cities, landscape features often have strong associations with the distribution and abundance of Aedes aegypti and Aedes albopictus, the dengue virus vectors (Landau and Leeuwen 2012); particularly as green spaces and water bodies are potential mosquito habitats and environmental risk factors (Ye and Qiu 2021). In contrast, a study in Indonesia reported that the risk of DF decreases with increasing forest cover (Husnina et al. 2019). In traditional epidemic areas, such as Yunnan and Myanmar, a larger proportion of water body may increase the risk of DF epidemics (Zheng et al. 2019).

However, in addition to vector and host habitats as risk indicators for mosquito-borne infections, the important infecting process of human exposure to vectors (human-mosquito contact) has been neglected (Thongsripong et al. 2021). The average flight distance for adult mosquito is approximately 100 m (Gubler et al. 2010), but the range of human mobility in large cities usually far exceeds that of vectors. Furthermore, the effect of human mobility on the transmission of mosquito-borne infections was just brought to the attention of epidemiologists (Guzzetta et al. 2020). Regions with large population densities have been used in some studies as proxies for population aggregation sites (Padmanabha et al. 2012), but the spatial and temporal variations in host activities have been ignored. Population movement may transport viruses from high-density vector areas to low-density vector areas (Lefebvre et al. 2022). Therefore, social behavior represented by population aggregation and mobility, is a more relevant factor. In the case of a DF outbreak in central Tokyo in 2014, almost all infected people visited the same park during the epidemic (Kutsuna et al. 2015). In Taiwan, urban SGPs influence DF transmission because social activity sites where vectors are frequently present should be incorporated into DF risk intervention programs (Wen et al. 2015). In cities, social behavioral habits are influenced by socio-spatial structure (Li et al. 2014). Similar studies have assessed the effects of the DF epidemic on SGPs, such as traditional markets, parks, and shopping malls, as a proxy for human social behavior (Favier et al. 2005), to avoid violation of the privacy of infected subjects.

In the most arbovirus epidemic region, there was high frequency of non-infected areas (zero cell) which will lead the bias in classical statistical models. Zero-inflated poisson regression (ZIPR) model was developed for the count data with excess zeros (Lambert 1992). Spatial heterogeneity and spatial autocorrelation analyses are common theoretical approaches for understanding the dominant landscape processes influencing infectious disease epidemics. The spatial heterogeneity of DF was studied in depth after correcting for the heterogeneity of the DF epidemics on Easter Island, and it was found that the epidemic model could be optimized (Favier et al. 2005). However, most studies of the relationship between landscape patterns and infectious disease epidemics applying global regression approaches, such as the generalize linear model cannot accurately distinguish the local spatial heterogeneity. This meant that it was not possible to use these findings directly to guide the prevention and control measures within the urban region. In this study, we applied the geographically weighted zero-inflated poisson regression (GWZIPR), a spatial modeling technique that includes spatial information to reveal spatial heterogeneity between the independent and dependent variables, considering both count and zero data (Kalogirou 2016).

Hangzhou was not traditional endemic to DF outbreaks, and the sporadic cases that do occur were mainly imported (Lai et al. 2015). However, in the summer of 2017, Hangzhou experienced its first DF outbreak in which 1151 autochthonous DF cases were reported in Hangzhou’s central urban area (Yan et al. 2018). Autochthonous cases of DF were initially reported centrally, likely due to its high population density and increased mobility and then quickly spread to neighboring areas and suburbs. Specifically, the indicated cases were linked to transmission via SGPs, with 14 infected individuals visiting the same fitness club (Yu et al. 2019).

Our study investigated the spatial pattern of DF cases in Hangzhou and analyzed the effects of land use and land cover (LULC) patterns, socioeconomic conditions, and distribution of SGPs on DF epidemics to enhance the understanding of the epidemic process of DF in urban landscape. We developed a new approach to detect the spatial relations between urban landscape pattern and DF epidemic using spatial autocorrelation analysis, GWZIPR and penalized regression algorithm. Our study also tried to assess the effects of the DF epidemic on SGPs as a proxy for human social behavior to avoid violation of the privacy of infected subjects within the urban area and provide targeted guidance for early warning and effective landscape governance in other new epidemic areas of DF in the future.

Methods

Study area

Hangzhou (118–120°E, 29–30°N) is the capital city of Zhejiang Province in eastern China. It is a core city in the Yangtze River Delta and one of the most developed regions in China. Hangzhou has a subtropical monsoon climate with an average annual temperature of 17.8℃ and annual precipitation of 1454 mm. The mosquito population usually peaks in August-October as a result of the constant high temperatures and abundant rainfall in the summer (Tian et al. 2016). Geographical variation of land use in Hangzhou City are obvious. In the eastern region, massive built-up areas occupy a larger proportion of the land and have a higher population density, while mountains and vegetated areas occupy a larger proportion of the western region (Ma, 2016). As DF cases were mainly distributed in the main urban area of Hangzhou, the study area was delineated within the main urban area, including Shangcheng District, Xiacheng District, Gongshu District, Jianggan District, Xihu District, Binjiang District, Yuhang District, and Xiaoshan District (Fig. 1).

Fig. 1
figure 1

Geographical locations of DF cases in Hangzhou at 2017

Full size image

Data collection

Hangzhou DF case characteristics

Basic data of DF cases observed in Hangzhou in 2017 was obtained for the analysis from the Hangzhou Center for Disease Control and Prevention, including the spatial location and type of case (autochthonous or imported). As the infected person may have been bitten in several places, it was difficult to determine the probable location of exposure. Here, we assumed that the daily movement of these infected individuals was centered on their residential address, and applied the residential address as the unique confirmed spatial location of the DF cases. Additional DF data in 2018 and 2019 were applied to verify the accuracy of the prediction model. Basic description of DF cases’ characteristics was list in Table S1. And all data were desensitized of personally identifiable information. The scale of the spatial analysis was determined based on the semi-variogram of the DF cases with a 2.1 km interval distance (Fig. S1). Based on distance, we divided the main urban area of Hangzhou into 2 km × 2 km grids for the following analysis. A previous study also showed that 2 km was the best spatial scale for understanding the spatial heterogeneity of the DF in Guangdong Province (Ren et al. 2017). The number of autochthonous DF cases in a spatial grid cell was used as an indicator of the potential spatial risk of DF in Hangzhou. According to the occurrences of DF cases, all the cells in the studied region were classified into count cells and zero cells.

Urban landscape pattern

In this study, urban landscape pattern was composed of LULC pattern and SGPs. LULC pattern reflected the plane pattern of landscape elements, while SGPs reflected the point pattern of public places in urban area.

Raw LULC data with 10 m resolution were obtained from the Finer Resolution Observation and Monitoring Global Land Cover database (http://data.ess.tsinghua.edu.cn/fromglc10_2017v01.html). The raw data were then reclassified to green space, water body, and built-up areas(Gao et al. 2021). We calculated two class-level landscape indices, the total class area (CA) and shape index (SI), for the three LULC types (Table 1) to represent the LULC patterns using Fragstats 4.2.

Table 1 Detailed description of the variables
Full size table

The POI data of the SGPs were obtained from the Gaode Map open platform (https://lbs.amap.com/) with spatial information. The SGPs included traditional markets (247 items), malls (366 items), parks (3249 items), restaurants (40240 items), services (20804 items), and entertainments (11117 items). We then calculated the density of each SGP type in each cell.

Socioeconomic condition

The population of Hangzhou with a 100 m resolution was obtained from the WorldPop database (https://hub.worldpop.org). Road density was obtained from OpenStreetMap (https://www.openstreetmap.org. openstreetmap. org). Property prices were obtained from the website of real estate agency, Chain Home(https://hz.lianjia.com). GDP was obtained from the Resource and Environmental Science and Data Center of the Chinese Academy of Sciences (https://www.resdc.cn/DOI/DOI.aspx?DOIID=33).

Statistical analysis

Spatial autocorrelation analysis

Global spatial autocorrelation analysis was performed using Moran’s I statistic for the spatial distribution of cases present per grid. The value of Moran’ s I was distributed between 1 and -1. Moran’s I > 0 indicated that similar values tended to occur in adjacent regions with a clustering tendency, whereas Moran’s I < 0 indicated that nearby locations tended to have different values with a dispersion tendency. To further assess the clustering patterns at the grid level, a local autocorrelation metric of spatial association was calculated. The results are shown in the LISA cluster maps, indicating the spatial distribution of clusters over local regions. This metric classifies important regions into high-high clusters, low-low clusters, high-low clusters, and low–high clusters. In this study, the high-high clustering areas were DF-epidemic areas. Moran’s I index and local autocorrelation metrics were calculated using GeoDa (v 1.18.0).

Spatial analysis for the effects of urban features on DF risk

ZIPR and GWZIPR analyses were used to quantify the effects of 16 variables (Table 1), including socioeconomic conditions, landscape patterns, and SGPs on DF risk, and to compare the performances of these two models. The comparison of these variables was listed in Table S2. Because of the high frequency of zero cell in the studied region, we used the ZIPR model as a global multivariate analysis (Lambert 1992). ZIPR model was applied using R package {pscl} to situations in which target data has relatively many of zero value, to go along with the other observed values (Zeileis et al. 2008). They are two-part models in ZIP, a logistic regression for whether an observation is zero or not, and a poisson regression for the other observations. GWZIPR was a geographically weighted version of the classical ZIPR that it allowed the effect of each factor to vary locally at each spatial location used R package {lctools} (Kalogirou 2016). By examining the spatial distributions of the coefficients, we can better understand the spatial heterogeneity of propagation processes (Nakaya et al. 2005). The GWZIPR can be expressed in mathematical form as follows:

$$logit({\psi }_{i})={\alpha }_{i0}({\mu }_{i}\text{, }{\nu }_{i})+{\sum }_{k=1}^{n}{\alpha }_{ik}({\mu }_{i}\text{, }{\nu }_{i}){x}_{k}+{\varepsilon }_{i}\;\left(\mathrm{Zero\;model}\right)$$
$$ln({\lambda }_{i})={\beta }_{i0}({\mu }_{i}\text{, }{\nu }_{i})+{\sum }_{k=1}^{n}{\beta }_{ik}({\mu }_{i}\text{, }{\nu }_{i}){x}_{k}+{\varepsilon }_{i }\;\left(\mathrm{Count\;model}\right)$$

where ψi was the proportion of zero case in cell i; λi was the number of DF cases in cell i; (μi, νi) was the central coordinate of cell i; αi0, βi0 were the intercepts; αik, βik were the correlation coefficients of the explanatory variable k in cell i; xk was the explanatory variable k; ε was the error. αi0, βi0, αik, βik, εi would change depending on the spatial location of cell i.

After standardization of the 16 variables, we constructed three independent datasets involving socioeconomic, LULC, and SGP variables, and applied them to the ZIPR and GWZIPR models to test their effects to DF risk. The relevant parameters of DF risk in Hangzhou were selected by a penalized regression, ZIPR-LASSO algorithm using R package {mpath} (Wang et al. 2016) to optimize the model. And the relative contributions of selected factors were calculated using R package {glmm.hp} (Lai et al. 2023). The predicted DF case number were obtained using the optimized GWZIPR model, and a risk level map was drawn. We superimposed the spatial location points of the DF cases in 2018 and 2019 to overlap on the risk level map to verify the accuracy of the model. A spatial autocorrelation test of the residual spatial distribution using R package {spdep} was conducted to compare the models for reasonableness. If Moran’s I value was large, indicating that the residuals were spatially autocorrelated, the assumptions of the model estimation were violated.

Results

Spatial distribution of DF risk

The spatial distribution of DF cases is plotted based on the number of cases in each cell (Fig. 2a). The DF occurrence in Hangzhou with 140 count cells and 822 zero cells was spread unevenly and showed a decreasing gradient from the city center to the outside. Based on the spatial autocorrelation analysis, the distribution of DF cases showed significant spatial autocorrelation (Moran’s I = 0.214 (p < 0.001), with similar levels of DF risk usually adjacent to each other. LISA cluster analysis (Fig. 2b) identified 34 high-high subdistrict clusters in 961 grids over Hangzhou, which were concentrated in the center of Hangzhou. Low–high clusters were found in nine grids, and these spatial outliers were adjacent to the high-high clusters. The three grids were high-low clusters. No “low-low” clusters were found. The spatial distribution of DF cases in the low-density areas was scattered.

Fig. 2
figure 2

Spatial distribution of DF cases (a) and LISA cluster map of DF density (b)

Full size image

Most high-high clusters were found in city center (Fig. 3a), indicating that most DF cases occurred in the cells with larger built-up area than in cells with more green and blue space. Property prices were not the highest in the high-high clusters, like other socioeconomic types, but in the low–high clusters (Fig. 3b). High-high clusters were dominant in most SGPs, except parks (Fig. 3c). The low–high clusters were mostly in new urbanizing areas with higher property prices and greenery.

Fig. 3
figure 3

Total areas of LULC types (a), socioeconomic condition (b), and SGP density (c) in LISA clusters

Full size image

Effects of landscape patterns, socioeconomic condition and SGP density

The comparison between GWZIPR and ZIPR with the three categories of variables showed that GWZIPR was a better-fitted model than the ZIPR (Table 2). A higher adjusted R2, lower AICc, and Moran’s I of the residuals were found in the GWZIPR model for each category. Variables related to SGPs had better explanatory power than LULC patterns and socioeconomic conditions in the GWZIPR model. Unexpectedly, the contribution of socioeconomic factors was lower than SGPs (Table 2).

Table 2 Comparison between GWZIPR and ZIPR with three explanatory categories for DF risk
Full size table

Variable selection for the optimized model

ZIPR-LASSO algorithm selected 9 variables for the following analysis (Fig. 4). The final optimized ZIPR model (Table 3) and GWZIPR model (Table 4) both contained 8 variables involving property price, GDP, population, entertainments, parks, malls, services, and total area of built-up area, removing one insignificant variable, restaurants. The optimized GWZIPR model (Table 4) considering spatial information was the best-fit model with the higher adjusted R2 and lower AICc and residual Moran's I than the global ZIPR models (Table 3). Compared with the initial models involving only one category of variables (Table 2), the performance of the optimized GWZIPR model considering the selected variables from three categories was also more powerful (Table 4).

Fig. 4
figure 4

Variables selection based the ZIPR-LASSO penalized algorithm. Dash line indicted the optimal lambda value in cross validation

Full size image
Table 3 Effects of the selected variables on the DF case number in the optimized ZIPR model
Full size table
Table 4 Descriptive statistics of the local parameter estimates of optimized GWZIPR model
Full size table

Contribution of selected variable to DF risk

We estimated the relative contribution of the eight selected variables in the optimized global ZIPR model, because the GWZIPR model was a local regression of which the contributions of variables were changed with the spatial locations. The total area of built-up area was the most powerful explanatory variable for DF risk (Fig. 5). Different with the comparison between categories, the property price which indicated the household income was the second one. Some special SGPs, such as entertainments and services, also played an important role in the spatial risk of DF transmission. Beyond our expectation, the DF risk was regardless of the greenspace and waterbody, which was thought to be closely related to mosquito density (Landau and Leeuwen 2012).

Fig. 5
figure 5

Relative contribution of each variable for DF risk in the ZIPR model

Full size image

Spatial heterogeneity of variables’ effects to DF risk

Based on the spatial visualization of the optimized GWZIPR model, we found that the effects of these selected variables were spatially dependent. In the count model, the density of the entertainments and park, built-up area (CA_B) and GDP had positive effects on the DF risk at the central part with high density of DF cases in Hangzhou (Fig. 6a, e, f, h). The density of the service and mall, and property price showed a reversed spatial pattern (Fig. 6b, c, g). Population had a positive effect on the eastern part, while a negative effect on the western areas (Fig. 6d). In the zero model, the malls and services were performed negative effects to the appearance of DF case (Table 4).

Fig. 6
figure 6

Spatial distribution of coefficients for each variable of count model of the optimized GWZIPR model descripted in Table 4

Full size image

The risk prediction of DF occurrence in Hangzhou was performed according to the optimized GWZIPR model, and the risk level was divided according to the predicted number of DF cases in each cell (Fig. 7a). In the studied region, 11 high-risk cells, 67 median-risk cells and 65 low-risk cells were predicted. In 2017, the predicted area with high and medium risk captured 91% of the observed DF cases in 2017 (Table 5, Fig. 7b). In 2018, this predicting model only captured 44% DF cases in the area with medium and high risk (Fig. 7c), which was due to the unusually stronger preventive and control measures in the city center following the 2017 DF outbreak. 83% cases were captured again in the high and medium risk area in 2019 (Fig. 7d) with a formerly preventive and control level.

Fig. 7
figure 7

DF risk evaluation in Hangzhou based GWZIPR. (a) Spatial prediction of DF risk; (b) Observed DF cases in 2017; (c) Observed DF cases in 2018; (d) Observed DF cases in 2019

Full size image
Table 5 Distribution of observed DF case number and proportion across the risk predicting area
Full size table

Discussion

Landscape pattern and socioeconomic condition due to urbanization dominantly regulated spatial risk of DF epidemic

The city of Hangzhou was located in the most urbanized region, the Yangtze River Delta in China, with high building density and property prices. In the present study, we found that the variables associated with high urbanization, such as more built-up area, high GDP and more entertainment places would increase the DF risk, especially at the city center. Mahmood et al. (2019) reported similar findings for mosquito habitats, where DF risk was associated with built-up area rather than particularly green space. In the optimized GWZIPR model, the built-up area positively affected DF transmission throughout most part of the city (Fig. 6f). This result suggested that human-mediated landscape patterns are strongly associated with DF epidemics. In the highly urbanized area, there were more water containers which provided safe breeding habitats for mosquito (Husnina et al. 2019), but were not suitable for its natural enemies, such as frogs or dragonflies. High population aggregation in central Hangzhou for social and recreational activities also increased exposure to mosquito bites and transmission of dengue (Patz et al., 2013). One study reported that residents who stayed at home had a lower DF risk than who did not in Vietnam (Anders et al. 2015). The eastern side of Hangzhou had a significantly higher level of urbanization than the western side (Ma, 2016). Our study found that the influence of built-up area and SGPs on DF risk was well matched with the urban structural consistency, which was consistent with the socioeconomic pattern of Hangzhou.

Among the related socioeconomic variables, property price which mostly represents the household income, had a negative relation with DF risk, which was supported by the study in Brazil (Bavia et al. 2020). High-income households could afford better anti-mosquito facilities and health care to reduce DF risk. We also found that property prices explained DF risk better than GDP. This revealed that DF risk was closely related to the local distribution of socioeconomic resources at a fine scale in cities (Mulligan et al. 2015). In contrast, the large-scale regional information, such as GDP and climate, typically explained epidemic disease across cities (Bhatt et al. 2013). The importance of property price also revealed the inequity of public health within cities (Smeltz and Carpenter 2022).

The key role of SGPs for DF risk

In our results, the effects of SGP’s density were spatial heterogeneity, with important explanatory power. Less consideration for the spatial dependence of SGPs effects has led to an underestimation of their roles in previous epidemical studies. However, in our study, their roles in determining spatial risk of DF within urban areas was beyond our expectation. The successful spread of DF mainly depends on the spatiotemporal overlap of mosquitoes and susceptible populations (Sumayyah et al. 2016). In exploring the impact of the six SGPs on DF risk, the places of entertainment had the greatest effect on DF because local people usually visited cinemas, karaoke houses, bars, fitness centers, etc., at late afternoon after work, when overlapped with the peak times of mosquito activity (Chen et al. 2014; Yan et al. 2018). The crowed and warm indoor environment of these entertainment places also activated the mosquito biting behavior. In contrast, the peak crowd hours in services houses were usually at the working hours, and in shopping malls at noon and evening. In addition, these were more enclosed environments with adequate air conditioning and cooling in the malls. These dry and cool environments were unsuitable for mosquitoes to inhabit and breed. Rare shopping malls also lead to a stronger aggregation effect. This might be the reasons for the negative impact of malls at the city center, even though they were also crowded places.

The results from the zero model of GWZIPR also illustrated that the density of shopping malls and services drove the dengue appearance. Previous classical epidemic studies proved that imported infection from the traditional epidemic areas (Southeast Asian countries) was the main driver for the outbreak in the new epidemic area (Yan et al. 2018). Fewer shopping malls increased the aggregation effect with both local and alien people. However, service facilities had the opposite effect in the count and zero model. We speculated that this was due to the different utilization of service facilities between alien and local people. The alien short-term visitors visited the service houses less than the local people who were closed in connection with the autochthonous dengue epidemic. Fewer service facilities also gather these visitors in a special narrow place to increase the connection between these potential infections.

Indirect effect of urban green–blue space

Despite the significant effect of urban green–blue space to mosquito density were reported in many studies, its effect was weaker than the SGPs and socioeconomic level to DF risk directly. It was not doubt that the positive relation between mosquito density and DF risk, but the spatiotemporal overlap of mosquitoes and susceptible populations was more important and direct for the DF epidemic in our study. We speculated that the green–blue space had a direct effect to mosquito density but indirect effect to DF risk without consideration of social activities.

Park was the only one variable referred with green–blue space in the final optimized models, which with more grassland, shrubs and pools providing suitable habitats for mosquito breeding (Manica et al. 2016), leading to higher mosquito density and a higher DF risk for visitors (Huang et al. 2018; Landau and Leeuwen 2012). In another word, some studies declared a negative effect of green–blue space due to mosquito-predator cascade in Thailand (Weterings et al. 2018). Urban green–blue space benefited the natural enemy’s population to inhabit mosquito population. The effect of green–blue space was unconfirmed and spatial heterogeneity. In our study, the effect of park to DF risk was positive at city center, but negative at city edge.

Implications for urban sustainability

The study framework and technical approach of landscape epidemiology applied in this study demonstrated the importance of the transdisciplinary for the urban public health and sustainable landscape governance. Beyond a better understanding of the socioeconomic contexts and ecological interactions that may determine disease transmission in a given location, landscape epidemiology with its transdisciplinary approaches can lead to the building of practical bridges for action between academic communities that otherwise rarely interact with each other, such as ecologists, epidemiologists and urban managers (Boischio et al. 2009; Yoshikawa et al. 2020). Furthermore, coupling the spatial heterogeneity with traditional epidemiological methods will support more precise predictions of spatial risk and more actionable prevention and control measures at fine scales.

At the policy and practice aspects, the impact of infectious diseases has always been missing in the framework of urban sustainable development. In the UN’s Sustainable Development Goals (SDGs), the content of health goal (SDG3) mostly focused on the woman and children health, and some traditional diseases but without most epidemic diseases. Most case studies on healthy sustainable urban development also ignored the epidemic diseases (Tonne et al. 2021). The catastrophe of global cities caused by new epidemic diseases in past decades has been underestimated in these urban sustainability studies. Bangert et al. (2017) illustrated that the tropical epidemic diseases would affect SDG3, but also affect and are affected by other SDGs. Based on our findings, the poor communities in the highly human-modified landscape in cities suffered the highest DF risk. We suggest that more parks should planned at the city center to reduce the high coverage of built-up area. More frequent surveillance of mosquitoes and infected people should be carried out in the low-income communities. The entertainment and services venues also should be strictly controlled during the epidemic period. Precise, effective and equitable prevention and control of these epidemic diseases (Weterings et al. 2018) with the framework of landscape epidemiology (Reisen 2010) can improve our prospects of achieving the urban sustainability.

Conclusions

In this study, we analyzed the urban spatial pattern of DF cases in Hangzhou, a new epidemic region in China, and explored the dominant factors regulating the spatial variation of the DF epidemic. We found that landscape patterns referred to urbanization including LULC and SGPs had a greater influence on DF transmission at the city level than socioeconomic conditions due to the spatiotemporal overlap of people and mosquitoes. The total area of built-up area at each cell had the greatest influence on DF risk, while the green–blue space had less direct effect on DF risk. These findings highlight the critical functions of landscape epidemiology in the management of public health and urban sustainable development. We suggest that DF prevention and control strategies should involve landscape patterns and SGPs in cities, which are rarely considered in classical epidemiology.