The distribution of COVID-19 in China was highly correlated with population emigration from Hubei, especially from Wuhan. To explore the county-level spatial distribution of the epidemic in Guangdong, county-level population emigration from Hubei to Guangdong was estimated using Baidu Qianxi’s daily population migration data. The county-level population migration of Guangdong was estimated by integrating the total confirmed COVID-19 cases of Hubei, the population migration data from Hubei to Guangdong, and the resident population data of Hubei and Guangdong. First, we calculated the COVID-19 incidence rate of each city in Hubei using the following mathematical expression: i n c i d e n c e _ r a t e i = c o n f i r m e d _ c a s e s i p o p u l a t i o n i , (1) where “i” represents the number of cities of Hubei province, confirmed cases _i represents the total confirmed COVID-19 cases, and population _i is the resident population of city “i.” Combining the incidence rate of COVID-19 and the total population migration data from each city of Hubei with other cities of Guangdong, we estimated the possible number of infected migrants as follows: i n f e c t e d   m i g r a n t s j = ∑ i = 1 m i n f e c t e d   m i g r a n t i , j (2) i n f e c t e d   m i g r a n t i , j = i n c i d e n c e _ r a t e i × m i g r a n t i , j , (3) where “i” and “j” represent the number of cities of Hubei and Guangdong provinces, respectively, infected migrants _j is a summation of the possible number of infected migrants traveling from Hubei to cities “j” of Guangdong, while infected migrants _j is the number of possible infected migrants from city “i” of Hubei to city “j” of Guangdong, and migrant _i,j stands for total population migration from city “i” of Hubei to city “j” of Guangdong, obtained from the population migration data from Baidu Qianxi. Last, the county-level, possibly infected migrants from Hubei to Guangdong were estimated using the following formula: i n f e c t e d   m i g r a n t s j k = i n f e c t e d   m i g r a n t s j × p o p _ r a t e j k (4) p o p _ r a t e j k = p o p j k p o p j , (5) where “j” and “k” represent the cities and counties of Guangdong Province, infected migrant s j k represents the final possible infected migrants from Hubei to county “k” of city “j” in Guangdong, p o p _ r a t e j k   is the population proportion of county “k” as compared with city “j” in Guangdong, p o p j k is the resident population of county “k” of city “j”, and pop _j is the resident population of city “j” in Guangdong. Here, the county-level, number of possibly infected migrants were used to explore COVID-19 transmission in Guangdong.

Spatial autocorrelation can always be used to measure spatial similarities between nearby observations; it describes the degree to which observations of a variable (e.g., epidemic disease) are similar to each other at a given spatial scale. Here, a global and local spatial autocorrelation analyses were applied to analyze the spatial agglomeration pattern of the confirmed COVID-19 cases at the county level in Guangdong, including the global and local Moran’s I indices. The global Moran’s I index was applied as shown in Eq. 6 (Moran 1950). Mora n ' s I = n ∑ i = 1 n ∑ j = 1 n w i j ∑ i = 1 n ∑ j = 1 n w i j ( x i − x ¯ ) ( x j − x ¯ ) ∑ i = 1 n ( x i − x ¯ ) 2 , (6) where n indicates the number of counties in Guangdong, x i   and x j indicate the confirmed COVID-19 cases of counties “i” and “j”, x ¯ is the average number of confirmed COVID-19 cases of Guangdong, w i j indicates the spatial matrix of counties “i” and “j”, which was defined as “1” if county “i” is contiguous with county “j” and “0” if otherwise. The Moran’s I value ranges from “-1” (strong positive spatial autocorrelation) to “1” (strong negative spatial autocorrelation), while a value of “0” indicates no spatial autocorrelation. In this study, p-value was used to indicate whether the index value was statistically significant or not.

Furthermore, the local indicator of spatial association (LISA) (Anselin 2010) was also applied to reflect the spatial correlation between the COVID-19 cases of a county and its adjacent counties. The LISA index was defined by (Anselin 2010): Local Mora n ' s I = n ( x i − x ¯ ) ∑ j = 1 m w i j ( x j − x ¯ ) ∑ i = 1 n ( x i − x ¯ ) 2 , (7) where m indicates the number of counties geographically adjacent to county “j.” LISA consists of five types of local spatial autocorrelation: high–high clusters (high number of COVID-19 cases surrounded by the same), low–low clusters (low number of COVID-19 cases surrounded by the same), high–low clusters (high number of COVID-19 cases surrounded by low number of cases), low–high clusters (low number of COVID-19 cases surrounded by high number of cases), and spatial randomness (no significant spatial patterns). In this study, the spatial autocorrelation analysis of the COVID-19 epidemic was implemented by using GeoDa software (Anselin et al., 2006).

Because of its advantages (without assuming linearity between geographical phenomenon and the driving factors and without being affected by multicollinearity), the geographical detector (Wang and Xu 2017) has been widely applied to detect spatial variation of geographical phenomena and reveal their potential driving forces; it contains four detectors: factor, risk, ecological, and interaction. The geographical detector was implemented using GeoDetector software (http://www.sssampling.org/Excel-Geodetector). An essential assumption of this method is that the spatial distribution of COVID-19 transmission is similar to that of its potential driving forces. Here, only the factor and interaction detector were involved in the driving force analysis.

Here, the factor detector was used to quantitatively measure the relative importance of the factors influencing COVID-19 transmission, which could be expressed with the power determinant (q). The mathematical expression of PD is as follows (Wang et al., 2010): q = 1 − 1 n σ 2 ∑ j = 1 m n j σ j 2 , (8) where m is the grade number of potential driving factors (j = 1, 2, … , m). n and n _j are the number of counties in the whole area and the j grade, respectively, and σ 2   and σ j 2 are the variances of the confirmed COVID-19 cases in all counties of Guangdong and the j grade, respectively. In general, the value of q ranges from “0” to “1”, which indicates that the driving factor can explain 100 × q% of the spatial pattern of the epidemic. A larger q value indicates the stronger influenced effects of the factors on COVID-19 transmission.

Furthermore, the interaction detector was used to investigate the interactive effects of two different factors x ₁ and x ₂ on COVID-19 transmission, which was assessed by comparing the values of q ( x 1 ∩ x 2 ) , q ( x 1 ) , and q ( x 2 ) . The value of q ( x 1 ∩ x 2 ) indicates the power determinant for a new factor created by overlaying the two factors x ₁ and x ₂. The interaction between factors x ₁ and x ₂ can be divided into the following categories (Wang et al., 2010; Wang and Xu 2017): { q ( x 1 ∩ x 2 ) < min ( q ( x 1 ) ,   q ( x 2 ) )                                                                                                                                               N o n l i n e r − w e a k e n min ( q ( x 1 ) ,   q ( x 2 ) ) < q ( x 1 ∩ x 2 ) < max ( q ( x 1 ) ,   q ( x 2 ) )                                     U n i − e n h a n c e / w e a k e n max ( q ( x 1 ) ,   q ( x 2 ) ) < q ( x 1 ∩ x 2 ) < ( q ( x 1 ) + q ( x 2 ) )                                                   B i − e n h a n c e q ( x 1 ∩ x 2 ) = ( q ( x 1 ) + q ( x 2 ) )                                                                                                                                                           I n d e p e n d e n t q ( x 1 ∩ x 2 ) > ( q ( x 1 ) + q ( x 2 ) )                                                                                                                                                                   N o n l i n e a r − e n h a n c e (9) where ∩ indicates the interaction between factors x ₁ and x ₂.

Nine factors, in terms of population migration and socioeconomic data at the county level in Guangdong, were selected. In particular, the socioeconomic data were represented by infected migrants (InfM), resident population (ResP), residential communities (ResC), companies (Comp), community facilities (ComF), restaurants (Rest), transport services (TranS), railways (RailL), and roadways (RoadL). According to the input requirements of the geographical detector, these factors were categorized on considering prior information as well as the range and distribution of the data. The number of categorized types of each factor is shown in Figure 2.

Figure 3 shows the spatial distribution of cumulative confirmed cases from 1 December 2019 to 17 February 2020 at the city level in Hubei and Guangdong. Cumulative confirmed cases of Hubei and Guangdong were up to 59,989 and 1,328, respectively. However, 24 cases in Shenzhen city of Guangdong Province that were confirmed at railroad stations and airports were not considered because the infected persons were immediately rushed to COVID-19–designated hospitals. Thus, the final number of cumulative confirmed cases of Guangdong was 1,304. In Figure 3A, Wuhan showed the largest number of COVID-19 cases, accounting for most cases in Hubei. There were over 1,000 cases in seven of seventeen cities in Hubei, including Wuhan, Xiaogan, Huanggang, Jingzhou, Ezhou, Suizhou, and Xiangyang. Figure 3B indicates that Shenzhen and Guangzhou had the most confirmed cases, with 392 and 339 cases, respectively, together accounting for 56% of all cases in Guangdong. COVID-19 mainly spread across the Pearl River Delta (PRD). However, Zhanjiang City had up to 22 confirmed cases, being a tourist-city located in the southernmost part of Guangdong. This could be as many people from Hubei province might have traveled to Zhanjiang City for the Spring Festival. Only five cities in Guangdong had less than 10 confirmed cases: Jieyang, Shanwei, Chaozhou, Heyuan, and Yunfu (no cases).

The incidence rate and the corresponding confirmed cases at the city level in Hubei are shown in Figure 4A where it can be seen that the largest incidence rate of COVID-19 occurred in Wuhan with up to 38.58 cases per 10,000 people, followed by Ezhou, Xiaogan, Suizhou, and Tianmen with incidence rates of 12.42, 6.75, 5.77, and 5.18, respectively. Although Ezhou had the second largest incidence rate, its confirmed cases were less than that of Xiaogan, Huanggang, and Jingzhou. Tianmen and Xiantao showed high incidence rates with few confirmed cases, i.e., 500 and 544, ranking 14th and 13th^, respectively, among 17 member cities in Hubei. Enshi had the lowest incidence rate (less than one) with 250 confirmed cases, as shown in Figure 4B. Shennongjia only had 10 confirmed cases, far less than 100, but its incidence rate was higher than one. It can be indicated from Figure 4 that most cities adjacent to Wuhan might have greater confirmed cases and higher incidence rates. This may be attributed to the population migration from Wuhan, which can be deduced from Baidu Qianxi.

Figure 5 depicts the spatial distribution of confirmed COVID-19 cases of each county in Guangdong. The epidemic mainly occurred in the PRD region, an economically developed area, with a good flow of people and convenient transportation. Among 121 counties of Guangdong province, only Bao’an District of Shenzhen city had confirmed cases greater than 100. In addition, eight counties had confirmed cases greater than 50, including Dongguan city; Nanshan, Futian, and Longgang districts of Shenzhen; Baiyun and Haizhu districts of Guangzhou city; Zhongshan city; and Xiangzhou District of Zhuhai city. These high-risk areas were mainly distributed in the Pearl River estuaries, except for Nansha District of Guangzhou city. In this study, Dongguan and Zhongshan cities have no county administrative units. Thus, counties in Dongguan and Zhongshan cities were not considered.

To examine the spatial autocorrelation of the COVID-19 epidemic at the county level in Guangdong, the global and local Moran’s I indices were estimated using the confirmed cases of each county. Figure 6 shows that the global Moran’s I value was 0.66 with a p-value of <0.0001, indicating significant positive autocorrelation with the COVID-19 epidemic. Furthermore, the local Moran’s I values at the county level were calculated using Eq. 9; the results represented detailed local clusters and local spatial outliers in relation to confirmed cases. Figure 7 shows cluster maps identified with the local Moran’s I values and significance of p < 0.05. High–high spatial clusters were mainly concentrated in the central Guangdong Province, along the Pearl River estuaries. Out of 121 counties, 15 showed significant high–high clusters, including five counties of Shenzhen, five of Guangzhou, two of Foshan, one of Zhuhai, one of Dongguan, and one of Zhongshan. However, other counties showed no significant spatial clustering patterns, thus illustrating a spatial stochastic distribution of the COVID-19 epidemic.

Integrating the COVID-19 incidence rate of Hubei, the population migration data of Baidu Qianxi, and the resident population of Guangdong, the county-level population migration from Hubei to each county of Guangdong was calculated using Eqs 1–5. First, the possible number of COVID-19 infected migrants from each Hubei City to that of Guangdong was calculated using Eqs 1, 3, shown in Tables 2, 3, from which it can be seen that most of the infected migrants traveled from Wuhan City in Hubei into Guangdong, especially the Pearl River cities. Most of the infected migrants moved into Shenzhen, Guangzhou, Dongguan, Foshan, Zhuhai, Huizhou, and Zhongshan. Shenzhen and Guangzhou had up to 39 infected migrants, followed by Dongguan and Foshan with about 10. However, cities such as Jieyang, Maoming, Yunfu, Chaozhou, Shanwei, and Yangjiang had fewer infected migrants from Wuhan or other Hubei cities. The results may be reliable since these cities are economically underdeveloped. In general, many people would tend to migrate to developed cities with convenient transportation while some would tend to travel to the PRD and tourist cities of Guangdong over the Spring Festival. However, the results are subject to the limited population migration data; Baidu Qianxi provided data only for the top 100 cities.

Then, the cumulative infected migrants from Hubei to each city of Guangdong Province were calculated using Eq. 2, as shown in Figure 8A, which showed that 4 out of 21 cities had >10 infected migrants, while 13 cities had <1. The highest number of infected migrants mainly occurred in the PRD, including Shenzhen, Guangzhou, Dongguan, Foshan, Huizhou, Zhuhai, and Zhongshan. However, Zhanjiang had 1.66 infected migrants, distant from that of the PRD, possibly as Zhanjiang is a tourist city that would attract many people during the Spring Festival.

Figure 9 shows the relationship between confirmed COVID-19 cases and population migration at the city level; the coefficient of determination of the regressive equation between the infected migrants and confirmed cases was as high as 0.981 with statistical significance larger than that of the regressive equation between the migrants and confirmed cases. This could be as Figure 9A considered not only the Baidu Qianxi migration data but also the risk of COVID-19 infection of each Hubei City. Figure 9B estimated the number of migrants using only the Baidu Qianxi migration data and assumed identical risk of infection for each Hubei City. However, this was unreasonable. In general, a high incidence rate could represent high risk of infection, i.e., migrants from Wuhan (with the highest incidence rate) should have the highest risk of infection, while migrants from Enshi (with a low incidence rate) should have a lower risk of infection.

Last, based on city-level infected migrants and the resident population, the county-level infected migrants from Hubei to Guangdong were estimated using Eqs 4, 5. County-level infected migrants in Guangdong are shown in Figure 8B. Dongguan city and Longgang and Bao’an counties of Shenzhen city showed the largest number of infected migrants, being >13, followed by eight counties from Guangzhou and Shenzhen cities with >5 infected migrants. In summary, 29 counties (located mainly in the PRD) had >1 infected migrants, and 22 counties (mainly in eastern and southwestern Guangdong) had <0.01 infected migrants. However, due to a lack of actual population migration data between counties of Guangdong Province, shortcomings exist in calculating the number of county-level infected migrants. In the future, population migration data between counties can be accurately estimated by integrating location data from mobile phones and social media.

Using the geographical detector technique, the power of determinant (q) values were calculated to evaluate the intensity of factors significantly influencing COVID-19 transmission (Table 4). As shown in Table 4, the q values of these factors were ranked as follows: from the highest to the lowest: infected migrants, company, transport services, residential communities, restaurants, community facilities, and railways and each of the p values was <0.01. The results indicated that infected migrants, companies, transport services, residential communities, restaurants, and community facilities had significantly promoted the spread of the COVID-19 epidemic, among which infected migrants had the greatest impact. These spatial risk factors can predominantly help interpret the spatial heterogeneity of the epidemic in Guangdong. Railways proved to have a weak association with the epidemic. Furthermore, the q value of roadways was 0.333 with a p value > 0.05, indicating that it had no significant correlation with COVID-19 transmission. Except the infected migrant factor, the company and transport service factors showed the second and third highest explanatory power in predicting the spread risk of the epidemic, respectively. This may be reasonable that these places of companies and transport services were a place of people gathering, where these persons may be across many locations. These may increase the risk of COVID-19 transmission. The transport services, including airport, railway station, subway station, bus station, and parks, were the essential components of traffic network, where some of the infected migrants from Hubei province may across to Guangdong Province to work or travel by car, train, and plane. A small gathering with the infected patients was also unsafe for other people, who may gather at the places of residential communities, restaurants, and community facilities. This may be because that the infected patients may pass on their infection to others through direct contact or close proximity. Thus, these places also belonged to high-risk areas, which may contribute to the result that the spread risk was high in the county.

The interaction detector was used to investigate the interactive effects of any two factors on COVID-19 transmission (Table 5). As roadway was not significantly related to COVID-19 transmission, it was not considered in this analysis. Table 5 shows that the interactive effects between any two factors were all bi-enhanced as compared to their independent influence. The geospatial risk of COVID-19 transmission was intricate and complicated by multiple spatial risk factors. According to the interaction detector analysis, we found that there were interactions between different factors. Among them, the infected migrant factor interacted strongly with the community facility factor and the company factor, and their q values were 0.895 and 0.891, respectively. This indicated that the combined effect produced by the intersecting infected migrant, community facility, and company factors had a strong explaining capacity on the COVID-19 transmission at the county level. Besides, Tables 4, 5 also demonstrated the interaction of infected migrants with restaurant, transport service, residential community, and railway factors could enhance the ability to explain the spatial distribution of the COVID-19 epidemic because they all had a high q value greater than 0.880. The places of company, restaurant, transport service, residential community, and railway were the public places of big population density and strong fluidity, which were important places for clustering transmission (Yang Z. et al., 2020). If the infected people carrying the COVID-19 virus gathered at these public places and were in close contact with other persons, which may make the healthy persons be exposed to the virus. This may result in accelerating the COVID-19 transmission. Furthermore, the combined effect produced by the intersecting railway and transport service factors may increase the explanatory power to the COVID-19 epidemic. The railway and transport services were a part of traffic network, which may increase the risk of the large-scale spread of COVID-19. Residents’ feasting, shopping, and other activities may increase the chances of contact between people, and there may be the possibility of mutual infection. Thus, to reduce the risk of COVID-19 transmission, measures such as social distancing, reducing crowds, reduce visits, and limit travel may be necessary (Hellewell et al., 2020; Linka et al., 2020; Fowler et al., 2021; Pei et al., 2021). Companies, restaurants, transport services, and other places were closely related to the spread of the virus; thus, disinfection should be carried out, and the humans should be reminded to pay attention to their own protection.

In general, this study quantitatively analyzed the spatial distribution characteristics of the COVID-19 epidemic at county scale by using a series of confirmed cases, population migration, and socioeconomic factors before the Spring Festival. The spatial risk factors of the COVID-19 epidemic were also summarized at the county level. The results showed that the regions with more community facilities, companies, and restaurants may have high-risk of COVID-19 transmission. Before the Spring Festival, although a large-scale population moved out of Guangdong to go home for the Spring Festival, most of the infected migrant population may still be back from a business trip and travel to Guangdong for the Spring Festival with their family and loved ones through Hubei. These infected migrant populations may mainly gather in companies, community facilities, and restaurants, and their high mobility also increased the chances of their contact to other persons. This may significantly increase the COVID-19 risk. These places with more community facilities and companies, where the infected migrant population have visited, should been paid close attention to and could be monitored by integrating the trajectory population-data from mobile phones and social media. However, limited by data-availability, only population migration and a few socioeconomic factors were used here, which may result in lack of in-depth research of spatiotemporal transmission mechanisms of COVID-19. Besides, the Baidu migration data may have a certain bias in the sample statistics, which may introduce potential errors in the estimation of the infected migrant population. In future, the mobile phone data will be integrated to adjust the Baidu migration data to improve the accuracy of the potential number of infected migrant population. Then, the spatiotemporal transmission process of COVID-19 and its mechanisms at county (even street) levels can be analyzed and mined by integrating spatiotemporal social, environmental, and economic data and big data of social media (Xiong et al., 2020; Zhou et al., 2020). Furthermore, to better understand the consistent relationship between the spatial risk factors and the COVID-19 epidemic situation, a more detailed spatiotemporal analysis approach may be conducted by integrating the infected population and the cross-regional migrant population.