Natural and human environment interactively drive spread pattern of COVID-19: A city-level modeling study in China

Abstract

A novel Coronavirus COVID-19 has caused high morbidity and mortality in China and worldwide. A few studies have explored the impact of climate change or human activity on the disease incidence in China or a city. The integrated study concerning environment impact on the emerging disease is rarely reported. Therefore, based on the two-stage modeling study, we investigate the effect of both natural and human environment on COVID-19 incidence at a city level. Besides, the interactive effect of different factors on COVID-19 incidence is analyzed using Geodetector; the impact of effective factors and interaction terms on COVID-19 is simulated with Geographically Weighted Regression (GWR) models. The results find that mean temperature (MeanT), destination proportion in population flow from Wuhan (WH), migration scale (MS), and WH*MeanT, are generally promoting for COVID-19 incidence before Wuhan's shutdown (T1); the WH and MeanT play a determinant role in the disease spread in T1. The effect of environment on COVID-19 incidence after Wuhan's shutdown (T2) includes more factors (including mean DEM, relative humidity, precipitation (Pre), travel intensity within a city (TC), and their interactive terms) than T1, and their effect shows distinct spatial heterogeneity. Interestingly, the dividing line of positive-negative effect of MeanT and Pre on COVID-19 incidence is 8.5°C and 1 mm, respectively. In T2, WH has weak impact, but the MS has the strongest effect. The COVID-19 incidence in T2 without quarantine is also modeled using the developed GWR model, and the modeled incidence shows an obvious increase for 75.6% cities compared with reported incidence in T2 especially for some mega cities. This evidences national quarantine and traffic control take determinant role in controlling the disease spread. The study indicates that both natural environment and human factors integratedly affect the spread pattern of COVID-19 in China.

Graphical abstract

Highlights

• •A city-level two-stage model quantifies the environmental impact on COVID-19 in China.
• •Population flow from Wuhan and temperature is determinant before Wuhan’s shutdown.
• •After Wuhan’s shutdown (T2), human factors impact more showing distinct spatial heterogeneity.
• •The dividing line of positive-negative effect of temperature and precipitation on COVID-19 incidence in T2 is 8.5°C and 1mm.
• •Modeled incidence is higher than reported one in T2 especially for some mega cities.

1. Introduction

A novel Coronavirus was identified in December 2019 in Wuhan city of Hubei Province, China. Thus the disease was officially named by the World Health Organization (WHO) as COVID-19 and later its virus was identified to be a new coronavirus SARS-CoV-2 (Zhu et al., 2020). Till now, no medical cure is available for this disease, with absence of both preventive vaccine and specific drug (Liu et al., 2020; Shi et al., 2020). Since mid January 2020, it has rapidly spread throughout China and 835 cases were reported on 23 January 2020 (China, 2020). Thus a shutdown of Wuhan city was started on that day and a travel ban from Hubei Province was enforced on the following day. Meantime, China also declared it as level 1 national emergency response, defined as an “extremely serious incident” (Chinadaily, 2020), thus a national quarantine was since promoted. Since January 2020, it was also spread worldwide and declared as an international public health emergency by the WHO on January 30, 2020 (Lai et al., 2020). Till 19 October 2020, nearly 40.13 million cases have been reported globally, with the newly confirmed case of 389,683 on the day (WHO, 2020).

People have struggled to control the transmission of COVID-19 worldwide, thus global researchers have conducted many studies in order to explain its spatial-temporal distribution pattern. Firstly, some studies have investigated the impact of climate factors on global (Bannister-Tyrrell et al., 2020; Ficetola and Rubolini, 2020; Sobral et al., 2020), national (Bariotakis et al., 2020; Prata et al., 2020; Qi et al., 2020; Shahzad et al., 2020; Shi et al., 2020; Xu et al., 2020) and municipal (Bashir et al., 2020; Ma et al., 2020; Tosepu et al., 2020) COVID-19 pandemic, respectively. An early global study based on cases from the first imported case until 29 February 2020 showed preliminary evidence that higher temperature was strongly associated with lower COVID-19 incidence for temperature of 1°C and higher (Bannister-Tyrrell et al., 2020). A recent country-level study incorporated more bioclimatic factors and determined two major predictive variables, including minimum temperature of the coldest month (27.4%) and mean temperature of the wettest quarter (20.9%), thereupon potential distribution of global SARS-CoV-2 infection was predicted for April 2020 (Bariotakis et al., 2020). The impact of climate factors on COVID-19 incidence in different countries is respectively determined as temperature and absolute humidity in China (Shi et al., 2020), the interactive effect between daily temperature and relative humidity in mainland China (Qi et al., 2020), air quality in 33 locations of China (Xu et al., 2020), temperature in ten provinces in China (Shahzad et al., 2020), and temperature and humidity in the US (Gupta et al., 2020). The association between daily mortality of COVID-19 and meteorological factors and air pollutant index in Wuhan city indicated that diurnal temperature range (r = 0.44) was positively and relative humidity(r = −0.32) was negatively associated with COVID-19 mortality (Ma et al., 2020). The correlation between temperature and COVID-19 incidence in New York and Jakarta, Indonesia was also be explored (Bashir et al., 2020; Tosepu et al., 2020). However, the effect of temperature on COVID-19 transmissions is debating while predicting the spread of the disease in certain warm countries (Shakil et al., 2020).

Secondly, the effects of human and control measures on COVID-19 spread have also been investigated. A study showed that the international travel and border control during the early stages of the epidemic could reduce the rate of case exportations, but it could not fully arrest the global expansion of COVID-19 (Wells et al., 2020). Another study conceived that social lockdown was the only preventive measure against COVID-19 (Paital, 2020).

According to review on the published literatures, the impact of natural environment on COVID-19 incidence is mainly studied from aspect of climate factors. Presently, an integrated assessment of both natural and human environments on the disease transmission in China is lacking. Especially, the interactive effect between natural and human environmental factors on the COVID-19 incidence is never concerned. Therefore, this study will conduct a city-level study in China to investigate the correlation between both natural and human environment and COVID-19 spread, then model the disease transmission in two stages, and finally to evaluate the role of environment in the disease spread. Besides, modeling and prediction of disease incidence with absence of quarantine is conducted to quantitatively reflect the effectiveness of such measures.

2. Methods

2.1. Study data

In the study, city-level data were collected among 341 cities of 31 provinces in the mainland China (Fig. 1 ), between Jan 20 and Feb 29, 2020. The collected data mainly include reported COVID-19 case, natural environment (climate factors and terrain indicator) and human environment factors. Daily data on the number of new confirmed COVID-19 cases was obtained from the DX Doctor (http://ncov.dxy.cn), which is a collection of data reported by 32 provincial Health Commissions in China. The accumulated confirmed COVID-19 cases at a city level were shown in Fig. 1. The daily climate data, including minimum temperature (MinT), maximum temperature (MaxT), mean temperature (MeanT), precipitation (Pre), relative humidity (Rh), wind velocity (Wind) and air pressure (AirP), was downloaded from World Weather (https://en.tutiempo.net/). The terrain data, including minimum DEM (Min DEM), maximum DEM (Max DEM), and mean DEM (Mean DEM) was downloaded from Geospatial Data Cloud site, Computer Network Information Center, Chinese Academy of Sciences. (http://www.gscloud.cn). Among the human factors, GDP per capita (GDP), and population density (PD) was downloaded from Wind Economic Database (https://www.wind.com.cn); destination proportion in population flow from Wuhan (WH), migration scale (MS), and travel intensity within a city (TC) were downloaded from Baidu (http://qianxi.baidu.com/). The data type, variables and data source are listed in Table 1 .

Fig. 1.

Distribution of accumulated COVID-19 cases (till Feb 29, 2020) at city level in mainland China.

Table 1.

Data type, variables and data source.

Data type		Variable	Scale	Data source
Disease	COVID-19 data	Confirmed case	Daily	http://ncov.dxy.cn
Natural environment	Climate data	Minimum temperature	Daily	https://en.tutiempo.net/
		Maximum temperature	Daily
		Mean temperature	Daily
		Precipitation	Daily
		Relative humidity	Daily
		Wind speed	Daily
		Air pressure	Daily
	Terrain data	Minimum DEM	Yearly	http://www.gscloud.cn
		Maximum DEM	Yearly
		Mean DEM	Yearly
Human environment	Economic data	GDP per capita	Yearly	https://www.wind.com.cn
	Population data	Population density	Yearly
		Destination proportion in population flow from Wuhan	Daily
	Human mobility	Travel intensity within a city	Daily	http://qianxi.baidu.com
		Migration scale	Daily

The lifetime of the COVID-19 in a city in China is characterized by two-stage process, uncontrolled infection in early times and decaying stage at later times once quarantines are being performed (Gross et al., 2020). Therefore, according to the date when the Wuhan was shutdown, we divided the study period into two stages: T1 (Jan 20 and Jan 24, 2020) and T2 (Jan 25 and Feb 29, 2020). Based on the reported cases during T1 and T2, 115 and 291 cities were included in the analysis, respectively.

2.2. Statistical analysis

Our study is performed following the order of factor determining, modeling and prediction (Fig. 2 ). Firstly, after a correlation analysis and multicollinearity test, correlation factors were selected; then the Ordinary Least Squares (OLS) and Geodetector were used to finally determine the effective factors and interaction terms. Secondly, a Geographically Weighted Regression (GWR) model was developed and validated to quantify the effect of environment on COVID-19 incidence. Thirdly, migration scale data in 2019 was input the model to predict the disease incidence without control. All variables were first standardized and then put into the model.

Fig. 2.

Flowchart of data processing in the study.

2.2.1. Geodetector analysis

A Geodetector is a method to detect spatial stratified heterogeneity and determine the factors that drive it (Wang et al., 2010). The theoretical basis of the method is that if the spatial variability of an independent variable Y is caused by a specific factor X, the similarity should exist between spatial distributions of X and Y. The Geodetector is applied to calculate the contribution of each factor to COVID-19 incidence and to detect synergies between factors with respect to COVID-19 incidence. The Geodetector determines the spatial correlation of factor X and Y based on the q-statistic calculated with the following equation:

$$\mathrm{q}=1-\frac{\sum _{h=1}^L{N}_h{\sigma }_h^2}{{\mathit{N\sigma }}^2}$$ (1)

where N is the number of samples in the study area, L is the number of categories of factor X, σ ² is the total variance of Y in the study area, σ _h ² is the variance of Y within category h of factor X. The larger the q value is, the stronger the factor X explains Y.

The Geodetector includes four main detectors, and we mainly consider the interaction detectors (Yin et al., 2019) to determine the interaction terms of factors on the COVID-19 incidence. In this study, the notation * refers to the product of two variables, which is the interaction term. For example, Y = β ₁ X ₁ + β ₂ X ₂ + β ₃ X ₁ ∗ X ₂ means that the effect of X ₁ on Y depends on the value of X ₂. If an interaction exists in the data, interaction terms can provide a better description of the relationship between the independent and dependent variables, and offer a more accurate estimation of the relationship and explain more of the variation in the dependent variable.

2.2.2. Spatial temporal analysis

To investigate the impact of the effective factors and interaction terms on the COVID-19 incidence, both OLS and GWR models were applied. As a global regression model, OLS captures the average strength and significance of the independent variables (Delmelle et al., 2016) and thus derives one-size-fits-all outcomes. In other words, it analyzes the impact of the effective factors on the COVID-19 incidence for the whole China. However, it fails to consider local variations of the effective factors for each city (Sumanasinghe et al., 2016). Global Moran's I statistic for the COVID-19 incidence was conducted to measure spatial autocorrelation, in order to test whether the COVID-19 incidence is stationary across the China. By contrast, GWR is a local regression model, captures spatially relationships between the dependent variable and independent variables, which is different in varying locations (Zhao et al., 2020). Actually, for each location, GWR bases data from adjacent neighborhoods to conduct a local regression, thus estimating local regression coefficients for each of the predictor variables (Delmelle et al., 2016). Using COVID-19 incidence as the dependent variable, the GWR model can be expressed as follows:

$$Y_i={\beta }_0\left(i\right)+{\sum }_{k=1}^N{\beta }_k\left(i\right)X_k\left(i\right)+{\epsilon }_i$$ (2)

where Y _i is the dependent variable, X _k(i) is the k-th independent variable (k = 1…N), β ₀(i) is the intercept at neighborhood i, β _k(i) is the coefficient of X _k(i), ε _i is the residual of the location i. The adaptive bi-square kernel function and cross-validation were used to determine the ideal number of neighbors.

In our study, we used OLS forward stepwise to select the variables entering the regression model. Following this approach, variables are added sequentially until the coefficient of determination (R²) does not increase significantly anymore. Dominant variables are used in the GWR model. We compare regression results to the ones obtained using an OLS approach.

3. Results

3.1. Factor identification and model development

3.1.1. Factor identification

Factor is determined based on correlation analysis, multicollinearity test, Geodetector analysis, and OLS analysis (Fig. 2). The correlation between COVID-19 incidence and all the factors (Table 1) are listed in Table 2 . For the T1 stage, most factors are correlated with COVID-19 incidence, except for Pre, Rh, Wind and Max DEM; and MaxT, MeanT, MinT, Min DEM, GDP, PD, WH, MS and TC are significantly correlated with COVID-19 incidence (p < 0.01). While for the T2 stage, all the factors are significantly correlated with COVID-19 incidence, except Wind. After a multicollinearity analysis, some linearly correlated factors are excluded, such as AirP. Then for T1, two natural factors (MeanT, Mean DEM) and five human factors (GDP, PD, WH, MS, TC) are left as dominant factors; while for T2, more natural factors (MeanT, Mean DEM, Pre, Rh) and the same human factors (GDP, PD, WH, MS, TC) are dominant ones as the T1. All variance inflation factor values are ranged between 1.05–2.80 and 1.14–2.17 respectively for T1 and T2, indicating there is no multicollinearity.

Table 2.

Pearson correlations between COVID-19 incidence and environmental factors.

Natural factor	T1	T2	Human factor	T1	T2
MinT	0.326⁎⁎	0.298⁎⁎	GDP	0.338⁎	0.301⁎⁎
MaxT	0.282⁎⁎	0.211⁎⁎	PD	0.518⁎⁎	0.438⁎⁎
MeanT	0.313⁎⁎	0.268⁎⁎	WH	0.413⁎⁎	0.443⁎⁎
Pre	0.010	0.300⁎	MS	0.574⁎⁎	0.489⁎⁎
Rh	0.127	0.419⁎⁎	TC	−0.371⁎⁎	−0.409⁎⁎
Wind	−0.037	0.080
AirP	0.215⁎	0.418⁎⁎
Min DEM	−0.247⁎⁎	−0.396⁎⁎
Max DEM	−0.073	−0.337⁎⁎
Mean DEM	−0.224⁎	−0.414⁎⁎

^⁎Correlation is significant at the 0.05 level.

^⁎⁎Correlation is significant at the 0.01 level.

The interactive effect of the above dominant factors on the COVID-19 incidence is analyzed using Geodetector (Fig. 3 ). Clearly, the interactions of most pairs of natural factors and human factors were nonlinearly enhanced. The interaction terms with q > 0.6 (in blue box in Fig. 3) are then input OLS models to do final selection (Fig. 2), such as WH*MeanT in T1 and MS*TC in T2. After the OLS analysis, the effective factors and interaction terms are finally determined (Table 4). For T1, the effective factors include MeanT, WH, and MS, and the effective interaction term is WH*MeanT. While for T2, the effective factors are MeanT, Pre, Rh, Mean DEM, WH, MS, and TC, and the effective interaction terms include WH*Pre and MS*TC. All the determined factors and interaction terms are used to develop the model.

Fig. 3.

The interactive effect of dominant factors on COVID-19 incidence, where the blue box are the interaction terms with q > 0.6. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 4.

Comparison between OLS and GWR model.

Stage	Variables	Model	AICc	R2	Adjusted R2
T1	Effective factors	OLS	241.39	0.56	0.55
		GWR	240.31	0.58	0.56
	Effective factors and interaction terms	OLS	236.86	0.59	0.57
		GWR	235.56	0.61	0.58
T2	Effective factors	OLS	557.84	0.59	0.57
		GWR	470.58	0.74	0.70
	Effective factors and interaction terms	OLS	517.32	0.64	0.63
		GWR	440.23	0.77	0.73

3.1.2. Model development and validation

For choosing proper model, the spatial autocorrelation analysis is conducted for the COVID-19 incidence itself. The calculated global Moran's index for ln (case) in the study period is 0.45, indicating high spatial aggregation. This imply that GWR model is proper since GWR can capture spatially non-stationary relationships between the dependent variable and predictor variables by incorporating geographical information (Zhao et al., 2020). Then, using ln (case) as the dependent variable and the determined effective factors and effective interaction terms (Table 4) respectively as the independent variables, we developed both the GWR models and OLS models. To validate the accuracy of the developed models, we calculated the R², the adjusted R², and the Akaike Information Criterion (AICc) (Li et al., 2019; Zhao et al., 2020) for the models. First, the developed GWR models are compared to the OLS regression models (Table 4 ). Obviously, for both T1 and T2, the GWR models have higher values of R², adjusted R² and lower values of AICc than those of the OLS models, regardless of the model developed with the effective factors or the effective terms (Table 4). This comparison fully indicates the GWR model, with higher accuracy, is significantly more appropriate than the OLS model. Second, for both OLS and GWR models, the developed models with both the effective factors and interaction terms are much better than those meantime only with effective factors; the former models have the higher R² and the adjusted R² and smaller values of AICc (Table 4). This implied that the interaction of different factors could greatly promote COVID-19 incidence in comparison with single factor, whenever and wherever.

Table 3.

Effective factor and interaction terms for COVID-19 incidence based on OLS analysis.

Stage	Effective factors	Effective interaction terms
T1	MeanT, WH, MS	WH*MeanT
T2	MeanT, Pre, Rh, Mean DEM, WH, MS, TC	WH*Pre, MS*TC

3.2. GWR analysis of environment impact on COVID-19 incidence

3.2.1. GWR analysis of environment impact on COVID-19 incidence in T1

The spatial distribution of local coefficient (β _k) for the four explanatory variables (Table 4) in the final GWR model for T1 (Table 4) is mapped in Fig. 4 . In general, the effect of the three factors on the COVID-19 incidence is similarly promoting in T1. Under the average destination proportion in population flow from Wuhan (WH), MeanT has the strongest positive correlation with the COVID-19 incidence, with the values of coefficient ranging 0.80– 1.04. In other words, the hotter, the higher disease incidence occurs. While under average temperature level, WH is also positively correlated with the COVID-19 incidence, with the values of coefficient ranging 0.48– 0.59. The migration scale (MS) is also positively correlated with the disease incidence, and its effect degree weakens gradually from north to the south. The interaction between WH and MeanT is the strongest, with the values of coefficient ranging 1.82– 3.59. This implies that the linkage effect of WH and MeanT is very strong in this stage, and it is necessary to analyze the impact of two factors on the COVID-19 incidence at the same time. Under the same WH (MeanT) level, the higher COVID-19 incidence tends to occur in hotter (higher-WH) areas. The results indicate during early outbreak of the COVID-19, WH and MeanT are key determinants promoting disease transmission. Therefore, shutdown of Wuhan city is important and effective at the time.

Fig. 4.

The spatial distribution of the local coefficient (β_k) for the four variables used in the GWR model for T1.

3.2.2. GWR analysis of environment impact on COVID-19 incidence in T2

The spatial distribution of local coefficient for the final explanatory variables (Table 4) used in the GWR model in T2 (Table 4) is mapped in Fig. 5a. In comparison with T1, more (including Rh, Mean DEM, Pre, TC, and MS*TC) and different (WH*Pre) factors have significant effect on the COVID-19 incidence and their effect shows distinct spatial heterogeneity in T2; that is, the same factor has different impact on the disease in different areas. And the heterogeneity of the effect of natural factors is stronger than that of human factors. Even though with a few exception, the effect of temperature on COVID-19 incidence is spatially divided by the temperature about 8.5°C (Fig. 5b). Specifically, in areas with MeanT < 8.5°C, temperature is positively correlated with the COVID-19 incidence, and its effect degree strengthens gradually from west to the east; but in areas with MeanT > 8.5°C, this effect is negative and strengthens gradually from north to the south. Under the average WH level, there is obvious spatial heterogeneity in the impact of precipitation on the COVID-19 incidence, with a positive relationship in the west and a negative relationship in the east and northeast (Fig. 5c). Except for the northeast region, this positive-negative relationship is approximately divided by the precipitation equal to 1 mm. Most areas with Pre < 1 mm have a positive relationship, and its effect degree weakens gradually from west to the east. The positive relationship is strongest in Xinjiang province. Most areas with Pre > 1 mm have a negative relationship, but this effect has no obvious spatial heterogeneity.

Fig. 5.

The spatial distribution of the local coefficient (β_k) for the nine variables used in the GWR model for T2. (a) All the factors; (b) MeanT divided by the temperature about 8.5°C; (c) Pre divided by the precipitation equal to 1 mm.

In T2, human factors generally have more significant effect on the COVID-19 incidence than natural factors. The MS has the strongest positive correlation with COVID-19 incidence; while its correlation with WH is not so strong. Because Wuhan has been shut down in this stage, and the MS has become a determinant promoting disease transmission. The effect of Mean DEM on the COVID-19 incidence is the weakest, and the relationship is mostly negative except in a few areas. For example, in rugged areas within the “third topography ladder” of China, such as Fujian Province, the positive relationship is relatively strong.

The interactive effect of environmental factors on the COVID-19 incidence is far more complex than in T1. The interaction between WH and Pre is always negative (Fig. 5a), which means that one factor will reduce the effect of the other on the COVID-19 incidence. In other words, under the same WH (Pre) level, the less precipitation (WH), the higher disease incidence occurs. The interaction between MS and TC is generally positive (Fig. 5a), with some exception in areas such as Akesu and Yili in XinJiang Province. Overall, under the same MS (TC) level, the higher TC (MS), the higher disease incidence occurs.

The above two-stage modeling not only examines the different impact of environment on COVID-19 incidence in early and late period, but also quantitatively reflects the inhibition effect of shutdown of Wuhan city and following national quarantine.

3.3. Modeling and risk analysis of COVID-19 in uncontrolled T2

According to the above results, MS is the most important factor affecting COVID-19 spread for T2 (Fig. 5). Therefore, a scenario is defined as “uncontrolled T2” to represent T2 with the meantime MS situation in 2019. Then MS data in 2019 is input as an alternative independent variable, and we model the COVID-19 incidence based on the developed model for T2 (Table 1). The spatial distribution of the reported incidence in T2 (Fig.6a) is compared with that of the modeled incidence under this scenario (Fig.6b). The difference between them (Fig.6c) indicates the changed MS alone could cause an increase in the COVID-19 incidence in 75.6% of cities nationwide. The modeled COVID-19 incidence has increased in almost all provincially capital cities (Fig.6b). Such increase is particularly prominent in some mega cities, such as Beijing, Shanghai, Guangzhou and Chengdu, where booming outbreak of the COVID-19 is likely to occur as predicted (Fig.6c). This evidences the effectiveness of national quarantine and traffic control in inhibiting the disease spread. In contrast, the predicted disease incidence is decreased in some areas, such as Wenzhou and cities around Wuhan. The population movement between Wenzhou and Wuhan is very frequent. According to the Wenzhou city government report on Jan 29, 2020, there are about 180,000 Wenzhou people doing business, working and studying in Wuhan. And a total of 433,000 people have been identified to return to Wenzhou from Wuhan and surrounding key areas (Yang et al., 2020). Similar phenomena exist in cities around Wuhan. These all imply the dominant factor for these cities is WH instead of MS, so the modeled incidence may not be a very accurate estimate.

Fig. 6.

Risk analysis of COVID-19 incidence based on T2 situation: (a) Reported incidence in T2; (b) Modeled incidence using GWR model in T2 with 2019 migration scale; (c) The difference between the modeled and reported incidence.

4. Discussion

4.1. Research implication

This study has implications for the two aspects. On the one hand, climate change is well known to have significant effect on infectious diseases (Li et al., 2018; Wu et al., 2020; Wu et al., 2016). Our study indicates that temperature, precipitation and relative humidity are three important climate indicators for COVID-19 incidence. The three factors are also found effective on the other coronavirus transmitted diseases, including SARS-CoV (Tan et al., 2005) and MERS (Alghamdi et al., 2014). Furthermore, temperature is usually a powerful indicator for these diseases (Tan et al., 2005) and the impact of temperature is proven to vary within different temperature ranges. Therefore, even though the temperature and humidity have significant effect on COVID-19 in China (Qi et al., 2020), Brazil (Auler et al., 2020), and the US (Gupta et al., 2020), the effect of temperature on COVID-19 is different among provinces (Qi et al., 2020; Shahzad et al., 2020) and cities in China. The improvement of our study lies in that it not only indicates that the effect of climate factors on COVID-19 has strong spatial heterogeneity, but also determines the dividing line of positive-negative effect of MeanT and Pre on COVID-19 incidence is 8.5°C and 1 mm, respectively. A study in mainland China finds that there is a potential interactive effect between daily temperature and relative humidity on COVID-19 incidence (Qi et al., 2020). Another study in 33 locations in China finds that the degree of the effect of air quality on COVID-19 incidence is related to climate factors, when the temperature of 10–20 °C or the relative humidity of 10–20%, the air quality has a stronger effect (Xu et al., 2020). The improvement of our study lies in that the former studies only consider the interaction between climate factors, while ours consider interactions between all factors and quantified these interactions. Furthermore, we find interaction between natural and human factors is stronger than the interaction between natural factors.

On the other hand, human activity has been reported as an important driver for the transmission of infectious diseases (Wu et al., 2014). Our study infers that destination proportion in population flow from Wuhan, travel intensity within a city, and migration scale were three contributing factors for the COVID-19. It is evidenced that the transportation network of highways and railways (Fang et al., 2009) and distribution of ring road (Wang et al., 2008; Wang et al., 2006) are important in spreading an SARS epidemic. This indicates traffic transportation is crucial to instant spread of emerging coronavirus diseases. Prior studies on COVID-19 find that lockdown of Wuhan has a great efficacy, the population outflow from Wuhan has a strong correlation with the confirmed cases in each county (Jia et al., 2020), and measures taken to close schools and workplaces are conducive to controlling the outbreak of COVID-19 in Wuhan (Prem et al., 2020). The selected five human factors in this study not only include those corresponding to the indicators in previous studies, but also consider more aspects than the previous studies. For example, “GDP” in our study contains both “the number of doctors” (Qiu et al., 2020) and “the number of companies” (Briz-Redón and Serrano-Aroca, 2020); and our “MS” is highly correlated with “the number of travelers”. Besides, the human factors we selected are more comprehensive and representative. For example, WH, MS, and TC can reflect the human mobility in key disease outbreak cities, between different cities, and within cities, respectively. A great significance lies in that our study find that the interactive effect of two human factors is much more significant than the single factor. We believe that the interaction between migration scale and travel intensity within a city is typical in reflecting human mobility, and thus both is determinant in the disease transmission in T2. Therefore effective preventive control on human mobility should be promoted in other pandemic areas.

4.2. Research contribution

This study has advanced our understanding of the COVID-19 from the four aspects. Firstly, this is a systematic study on integrated assessment of environment on COVID-19 in China. Natural and human environment is geographically correlated, which is rarely detected. Compared with single-perspective environmental study, such a comprehensive evaluation could truly reflect the underlying environment driving of the newly emerging disease. This could also shed light on the impact of environment on other diseases. Secondly, the interactive effect of natural and human environmental factors has been innovatively addressed. This investigation contributes to explaining the spatial-temporal spread pattern of COVID-19 in China. Thirdly, this is a novel spatial modeling of COVID-19 incidence in China using GWR model. Environment and disease distribution are spatially correlated in essence and our model quantitatively depicts this. Fourthly, we have conducted two-stage investigation on impact of environment on COVID-19 incidence at city level, considering the transmission and virulence of the SARS-CoV-2 virus varies in different conditions (Shi et al., 2020). The stronger interactive effect of natural and human environment in some zone has further proven and evidenced this.

5. Conclusion

In this study, we investigate the effect of both natural and human environment on COVID-19 pattern at a city level, based on the two-stage modeling using GWR models. The study finds four effective factors (MeanT, WH, MS) and interaction terms (WH*MeanT) are generally promoting for COVID-19 incidence before Wuhan's shutdown (T1), and the WH and MeanT play a determinant role in the disease spread in T1. The effect of environment on COVID-19 incidence after Wuhan's shutdown (T2) includes more factors than T1, and their effect shows distinct spatial heterogeneity. The impact is spatially divided by temperature about 8.5°C and precipitation of 1 mm. In T2, human factors have more impact than the natural ones. WH has weak impact than in T1, but the MS has the strongest effect. The COVID-19 incidence in T2 without quarantine is also modeled using the developed GWR model, and the modeled incidence show an obvious increase compared with reported incidence in T2 especially for some mega cities. The results indicate that both natural environment and human factors integratedly affect the spread pattern of COVID-19 in China, and national quarantine and traffic control take determinant role in controlling the disease spread. Therefore, our future research will investigate the effect of different social and traffic control measures on controlling the COVID-19 spread, so as to provide a reference for other pandemic countries.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Editor: Jay Gan