A review of GIS methodologies to analyze the dynamics of COVID‐19 in the second half of 2020

3.1.1. Spatial regressions

In spatial modeling studies, it is common to start with ordinary least squares (OLS) regression to identify significant relationships between the dependent and independent variables. If the residuals of an OLS model are spatially autocorrelated, then it is appropriate to use spatial regression‐based methods (Delmelle, Hagenlocher, Kienberger, & Cases, 2016). For example, a spatial lag model (SLM) can be used to examine how events at a location influence similar events in surrounding locations (i.e., spatial interaction); and a spatial error model (SEM) can be applied to account for autocorrelation of the residuals (Iyanda et al., 2020; Maiti et al., 2020; Mollalo, Vahedi, et al., 2020; Nian et al., 2020; Sannigrahi et al., 2020; Sun, Di, Sprigg, Tong, & Casal, 2020; Urban & Nakada, 2021); see Table 1. For COVID‐19, spatially combined autoregressive models (SAC) have also been used as a combination of the previous models to simultaneously consider spatial lag and spatial error parameters (Sun, Di, et al., 2020; Zulkarnain & Ramadani, 2020).

Another common method is geographically weighted regression (GWR), using the variables previously included in OLS regression (Alkhaldy, 2020; Fan, Zhan, Yang, Liu, & Zhan, 2020; Iyanda et al., 2020; Karaye & Horney, 2020; Shariati, Jahangiri‐rad, Mahmud Muhammad, & Shariati, 2020; Urban & Nakada, 2021; Wu et al., 2020). GWR creates a local model and calculates the parameters for all points of the sample considering the spatial variation in the relationships (Brunsdon, Fotheringham, & Charlton, 1996; Maiti et al., 2020). It can consider non‐stationary variables (such as climate, demographic factors, and environmental factors) and models the local relationships between those predictors and the patterns under study. It facilitates the analysis of spatial variation in a phenomenon in a given place (Murgante et al., 2020), following Tobler's first law of geography—that “everything is related to everything else, but near things are more related than distant things” (Tobler, 1970). Regarding COVID‐19, GWR has been used to examine the relationships between the disease and air quality and a variety of socioeconomic variables. For example, Fan et al. (2020) applied GWR and other models to assess the evolution of air pollution during 2020 in urban contexts in China. Murgante et al. (2020) examined the geographic parallels between affected areas in the Po Valley, Italy, and Wuhan, China, where they found that pollution and land use play an important role in the distribution of COVID‐19 in both regions. Mansour, Al Kindi, Al‐Said, Al‐Said, and Atkinson (2020) used GWR to identify relationships between sociodemographic variables (population density, age groups, diabetics) and COVID‐19 in Oman.

GWR has the limitation of assuming that all the processes that are modeled operate on the same spatial scale, which led to the development of multiscale geographically weighted regression (MGWR) where modeling can be applied at different spatial scales (Fotheringham, Yang, & Kang, 2017; Maiti et al., 2020). Iyanda et al. (2020) used MGWR to identify significant sociodemographic and economic factors of COVID‐19 at the country level. Sannigrahi et al. (2020) estimated local spatial correlation coefficients between sociodemographic variables and COVID‐19 data in Europe. Mollalo, Vahedi, et al. (2020), on the other hand, used MGWR for a local‐level examination of the spatial non‐stationarity between 35 environmental, socioeconomic, topographic, and demographic variables in the United States.

Geographically weighted principal component analysis (GWPCA) is an extension of the classic principal component analysis that adapts the approach for use with geographic data by considering spatial heterogeneity (Fernández, Cotos‐Yáñez, Roca‐Pardiñas, & Ordóñez, 2018). For COVID‐19, Das et al. (2021) use GWPCA in developing their improved Index of Multiple Deprivations in districts of Kolkata, India, using housing conditions, household amenities, water, sanitation, and hygiene, asset possession, and gender disparity.

Nomura et al. (2020), in the prefecture of Fukuoka, Japan, used a modified test of spatial correlation coefficients (Clifford, Richardson, & Hémon, 1989) to relate the number of PCR‐confirmed COVID‐19 cases to a social network application that provides real‐time monitoring of self‐reported COVID‐19 symptoms. The highly significant correlation they report indicates the utility of crowdsourced data in spatial analysis, which can assist in policy evaluations such as emergency declarations.

3.1.2. Hotspots and clustering

Another common method utilized to study COVID‐19 is hotspot analysis (Table 1), which can facilitate targeted interventions by local, state, and federal agencies. Two of the most common tools to measure local clustering is the Getis–Ord Gi statistic and kernel density. The Gi statistic uses a global index to measure the level of spatial autocorrelation, that is, the degree to which objects or activities in a geographic unit are similar to other objects or activities in nearby geographic units (Goodchild, 1987). It has been used in numerous studies to understand the distribution of COVID‐19 cases, their spatial evolution over time, and for vulnerability and risk maps (e.g., Al‐Kindi et al., 2020; Barboza, Schiamberg, & Pachl, 2021; Maroko, Nash, & Pavilonis, 2020; Mollalo, Vahedi, Bhattarai, et al., 2020; Rahman, Islam, & Islam, 2020).

Kernel density estimation (KDE) is used to estimate data densities that do not have parametric statistical behaviors, that is, do not follow normal, binomial, or exponential distributions (Okabe, Satoh, & Sugihara, 2009). Rex, Borges, and Käfer (2020) applied KDE to identify areas with a high density of COVID‐19 cases; and Nian et al. (2020) examined public transport mobility data.

Following the review of spatial autocorrelation methods, the global univariate Moran's I method is the most widely used, although it has mainly been utilized with socioeconomic and COVID‐19 data. It can evaluate whether the data tend to be clustered, dispersed, or spatially random. It has been used to identify clustering of COVID‐19 (Alcântara et al., 2020; Baum & Henry, 2020; Kang, Choi, Kim, & Choi, 2020; Liu, Fang, & Gao, 2020; Wu et al., 2020) and facilitate the production of vulnerability and risk maps (e.g., Andrades‐Grassi et al., 2020; Gomes et al., 2020; Shariati, Mesgari, Kasraee, & Jahangiri‐Rad, 2020).

As a global indicator, Moran's I neglects the instability of local spatial processes, which led to the development of the local version of Moran's I (Anselin, 1995) which identifies both the spatial clustering of entities with similar values and the occurrence of divergent values. This latter version is also known as a local indicator of spatial association (LISA). Xie et al. (2020) and Wu et al. (2020) in China, Murgante et al. (2020) and Sun, Di, et al. (2020) in the United States, and Santana Juárez, Castañeda, Carillo, Carillo, and Alcántara (2020) in Mexico produced LISA cluster maps to analyze the characteristics of COVID‐19 at various spatial levels of aggregation.

Other prominent clustering techniques are:

• Self‐organizing maps, also known as Kohonen networks, a particular type of unsupervised neural network that performs spatial groupings of data based on similar behaviors. Melin et al. (2020) used a self‐organizing map to examine COVID‐19 cases in various countries.

• Spatial scan statistics (Kulldorf, 2018) to identify significant clusters of cumulative COVID‐19 cases (Andersen, Harden, Sugg, Runkle, & Lundquist, 2020).

• GeoMEDD (Curtis et al., 2020) is a new real‐time cluster detection methodology that provides indicators on the spatial evolution of the disease, based on access to various public sources that account for the location and timing of cases.

3.1.3. Interpolation and geostatistics

Here we focus on the interpolation methods used in works that address the spatial and spatiotemporal patterns of the pandemic, associated with atmospheric themes (pollution and climate), but also socioeconomic ones (Table 1). Many works pay special attention to the severe lockdowns that have been imposed in different countries, mainly in Europe and Asia.

• Inverse distance weighting interpolation: Saha, Barman, and Chouhan (2020) analyzed the impact of the COVID‐19 lockdown on community mobility in different Indian states and used IDW to show movement trends before and after the lockdown.

• Voronoi: Bherwani et al. (2020) used Bayesian probabilistic modeling to understand the relationship between COVID‐19 cases and population density in a given region together with GIS‐based Voronoi diagrams to identify high‐risk areas. Thiessen polygons delineate the risk zone boundaries.

• Geostatistics estimate values of phenomena in areas where the values are uncertain or missing based on the covariance (how two random variables change) or variogram (spatial dependence among stochastic (random) processes); see Moral García (2004). Spline interpolations are another technique using discrete data points to model a continuous variable. Sui et al. (2020) used a cubic spline technique to estimate the second‐by‐second speed of buses and taxis through vehicle GPS data devices; they discuss the potential change in emissions in a post‐COVID period.

• Kriging is probably the most commonly used geostatistical method and has been used in the COVID‐19 literature to predict climatic variables (Sarfo & Karuppannan, 2020; Wei et al., 2020; Yao, Zuo, et al., 2020; Yao, Pan, et al., 2020) and common atmospheric contaminants (Huang & Brown, 2021; Ran et al., 2020). In the latter case, research has used kriging to identify associations between air pollution and COVID‐19. A related technique is cokriging. Kerimray et al. (2020) use cokriging to map distributions of PM 2.5 and benzene in Almaty, Kazakhstan, in 2018–2019 and 2020, respectively. In particular, they examined the effect of government‐mandated lockdowns on concentrations of the pollutants.

3.1.4. Other spatial models

We mention other models used (Table 1).

• Spatial Markov (Dehghan Shabani & Shahnazi, 2020; Gayawan et al., 2020; Krisztin, Piribauer, & Wögerer, 2020) is a method where the probability of an event occurring depends only on the immediately preceding event, characterizing the spatial evolution of COVID‐19.

• Geodetector (Hu, Qiu, et al., 2020; Wu et al., 2020; Xie et al., 2020) is a method to detect stratified spatial heterogeneity and determine related factors, in this case with COVID‐19.

• Spatial agent‐based models (ABM), which are popular in the study of epidemics, simulate the spatial behavior of the agents (population in this case), identifying the areas with the greatest concurrence, the population mobility, and the interactions between them. This can re‐create and predict the appearance of COVID‐19 cases based on expected social behavior. It has been applied in urban environments (Gharakhanlou & Hooshangi, 2020; Prudhomme, Cruz, & Cherifi, 2020)

• The two‐step floating catchment area (2SFCA) method was used by Tao, Downs, Beckie, Chen, and McNelley (2020) to calculate the accessibility of COVID‐19 testing sites in Florida.

• The birthday paradox is a probability model which was used to estimate the risks of exposure of populations to contagion (Sun, Di, et al., 2020).

• Topological weighted centroid (TWC; Buscema, Della Torre, Breda, Massini, & Grossi, 2020) is a new algorithm used in Italy to analyze the evolution of the outbreak and to predict future epidemic processes.

• The spatial seemingly unrelated regression (SUR) equations model can be used when the structure of the data consists of cross‐sections for different time periods. Saez, Tobias, and Barceló (2020) analyzes the possible seasonality of COVID‐19.

• The spatiotemporal refined risk model was used by Michalak et al. (2020) for the unbiased identification of time periods with high risk of COVID‐19 in Poland.

3.1.5. Aspatial models in GIS

There are studies that use aspatial models and methods in conjunction with GIS, which is also important to mention due to its geographical contribution to the study of the pandemic (Table 1).

• Poisson regression has been applied for COVID‐19 in socioeconomic studies with health services infrastructure, especially in urban contexts (Ballesteros, Salazar, Sánchez, & Bolaños, 2021; DiMaggio, Klein, Berry, & Frangos, 2020; Saez et al., 2020; Yip, Huang, & Liang, 2020).

• Pearson correlation has been used with all kinds of variables, but especially with socioeconomic data, for spatiotemporal analysis, risk maps, health accessibility, and environmental repercussions due to the pandemic (Chatterjee et al., 2020; Li, Wang, Huang, & Lu, 2020; Marquès, Rovira, Nadal, & Domingo, 2020; Meraj et al., 2021; Tello‐Leal & Macías‐Hernández, 2020; Vergara‐Perucich, Correa‐Parra, & Aguirre‐Nuñez, 2020).

• Spearman and Kendall tests have been used with confirmed cases of COVID‐19 and socioeconomic variables, as well as with climate and air quality, to analyze the spatiotemporal evolution of the pandemic, mainly in urban contexts (Kuzmenko, Vasylieva, Vojtovič, Chygryn, & Snieška, 2020; Nakada & Urban, 2020; Pani, Lin, & RavindraBabu, 2020).

• Cox regression has been used to produce maps of vulnerability and risk in urban contexts (Fortaleza, Guimarães, de Almeida, Pronunciate, & Ferreira, 2020; Hutter et al., 2020).

• K‐means is an unsupervised clustering algorithm that partitions the data based on the closest mean. In this review we identify those works that used it to perform clustering as part of the process of a spatial analysis. For example, Lai, Charpignon, Ebner, and Celi (2020) used it to group US counties based on sociodemographic characteristics and COVID‐19 data; and Abdallah, Khafagy, and Omara (2020) for GPS location data.

• SIR models can add explicit geographical variables to study epidemic dynamics (Geng et al., 2020; O'Sullivan, Gahegan, Exter, & Adams, 2020; Thomas et al., 2020).

3.3.1. The atmosphere

Relevant atmospheric analyses (Table 3) tend to fall into one of two categories: climatic conditions or pollution. The spatiotemporal dynamics of climatic conditions and their relationship to COVID‐19 have been analyzed at many scales (Aman, Salman, & Yunus, 2020; El‐Magd & Zanaty, 2020; Ghahremanloo, Lops, Choi, & Mousavinezhad, 2020; Mishra et al., 2020; Pei, Han, Ma, Su, & Gong, 2020; Rahman, Azad, et al., 2020; Wyche et al., 2020). Spatiotemporal analyses of pollution levels during the pandemic tend to focus on the consequence of reductions in mobility and economic production (Filonchyk, Hurynovich, Yan, Gusev, & Shpilevskaya, 2020; Sandifer et al., 2020; Wu et al., 2020). Special attention is paid to countries that imposed lockdowns to control the pandemic, as well as to issues specifically affecting urban areas (Das et al., 2021; Guida & Carpentieri, 2020; Kerimray et al., 2020; Prunet, Lezeaux, Camy‐Payret, & Thevenon, 2020; Ran et al., 2020; Yao, Pan, et al., 2020). PM 2.5, PM 10, SO₂, NO₂, CO₂, O₃, O _x , and black carbon have been the most studied particles.

3.3.2. The lithosphere

On the Earth's surface, remote sensing (RS) and UAVs have been used for studies of land use change, with significant examples such as the case of Yao, Zuo, et al. (2020) who analyze the potential distribution of soya crops and how COVID‐19 has affected the soya market. Wang, Peng, et al. (2020) conducted a land use analysis of crops in China to evaluate whether the pandemic led to increased cultivation and land exploitation; their findings are relevant to the development of agricultural policy to improve food security. RS and UAVs also have applications for the real estate market as they are used for property inspection in a study that analyzes the consequences of COVID‐19 on the real estate sector (Renigier‐Biłozor, Zróbek, Walacik, & Janowski, 2020).

Several studies have analyzed pollutant levels in urban environments (Das et al., 2021; Prunet et al., 2020). Lighting is considered an indicator of economic recovery or recession, as in the case of several Chinese cities, where lighting levels are compared during peak closure with the same variable a year earlier (Elvidge, Ghosh, Hsu, Zhizhin, & Bazilian, 2020).

Research on the dynamics of human movement during the pandemic demonstrates important applications of RS and UAVs. Minetto, Segundo, Rotich, and Sarkar (2020) employ a deep learning technique to automatically detect objects, such as cars and aircraft, from satellite imagery. They suggest that the ability to automatically identify these objects in a time series of images will allow for temporal analysis of societal indicators (Minetto et al., 2020). Similarly, Wu et al. (2020) use deep learning to identify vehicles within Wuhan, China, from RS imagery and thus evaluate the effect of a transportation ban on the city. In a very different type of analysis, Okyere et al. (2020) use UAVs to monitor fishing boats and assess adherence the effects of physical distancing mandates and risks of exposure in the fishing sector.

There are also studies in areas with armed conflict, as in the case of Syria (Mobaied, 2020). Using RS and spatial models, the authors designed the “Risk of Vulnerability to COVID‐19 in War Zones Index” to identify areas that are vulnerable to the pandemic and thus help decision‐makers to limit risk and avoid and/or manage widespread infection.

3.3.3. The hydrosphere, with special attention to lockdowns

Addressing the concepts of rivers and pollution: changes in water quality as a result of a major lockdown have been evaluated in both the Ganges River (Garg, Aggarwal, & Chauhan, 2020) and the Sabarmati River (Aman et al., 2020) in India using Landsat 8 and Sentinel‐2 imagery, respectively. Similar methods have been applied to measure lake pollution, such as by Yunus, Masago, and Hijioka (2020) who use Landsat 8 to measure water turbidity in Vembanad Lake in India, finding an improvement in water quality as a result of lockdown.

In another hemisphere, Sentinel‐2 and Sentinel‐3 imagery was applied to detect a harmful algal bloom in salmonid aquaculture in Chile. The analysis technique, combined with rapid delivery of the high‐resolution satellite imagery, allowed for near real‐time monitoring and decision‐making when in‐situ sampling was restricted by a mandated lockdown (Rodríguez‐Benito, Navarro, & Caballero, 2020).

RS and UAVs are also highly relevant to topics related to ocean economic activities. For example, a UAV was used in Ghana to monitor water‐based activity during COVID‐19, providing solid scientific evidence as a basis for decision‐making in the artisanal fishing sector (Okyere et al., 2020).