Compound impact of COVID-19, economy and climate on the spatial distribution of global agriculture and food security

Abstract

As the COVID-19 pandemic continues to unfold around the world, the per unit area yield of the world's three major crops (i.e. maize, rice and wheat) decreased simultaneously for the first time in 20 years, and nearly 2.37 billion people faced food insecurity in 2020. Around 119–124 million people were pushed back into extreme poverty. Drought is one of the natural hazards that mostly affect agricultural production, and 2020 is one of the three warmest years on record. When the pandemic, economic recession and extreme climate change occur simultaneously, food crisis will often be exacerbated. Due to the limited research on the geographic modelling of crops and food security at the country level, we investigated the effects of the COVID-19 pandemic (COVID-19 incidence and mortality rate), economic (GDP and per capita GDP), climate (temperature change and drought), and their compound effects on three crop yields and food security in the world. On the basis of verifying the spatial autocorrelation, we used the global ordinary least squares model to select the explanatory variables. Then, geographically weighted regression (GWR) and multi-scale GWR (MGWR), were utilised to explore spatial non-stationary relationships. Results indicated that the MGWR was more efficient than the traditional GWR. On the whole, per capita GDP was the most important explanatory variable for most countries. However, the direct threats of COVID-19, temperature change and drought on crops and food security were small and localised. This study is the first to utilise advanced spatial methods to analyse the impacts of natural and human disasters on agriculture and food security in various countries, which can serve as a geographical guide for the World Food Organization, other relief agencies and policymakers to conduct food aid, health and medical assistance, financial support, climate change policy formulation, and anti-epidemic policy formulation.

Graphical abstract

1. Introduction

The United Nations Sustainable Development Goals proposes to eliminate all forms of poverty, hunger and malnutrition by 2030 (United Nations, 2015). However, the COVID-19 pandemic has made this significantly challenging. By the end of 2020, the COVID-19 pandemic has resulted in 81,475,053 confirmed cases and 1,798,050 deaths globally (World Health Organization (WHO), 2020, World Health Organization (WHO), 2021). The population lockdown caused by the epidemic has led to the emergence of global food security alarms (Galanakis, 2020). The surge in this infection may directly lead to labour shortages, resulting in a decline in crop yields in affected areas (Yaddanapudi and Mishra, 2022). In addition, the epidemic has directly disrupted global transportation, which may affect the import and export of grain and interrupt the world agricultural industry chain (Lin and Zhang, 2020). Trade restrictions and precautionary procurement due to the epidemic can cause a spike in global food prices and severe local food shortages, especially in low- and middle-income countries (Falkendal et al., 2021).

The COVID-19 pandemic has a devastating impact on the world's economy, leading to an unprecedented economic recession since the Second World War. Nearly all low- and middle-income countries were hit by the economic downturns that were twice as severe as any previously recorded in the same period (Ibn-Mohammed et al., 2021). The World Bank estimated that the COVID-19 pandemic in 2020 has increased the number of extremely poor people by 119 million to 124 million (Food and Agriculture Organization of the United Nations, 2021). A weak economic condition can negatively affect the local agriculture production. For example, it leads to low wages for agriculture labour, higher unemployment and continued shrinking of farm sizes (Potop, 2011). To prevent the spread of the epidemic, countries around the world have adopted a series of protective measures, including the reduction of travel and gatherings, which have caused the collapse of some hotels and restaurants, resulting in a 20 % drop in the price of agricultural commodities (Nicola et al., 2020). In addition, the economic slowdowns and downturns have sparked some panic amongst consumers and enterprises, distorted the usual consumption and production patterns and triggered a decline in people's affordability of healthy diets (McKibbin and Fernando, 2021).

Food insecurity prevalence has risen since 2014 after decades of decline driven by economic slowdown, climate change and conflict (World Meteorological Organization (WMO), 2021). In comparison with other sectors of the economy, agriculture is extremely vulnerable to overexposure to adverse natural hazards, especially climate-related ones. On the one hand, climate changes can cause short-term shocks, such as extreme weather events; on the other hand, it causes long-term pressures, such as rising temperatures and loss of biodiversity (Food and Agriculture Organization of the United Nations, 2021). Research has found that climate change reduces the total factor productivity of global agriculture, and global agriculture is increasingly vulnerable to the effect of ongoing climate change (Ortiz-Bobea et al., 2021).

The State of the Global Climate 2020 report emphasised that 2020 was a year of extreme weather and climate disruption. The global mean temperature in 2020 was about 1.2 °C higher than pre-industrial times, which was one of the three warmest years on record (World Meteorological Organization (WMO), 2021). For most crops and regions, investigating the sensitivity of crops to temperature change is one of the most important needs for assessing the impact of climate change on agriculture (Lobell and Burke, 2008). Liu et al. (2016) concluded that global temperature increase will decrease wheat yield according to grid- and point-based simulations and statistical regressions.

Drought is one of the most complex global natural hazards that has a serious influence on agriculture (Tian et al., 2018). Severe droughts can reduce water supply and cause a great loss of grain production (Zhao et al., 2017). With the development of remote sensing and earth observation technology, drought events can be detected over broad regions, and the ability of agricultural system to defend and recover from drought has been improved (Sur et al., 2015). However, the COVID-19 pandemic has brought unprecedented challenges to the agricultural system, and the impact of drought on agricultural production may become more complex (Mishra et al., 2021).

The pandemic, economic recession and climate disasters with increasing frequency and severity are main reasons for the decline of agriculture yield and global food insecurity (Food and Agriculture Organization of the United Nations, 2021). At the same time, they also form a compound disaster, which can seriously weaken the resilience of agricultural sector (Mishra et al., 2021). Accordingly, the main aims of this research are: (1) to understand the spatial pattern of the world's major crop yields and food insecurity prevalence after the epidemic outbreak; (2) to determine which social and natural factors will affect the world's major crop yields and food insecurity prevalence; (3) to investigate the severity and geographical differences in the impact of each influencing factor on world's major crop yields and food insecurity prevalence; and (4) to test for regional differences in the effects of compounding factors on world's major crop yields and food insecurity prevalence.

A previous study investigated the compound disruption of natural (drought) and human disasters (COVID-19 pandemic) on the global food supply chain (Mishra et al., 2021). However, this study only qualitatively analysed the potential impact of the epidemic and drought disasters on agriculture, and did not conduct quantitative research on the effects of disasters (e.g. impact on crop yields). Spatial regression models can be used to quantify the influence of the confounding effects of variables. Yaddanapudi and Mishra (2022) utilised geographically weighted regression (GWR) model to evaluate the potential impacts of drought and the COVID-19 pandemic, and their compound impact on three major crop yields amongst counties in the United States. They found that GWR can better capture the local scale variability than the traditional ordinary least squares (OLS) model. Although this study focused on identifying the potential impact on a local scale of crop yields, the study area was limited to the United States. Iyanda (2020) used global (OLS) and local (GWR and multi-scale GWR (MGWR)) models to examine three health and social determinants of the COVID-19 outbreak in 175 countries, and found that MGWR performed better than GWR. Mollalo et al. (2020) identified the effects of four explanatory variables on the spatial variability of COVID-19 incidence rates in the continental United States, and found that MGWR could explain the highest variations (adj. R²: 68.1 %) with the lowest Akaike information criterion with a correction (AICc) compared with other spatial models. Maiti et al. (2021) also proved that MGWR is the best spatial regression model in explaining the spatiotemporal impacts of the driving factors on COVID-19 incidences in the contiguous United States. In another study, Mansour et al. (2021) selected four sociodemographic determinants to explain the COVID-19 incidence rates in Oman, and the results suggested that MGWR can provide the best overall fit. Oshan (2020) found that MGWR provides a richer and more parsimonious quantitative representation of obesity rate determinants than GWR and OLS. Liu et al. (2020) applied MGWR to reveal the relationships between landscape patterns and the socio-economic factors of urbanisation in Shenzhen, and the results indicated that MGWR is better than GWR. Liu et al. (2021a) analysed the scale and spatial differences of the driving factors of China's industrial green development efficiency based on MGWR. Niu et al. (2021) utilised MGWR to identify the spatial heterogeneity of surface urban heat island drivers in 281 cities in China. Liu et al. (2021b) explored the influencing factors of intra-urban surface thermal environment and demonstrated the advantages of MGWR in capturing multi-scale spatial heterogeneity during the environmental responsive process. Forati et al. (2021) employed MGWR to examine the determinants of opioid overdose death and endorsed the superiority of MGWR in analysing epidemiology to guide epidemic responses to public health challenges.

Thus far, the combined effects of the COVID-19 epidemic, economy and climate change on global agricultural and food security at the country level have not yet been investigated nor has the MGWR model been introduced to assess the impacts of complex factors in this field. Therefore, to explicate global food crisis, we examined it at the country level. We utilised GWR and MGWR for modelling, which could explain the geospatial patterns varies, thereby enabling us to examine the differences of influencing factors. We initially compiled a data set of six candidate variables, considering epidemic, economic and climate change factors. Then, we calculated the Moran's I index to examine the spatial autocorrelation of the dependent variables (i.e. maize, rice and wheat yields and food insecurity prevalence). After determining the spatial autocorrelation of dependent variables, we utilised a global regression model (i.e. OLS) to select the explanatory variables associated with each dependent variable from six indicators to prepare for modelling. On this basis, local regression models (i.e. GWR and MGWR) were applied to capture local variation in each influencing factor, as well as the impacts of the compound factors on crop yields and food insecurity prevalence. The regression results could determine the dominant factor and the severity of composite factors in different countries. This research could help find the hotspots and causes affected by human and natural disasters around the world, which could aid governments in taking effective measures to deal with the food crisis.

2. Materials

This research has four dependent variables. We selected three crops, namely, maize, rice and wheat, as they have the largest production quantity and harvest area in the world. The prevalence of moderate or severe food insecurity is an estimate of the percentage of the population living in moderately or severely food insecure households. The national crop yields in 2020 and the prevalence of moderate or severe food insecurity (2018–2020, 3-year average) were derived from the Food and Agriculture Organization of the United Nations Statistical Database (FAOSTAT, https://www.fao.org/faostat/en/#home).

The National Earth System Science Data Centre, National Science and Technology Infrastructure of China (http://www.geodata.cn) provided the global administrative region boundary data. These data were established and improved by the United Nations Environment Programme, including the administrative boundaries of >200 countries (The boundaries and names of these administrative regions do not express any opinion of the authors on the legal status of any country, territory, city or area.).

The explanatory variables database contains six variables for three themes (Table 1 ). The global COVID-19 case and death data in 2020 were derived from the World Health Organization (https://covid19.who.int). We calculated the population of each country based on the unconstrained global mosaics 2020 (1-km resolution) population counts data provided by the WorldPop (www.worldpop.org, School of Geography and Environmental Science, University of Southampton; Department of Geography and Geosciences, University of Louisville; Departement de Geographie, Universite de Namur) and the Centre for International Earth Science Information Network, Columbia University (2018). The incidence and mortality rate of COVID-19 in various countries were obtained by dividing the COVID-19 case and death data by the population count in each country.

Table 1.

Explanatory variables and data sources.

Theme	Variable name	Data name	Data description	Data source	Acquisition date
COVID-19	(1) COVID-19 incidence (2) COVID-19 mortality	(a) Cumulative cases (b) Cumulative deaths (c) Population counts	(a) Cumulative confirmed cases reported to WHO to date (b) Cumulative confirmed deaths reported to WHO to date (c) Consistent 1 km resolution population count datasets for all countries of the world for each year 2000–2020	(a-b) WHO (c) World POP	(a-b) 2020-12-31 (c) 2020
Economy	(1) GDP (2) Per capita GDP	(a) GDP in USD (b) Per capita GDP in USD	(a-b) The annual national accounts tables from 1970 onwards for >200 countries of the world	(a-b) UNSD AMA	(a-b) 2020
Climate	(1) Temperature change (2) Drought indices	(a) Temperature change	(a) The mean surface temperature change by country with annual updates since 1961	(a) NASA GISS	(a) 2020
		(b) scPDSI	(b) scPDSI data for the global land surface, spanning the period 1901–2021 with monthly resolution. The spatial resolution is 0.5 × 0.5 degrees	(b) UEA CRU	(b) 2020 (monthly)

We used GDP and per capita GDP in US dollars (USD) to measure the economic level around the world. The FAOSTAT Macro Indicators database provides these country-level macroeconomic indicators, which were collected from the United Nations Statistic Division National Accounts Estimate of Main Aggregates (UNSD AMA) database (http://unstats.un.org/unsd/snaama/Introduction.asp).

The FAOSTAT country-level temperature change domain disseminates statistics based on the Global Surface Temperature Change data distributed by the National Aeronautics and Space Administration Goddard Institute for Space Studies (NASA GISS) (https://www.giss.nasa.gov).

Given the differences in water supply and demand around the world, standards of drought are difficult to harmonise (Dai et al., 2004). The self-calibrating Palmer Drought Severity Index (scPDSI) has been widely used in drought-related research, which was calculated based on time series of precipitation and temperature, together with fixed parameters related to the soil/surface characteristics at each location. (Wells et al., 2004). It has a similar variation range under different climatic conditions, which makes it suitable for comparing the availability of relative moisture in different regions (van der Schrier et al., 2013). The global monthly scPDSI data product (excluding Antarctica) is from the Climate Research Unit of the University of East Anglia (UEA CRU). We calculated the country-level annual average scPDSI based on the 0.5° × 0.5° spatial resolution monthly data set (https://crudata.uea.ac.uk/cru/data/drought).

3. Methodology

Fig. 1 demonstrates the methodological procedure steps and analysis stages of this research. The global and local spatial regression analysis has been carried out through three separate models.

Fig. 1.

Flowchart of the research methodological procedure steps and analysis stages.

3.1. Model 1: ordinary least squares (OLS)

The OLS model is a classical regression method, which has been widely used to explore the relationship between a group of explanatory variables and a dependent variable (Pohlman and Leitner, 2003). In this research, the relationship between a dependent variable and a set of explanatory variables using the OLS model is presented as a line of best fit, which can be written as follows:

$$y_i={\beta }_0+\beta x_i+{\epsilon }_i,$$ (1)

where $y_i$ represents each dependent variable (maize yield, rice yield, wheat yield and the food insecurity prevalence) in country i; ${\beta }_0$ is the intercept; $\beta$ is the vector of regression coefficients; and $x_i$ signifies the explanatory variables (i.e. COVID-19 incidence rate, COVID-19 mortality rate, GDP, per capita GDP, temperature change and scPDSI). Given that there exists a collinearity between COVID-19 incidence and mortality rate, they were not used as explanatory variables in the models of this study at the same time; and ${\epsilon }_i$ denotes the random error (Anselin and Arribas-Bel, 2013).

3.2. Model 2: geographically weighted regression (GWR)

According to the First Law of Geography, all objects on the geographical surface are interrelated, and near objects are more related to each other (Tobler, 1970). However, the OLS model assumes a stable and constant spatial relationship, which does not consider the spatial heterogeneity. To make the parameters alter over space, Brunsdon et al. (2010) introduced the GWR model based on kernel regression. The GWR model can be denoted by:

$$y_i={\beta }_{i0}+{\sum }_{j=1}^m{\beta }_{\mathit{ij}}{x}_{\mathit{ij}}+{\epsilon }_i,i=1,\dots ,n,$$ (2)

where in country i, $y_i$ is the dependent variable; ${\beta }_{i0}$ is the intercept; ${\beta }_{\mathit{ij}}$ is the value of the jth regression coefficient; $x_{\mathit{ij}}$ is the jth explanatory variable; and ${\epsilon }_i$ is the random error (Fotheringham and Oshan, 2016). The parameter estimates in matrix form for each explanatory variable in each country is as follows:

$$\widehat{\beta }\left(i\right)={\left(X^{\prime }W\left(i\right)X\right)}^{-1}{X}^{\prime }W\left(i\right)y,$$ (3)

where $\widehat{\beta }\left(i\right)$ is a j × 1 vector of the parameter estimates in country i; $X$ is an n × j matrix of explanatory variables; $W\left(i\right)$ = diag[$w_1\left(i\right)$, …, $w_n\left(i\right)$] is the diagonal weight matrix, which is weighted according to the distance between each observation and position i; and y is an n × 1 observation vector of the dependent variable. To calculate $W\left(i\right)$, a kernel function should be applied to emphasise the closer observations in space (Fotheringham and Oshan, 2016). The model type used in this study is Gaussian, the spatial kernel is adaptive bisquare, the bandwidth searching method is golden section, and the optimisation criterion is Akaike information criterion with a correction (AICc).

3.3. Model 3: multiscale geographically weighted regression (MGWR)

The classical GWR assumes that all the spatial changes processed in the model run on the same spatial scale. However, in many cases, when a phenomenon involves many spatial processes with various spatial scales, the fixed spatial scale is invalid (Fotheringham et al., 2017). MGWR allows the conditional relationship between dependent and explanatory variables to vary on different spatial scales (Yang et al., 2014). The form of MGWR can be written as

$$y_i={\sum }_{j=0}^m{\beta }_{\mathit{bwj}}\left(u_i,v_i\right)x_{\mathit{ij}}+{\epsilon }_i,$$ (4)

where $y_i$ represents the dependent variable in country i; ${\beta }_{\mathit{bwj}}$ indicates the coefficient describing the relationship between $y_i$ and explanatory variable $x_{\mathit{ij}}$ around the location $\left(u_i,v_i\right)$; bwj is the bandwidth used to calibrate the jth conditional relationship; and ${\epsilon }_i$ is the random error (Fotheringham et al., 2017). MGWR has less restrictive assumption compared with GWR, which can accurately reflect spatial heterogeneity. In addition, MGWR can diminish bias in parameter estimates and reduce collinearity (Oshan et al., 2019).

4. Experiments and results

4.1. Temporal and spatial patterns of three major crop yields and food insecurity prevalence

As shown in Fig. 2 , in the past 20 years, the yield of the world's three major crops has risen steadily in fluctuations. However, in 2020, these crop yields decreased simultaneously in one year (maize, rice and wheat decreased by 580, 223 and 688 hg/ha, respectively) for the first time since 2000. Although the prevalence of moderate or severe food insecurity in the total population has increased slowly since 2014, the estimated increase prevalence in 2020 (4.1 %) is almost equal to the sum of the previous five years (4.2 %).

Fig. 2.

Yield of three major crops (2000−2020) and prevalence of moderate or severe food insecurity (2014–2020) in the world.

According to the statistics of the United Nations, in 2020, 44 of the 166 countries growing maize had lower maize yield than recorded last year; 54 of 115 countries that planted rice recorded have a decline in rice yield; and wheat yield declined in 50 of the 124 wheat growing countries. At the same time, the prevalence of moderate or severe food insecurity risen in 81 of 127 countries recorded. Fig. 3 shows the spatial distribution of the world's three major crop yields and moderate or severe food insecurity prevalence at the country level. In 2020, North America had the highest maize yield with an average of 107,474 hg/ha, and Middle Africa had the lowest maize yield (10,776 hg/ha); Australia and New Zealand had the highest rice yield (100,312 hg/ha), and Middle Africa had the lowest rice yield (11,161 hg/ha); Western Europe had the highest wheat yield (71,310 hg/ha), and Melanesia had the lowest wheat yield (7500 hg/ha). The highest prevalence of moderate or severe food insecurity prevalence in 2018–2020 was in Western Africa, with an average of 68.3 %, and the lowest was in Western Europe (3.7 %).

Fig. 3.

Spatial distribution of the world's three major crop yields (2020) and moderate or severe food insecurity prevalence (2018–2020, 3-year average) at the country level.

We calculated the Moran's I index to examine the spatial autocorrelation of the dependent variables across the study areas. As listed in Table 2 , the global Moran's I of maize, rice and wheat yields and food insecurity prevalence are 0.153, 0.164, 0.452 and 0.724, respectively, which are considerably higher than the expected results. Moreover, the critical values (z-scores) are higher than 2.58, and their p-values are lower than 0.01, which implies that they all passed the significance test at 99 %. In conclusion, these values indicate a significantly positive global spatial autocorrelation of each dependent variable across the entire study area, and the probability that the clustered pattern of each dependent variable can be the result of random chance is <1 %.

Table 2.

Global Moran's I of three crop yields (2020) and moderate or severe food insecurity prevalence (2018–2020, 3-year average).

Variable	Moran's index	Expected index	z-score	p-value	Pattern
Maize yield	0.153	−0.006	10.340	0.000	Clustered
Rice yield	0.164	−0.009	7.578	0.000	Clustered
Wheat yield	0.452	−0.008	15.124	0.000	Clustered
Food insecurity prevalence	0.724	−0.008	20.694	0.000	Clustered

In addition, we utilised the local Moran's I to depict the local spatial correlation, which shows the local spatial clustering pattern about the variation of each dependent variable. High–high clusters of maize yield are mainly distributed in Western Europe, whereas, low–low clusters are mainly located in Southern Africa (Fig. 4(a)). High–high clusters of rice yield are located in Russia, China and Japan; whereas low–low clusters are distributed in Africa (Fig. 4(b)). High–high clusters of wheat yield are distributed in Western Europe, whereas low–low clusters are located in Africa (Fig. 4(c)). Most countries in Africa are suffering from food insecurity. On the contrary, low food insecurity prevalence is concentrated in Western Europe (Fig. 4(d)).

Fig. 4.

Local Moran's I clusters of the world's three crop yields (2020) and moderate or severe food insecurity prevalence (2018–2020, 3-year average) at the country level.

4.2. Selection of factors influencing three major crop yields and food insecurity prevalence in the world

Maps of six selected explanatory variables in 2020 are shown in Fig. 5 . The spatial distribution of COVID-19 incidence and mortality rate is similar. Most countries in Europe and North and South America have experienced severe epidemic, of which 19 countries have an incidence rate exceeding 5 % and 16 countries have a mortality of >0.1 % (e.g. Andorra, Belgium, Czech Republic, Liechtenstein, Montenegro, San Marino, Slovenia and the United States). GDP has decreased in 144 of 211 countries recorded, and per capita GDP has decreased in 156 of 211 countries compared with last year. Global GDP and GDP per capita have declined for the first time in five years. GDP and per capita GDP are inconsistent in geographic pattern. For example, China and India have a high GDP and low per capita GDP. The average temperature of 214 countries in the world has increased in 2020. Particularly, the temperature in Russia, Estonia, Latvia, Belarus, Lithuania, Finland, Moldova and Ukraine has increased by >3 °C. Amongst 232 countries, 128 have suffered from drought in 2020, including 32 countries with moderate drought (−3 < scPDSI ≤ −2) and 20 countries with severe or extreme drought (scPDSI ≤ −3; van der Schrier et al., 2013).

Fig. 5.

Spatial distribution of six explanatory variables (COVID-19 incidence rate, COVID-19 mortality rate, GDP, per capita GDP, temperature change and scPDSI) in 2020 at the country level across the world.

As shown in Table 3 , we selected the explanatory variables that could explain the variation of each dependent variable at the country level around the world according to the OLS model. The p-value reveals the statistical significance between the explanatory and dependent variables. The coefficient estimates are the mixture of positive and negative values. The results of this non-spatial model indicated that the combination of COVID-19 incidence and per capita GDP significantly has influenced maize yield, and per capita GDP is more influential. COVID-19 mortality, GDP and per capita GDP have significantly influenced rice yield, and COVID-19 mortality is the most influential variable, followed by GDP and per capita GDP. COVID-19 incidence, GDP, per capita GDP and temperature change have significantly influenced wheat yield, and per capita GDP is the most influential variable, followed by GDP, COVID-19 incidence and temperature change. COVID-19 incidence, per capita GDP and scPDSI have significantly influenced food insecurity prevalence, and per capita GDP is the most influential variable, followed by COVID-19 incidence and scPDSI. Tests for multicollinearity were also performed using the variance inflation factor (VIF) as diagnostic data to determine multicollinearity. The results show that multicollinearity was not an issue as the VIF for each explanatory variable was below 10.

Table 3.

Summary statistics of global OLS model.

Dependent variable	Explanatory variable	Coefficient	St. Error	T-statistic	p-value	VIF
Maize yield	(1) COVID-19 incidence	0.224	0.083	2.711	0.007***	1.458
	(2) Per capita GDP	0.304	0.083	3.675	0.000***	1.458
Rice yield	(1) COVID-19 mortality	0.291	0.083	3.505	0.000***	1.336
	(2) GDP	0.254	0.086	2.945	0.003***	1.421
	(3) Per capita GDP	0.241	0.092	2.629	0.009**	1.736
Wheat yield	(1) COVID-19 incidence	0.152	0.091	1.680	0.093*	1.532
	(2) GDP	0.164	0.075	2.197	0.028**	1.080
	(3) Per capita GDP	0.384	0.086	4.468	0.000***	1.402
	(4) Temperature change	0.141	0.084	1.680	0.093*	1.337
Food insecurity prevalence	(1) COVID-19 incidence	−0.250	0.086	−2.925	0.003***	1.573
	(2) Per capita GDP	−0.439	0.079	−5.582	0.000***	1.328
	(3) scPDSI	0.185	0.075	2.465	0.014**	1.216

Notes: ***, ** and * denote significance at the 1 %, 5 % and 10 % level, respectively.

4.3. Comparison of OLS, GWR and MGWR in modelling three major crop yields and food insecurity prevalence in the world

OLS model was used to perform a global analysis of the explanatory variables. GWR and MGWR models were applied to measure the local spatial variations. According to Table 4 , the performance of the OLS model is the weakest amongst the three models apparently, with the lowest R² and Adj. R² and highest AICc and residual sum of squares. For maize yield, the R² (Adj. R²) of GWR and MGWR increases to 0.316 (0.280) and 0.310 (0.280), the AICc decreases to 433.016 and 431.361, and the residual sum of squares decreased to 114.907 and 115.852 compared with the results of OLS. For rice yield, the R² (Adj. R²) of GWR and MGWR increases to 0.469 (0.411) and 0.500 (0.450), the AICc decreases to 286.114 and 277.024, and the residual sum of squares decreased to 62.139 and 58.462, respectively. For wheat yield, the R² (Adj. R²) of GWR and MGWR increases to 0.417 (0.384) and 0.472 (0.420), the AICc decreases to 303.722 and 301.921, and the residual sum of squares decreased to 72.823 and 65.970, respectively. For Food insecurity prevalence, the R² (Adj. R²) of GWR and MGWR increases to 0.729 (0.681) and 0.702 (0.669), the AICc decreases to 233.417 and 228.330, and the residual sum of squares decreased to 33.013 and 36.316, respectively. Overall, the MGWR model shows higher R² (Adj. R²) and lower AICc and residual sum of squares for rice and wheat yields. This result suggests that MGWR better captures the local scale variability of the two dependent variables than GWR does. For maize yield and food insecurity prevalence, the GWR model shows slightly higher R² and lower residual sum of squares, which reveals that GWR has higher measures of goodness of fit. However, the AICc values of the MGWR model are all lower than those of the GWR, indicating that MGWR is more parsimonious.

Table 4.

Comparison of goodness-of-fit measures for OLS, GWR and MGWR models.

Model	Dependent variable	R2	Adj. R2	AICc	Residual sum of squares
OLS	(1) Maize yield	0.219	0.214	441.469	131.275
	(2) Rice yield	0.318	0.306	295.652	79.823
	(3) Wheat yield	0.350	0.334	311.388	81.248
	(4) Food insecurity prevalence	0.446	0.437	282.431	67.544
GWR	(1) Maize yield	0.316	0.280	433.016	114.907
	(2) Rice yield	0.469	0.411	286.114	62.139
	(3) Wheat yield	0.417	0.384	303.722	72.823
	(4) Food insecurity prevalence	0.729	0.681	233.417	33.013
MGWR	(1) Maize yield	0.310	0.280	431.361	115.852
	(2) Rice yield	0.500	0.450	277.024	58.462
	(3) Wheat yield	0.472	0.420	301.921	65.970
	(4) Food insecurity prevalence	0.702	0.669	228.330	36.316

The bandwidths comparison of GWR and MGWR is shown in Table 5 . MGWR can directly reflect the differential effect scales of different variables, while classical GWR can only reflect the average of the effect scales of all variables. The GWR regression result of each variable was fixed. In contrast, MGWR results showed that the bandwidth of different variables varied greatly. For example, the explanatory variables' bandwidths of food insecurity prevalence were 61 (COVID-19 incidence), 43 (Per capita GDP) and 119 (scPDSI), respectively, indicating that scPDSI has the largest effect scale, and the coefficients are spatially smoother with minimal spatial heterogeneity; Per capita GDP has the smallest effect scale with greatest spatial heterogeneity amongst the three explanatory variables. Overall, MGWR performed better than GWR because the main advantage of MGWR over the fixed scale of classical GWR is that the different variables are scaled differently and the bandwidth of each variable is specific.

Table 5.

GWR and MGWR bandwidth comparison.

Dependent variable	Explanatory variable	The bandwidth of GWR	The bandwidth of MGWR
Maize yield	(1) COVID-19 incidence	89	115
	(2) Per capita GDP	89	90
Rice yield	(1) COVID-19 mortality	63	116
	(2) GDP	63	55
	(3) Per capita GDP	63	49
Wheat yield	(1) COVID-19 incidence	120	49
	(2) GDP	120	120
	(3) Per capita GDP	120	124
	(4) Temperature change	120	106
Food insecurity prevalence	(1) COVID-19 incidence	47	61
	(2) Per capita GDP	47	43
	(3) scPDSI	47	119

4.4. Local effects of each driving factor on three major crop yields and food insecurity prevalence in the world derived from GWR and MGWR

Fig. 6 shows the GWR and MGWR coefficient results for maize yield. The maize yield has a stronger spatial relationship with per capita GDP in 2020 than that with COVID-19 incidence. Both models fail to capture the obvious threat of the COVID-19 incidence to maize yield, as the coefficient estimates are positive values overall. Per capita GDP shows almost similar patterns in explaining the geographical distribution of maize yield at the country level in GWR and MGWR. Generally, per capita GDP is positively correlated with maize yield at the country level around the world. Many countries in Southern and Eastern Africa (e.g. Comoros, Djibouti, Lesotho, Madagascar, Malawi, Mauritius, Mozambique, Somalia, South Africa and Swaziland) with lower maize yields show stronger positive relationship with per capita GDP in GWR and MGWR models compared with other more affluent countries.

Fig. 6.

Effects of COVID-19 incidence (above) and per capita GDP (below) in describing maize yield utilising GWR (left) and MGWR (right) models at the country level across the world.

As shown in Fig. 7(a), COVID-19 mortality negatively affects the rice yield of China, Russia and many countries in Southeast Asia. However, in other countries, COVID-19 mortality is positively related to rice yield in GWR. According to the small coefficient values shown in Fig. 7(b), MGWR represents poor performance in exploring the geographic impact of COVID-19 mortality on rice yield. GDP and per capita GDP show similar results in describing the spatial distribution of rice yield at the country level in GWR and MGWR. GDP is a substantial factor in explaining the geographic distribution of rice yield in many countries in Western and Southern Africa (Fig. 7(c) and (d)). Per capita GDP is an influential factor in describing the rice yield across countries in Eastern and Southern Africa (Fig. 7(e) and (f)).

Fig. 7.

Effects of COVID-19 mortality (above), GDP (middle) and per capita GDP (below) in describing rice yield utilising GWR (left) and MGWR (right) models at the country level across the world.

The coefficient estimates in Fig. 8 illustrate that the per capita GDP is the most influential variable, followed by GDP, COVID-19 incidence and temperature change. In MGWR, COVID-19 incidence has a negative effect on the wheat yield of North America, Western Asia and several countries in South America and Northern Africa (Fig. 8(b)). However, the threat of COVID-19 incidence on wheat yield is not captured in the GWR model (Fig. 8(a)). As shown in Fig. 8(c) and (d), GDP is an influential factor in explaining the spatial distribution of wheat yield across Africa, specifically in Luxembourg, Netherlands, Switzerland, Belgium, Italy. Per capita GDP is a significant regressor in describing the variation of wheat yield across North and South America (Fig. 8(e) and (f)). Small coefficient estimates are found in GWR and MGWR, indicating a weak relationship between temperature change and wheat yield (Fig. 8(g) and (h)).

Fig. 8.

Effects of COVID-19 incidence (a and b), GDP (c and d), per capita GDP (e and f) and temperature change (g and h) in describing wheat yield utilising GWR (left) and MGWR (right) models at the country level across the world.

Fig. 9 shows the GWR and MGWR coefficient results for food insecurity prevalence. As shown in Fig. 9(a) and (b), COVID-19 incidence demonstrates different patterns in describing the spatial distribution of food insecurity prevalence at the country level in GWR and MGWR. Both models reveal that in Australia, Bangladesh, Cambodia, Fiji, Indonesia, Japan, Kiribati, Laos, Malaysia, Mongolia, Myanmar, Nepal, New Zealand, Philippines, Russia, Singapore, South Korea, Thailand and Vanuatu, the food insecurity prevalence increases with the COVID-19 incidence rate. Fig. 9(c) and (d) illustrate that per capita GDP is negatively correlated with the food insecurity prevalence in almost all countries around the world, especially in Africa and South America, which have a lower GDP. Fig. 9(e) shows that scPDSI is negatively correlated with food insecurity prevalence in North American countries, Peru and Ecuador, indicating that drought will increase food insecurity prevalence in these countries. The smaller positive value estimates in Fig. 9(f) indicate a weaker positive relationship between drought and food insecurity prevalence in MGWR compared with that in GWR.

Fig. 9.

Effects of COVID-19 incidence (above), per capita GDP (middle) and scPDSI (below) in describing food insecurity prevalence utilising GWR (left) and MGWR (right) models at the country level across the world.

4.5. Compound impact of driving factors on three major crop yields and food insecurity prevalence in the world derived from GWR and MGWR

Fig. 10 reveals the geographical distributions of local R² in GWR and MGWR models. Both models perform the best prediction on food insecurity prevalence, followed by rice, wheat and maize yields. As shown in Fig. 10(a) and (b), for maize yield, most of countries in Africa have a high local R², indicating a decent performance of both models in these countries. By contrast, the local R² is low in Russia, South America and most countries of Eastern Europe, indicating the poor prediction results of the models in these countries. For rice yield, both models perform well in most countries of Southeast Asia and Oceania. However, in Central and Western Asia countries, both models perform poor results (Fig. 10(c) and (d)); For wheat yield, both models have superior performance in Canada, Morocco, Algeria, Tunisia, Libya and many countries of Western Europe and Central Asia. In comparison with other countries, both models perform poorly in New Zealand, New Caledonia and Australia (Fig. 10(e) and (f). For food insecurity prevalence, countries in North America, South America and Northern Africa have a very high local R², showing great performance of the models in these areas. Conversely, both models have poor prediction results in Asian and European countries. Overall, the local goodness of fit of each dependent variable are almost similar in the spatial patterns in GWR and MGWR models, and the performance of MGWR is slightly better than that of GWR.

Fig. 10.

Geographic distribution of local R² of GWR and MGWR models for dependent variables associated with significant explanatory variables at the country level across the world.

5. Discussion

The year 2020 was an unprecedented time for human and the earth. In this year, we experienced a global pandemic that has never been seen in more than a century; a deep global economic recession, facing the greatest challenge since the new century; and the highest global temperature in the past millennium (World Meteorological Organization, 2021; International Monetary Fund, 2020). The per unit area yield of the three major crops in the world, namely, maize, rice and wheat, decreased simultaneously for the first time in 20 years. At the same time, the growth and prevalence rates of global food insecurity have reached the highest level on record.

To assess the location of disaster hotspots and clusters and understand the impacts of the compound disasters on world's major crops and food insecurity, this study investigated the potential influence of COVID-19, economy and climate on three major corps yields and food insecurity prevalence of the world. Due to the spatial heterogeneity of agricultural management and production conditions in different countries (Mok et al., 2014), the separate and compounding effects of COVID-19 epidemic, economic and climate factors on crop yields and food insecurity vary across countries.

For maize yield, both GWR and MGWR perform best in Africa, that is, the regions where the combined factors (COVID-19 incidence and per capita GDP) have the most severe effect on maize yield are African countries. COVID-19 incidence has almost no negative impact on maize yields. As the research mentioned, the pandemic has not seriously cut agriculture production (Bai et al., 2022). Africa has a low incidence rate of COVID-19 in 2020, and its rapidly growing population also brings a large labour force to agricultural production (Epule et al., 2022). Eastern Africa, with the lowest per capita GDP in the world, has the strongest relationship between per capita GDP and maize yield. This is due to the higher reliance of the agricultural sector on economic income in low- and middle-income countries with limited resources (Hatefi et al., 2022).

The impact of the COVID-19 mortality on rice yield reflects a stronger impact on Western countries and a weaker impact on Eastern countries in GWR model. In China, for example, as the first developing country heavily struck by the COVID-19 pandemic, the government has adopted a series of epidemic prevention policies, so the incidence rate of the epidemic may have a negative impact on rice production (J. Wu et al., 2023). However, MGWR did not capture this negative impact. The rice production is dominated by poor smallholder farmers (H. Wu et al., 2023). Thus, economic factors are of paramount importance for rice production. The impact of GDP on rice yield shows the strongest correlation in Africa where GDP is lower. The correlation between per capita GDP and rice yield is also the strongest in Eastern Africa. For rice yield, the combined factors (COVID-19 mortality, GDP and per capita GDP) have the most severe effect in Southeast Asian and Oceanian countries.

COVID-19 does not pose a major threat to wheat production in GWR model. While MGWR captured the negative impact of COVID incidence on wheat yield in a very small number of countries, such as the United States, COVID-19 could potentially affect wheat yield (Yaddanapudi and Mishra, 2022). Per capita GDP has the least impact on wheat yield compared with maize and rice yields. GWR and MGWR do not capture the significantly negative impact of temperature change on crop yields, which corroborates previous findings that declaimed continuing technology trends will offset much of the impact of climate change on cereal yields (Gammans et al., 2017). The combined factors (COVID-19 incidence, GDP, per capita GDP and temperature change) have the most severe effect in Western European and Central Asian countries for wheat yield.

For food insecurity prevalence, the regions where the combined factors (COVID-19 incidence, GDP, per capita GDP and temperature change) have the most severe effect are North American, South American and Northern African countries. The impact of the COVID-19 incidence on food insecurity prevalence is notable in Russia and many countries of Central Asia, Western Asia, Southeast Asia and Oceania. COVID-19 was the main reason of the significant increase in the average food price in countries with a high number of epidemic cases. For example, in Russia and Australia, the food price index has risen almost every month in 2020, which endangers food security (Bai et al., 2022). Food insecurity prevalence is higher where per capita GDP is lower, especially in Southern African countries. Our results are similar with the findings in previous studies, which highlight that the economic slowdown since the outbreak of the pandemic has exacerbated food insecurity in many developing countries due to the income losses (Elleby et al., 2020). The economic downturn caused by the global epidemic is one of the reasons for an increase of 320 million in the number of people that did not have access to adequate food in only one year, affecting almost all low- and middle-income countries (Food and Agriculture Organization of the United Nations, 2021). In GWR, scPDSI shows a negative relationship with food insecurity prevalence in North American countries. Drought aggravates the universality of food insecurity, because 2020 was regarded as the above-average drought year; and extreme, flash and interannual droughts have gradually become the new normal in North America (Overpeck and Udall, 2020).

This is the first study using advanced geographic technology to examine the country-level influencing factors of the world's three major crops and food security. The robust geospatial method provided us with a rich spatial quantitative representation of the selected variables. The spatial modelling revealed the hot spots affected by individual and complex factors, which provided us with a profound understanding of the food crisis. The research results have policy implications that can serve as a space guide for the World Food Organization and other international relief agencies to provide food aid or other intervention to countries suffering from serious compound disasters. The spatial impact intensity results of the selected variables in this study will alert organizations such as the World Health Organization and the World Meteorological Organization to provide health and medical assistance and suggestions on climate change policy formulation for specific countries. Countries suffering from severe impacts can also recognize the main factors leading to their food crisis through this study, and timely adjust their agricultural, anti-epidemic, economic and climate change policies. Our models do not capture the obvious threat of temperature change and drought to agriculture and food security, and other climate explanatory variables should be further considered in subsequent studies. Finally, as an example, this study provides ideas for analysing the determinants of countries with severe food crisis. The MGWR, introduced as a parsimonious model, performs well in explaining the spatial variability of variables. In comparison with using GWR alone, MGWR can solve the modifiable areal unit problem, which is one of the most challenging problems related to the use of areal data (Fotheringham and Wong, 1991).

6. Conclusion

This study aims to better understand the impacts of epidemic, economy and climate change on major world staple crops and food security and help managers and planners make timely policies to deal with compound disasters and mitigate food crises. We used the incidence and mortality rate of COVID-19 in each country to reflect the severity of the epidemic. We used GDP and GDP per capita to measure the economic level of each country. We also obtained the country-level temperature change data, and used scPDSI as the drought index. On this basis, we analysed the spatial characteristics of three crop yields and food insecurity prevalence in the world after the outbreak, and we then applied local regression models (GWR and MGWR) to determine the significance and direction of the driving factors.

The performance of GWR and MGWR models shows high spatial consistency. On the whole, MGWR is more parsimonious and shows better performance. Our results indicate that the compound crisis posed a challenge to humanity worldwide, with a wide variation in the impacts of complex factors on agriculture and food security across the world in the year following the outbreak of the pandemic. Generally, crop yields and food insecurity prevalence are mostly affected by per capita GDP. The impacts of COVID-19 and climate on agriculture and food security are local. On the basis of the local scale variations captured by GWR and MGWR models, COVID-19 incidence and per capita GDP can explain >40 % of the changes in maize yield of African countries. COVID-19 mortality, GDP and per capita GDP can explain >50 % of the changes in rice yield of South Asian and Oceanian countries. COVID-19 incidence, GDP, per capita GDP and temperature change can explain >45 % of the changes in wheat yield of North American, Western European, Western Asian, Central Asian and Northern African countries. COVID-19 incidence, GDP, per capita GDP and temperature change can determine the reason for the changes of food insecurity prevalence of >60 % in North American, South American and Northern African countries.

These results will help identify areas where agricultural production and food security are threatened by various factors under complex disasters, and help policymakers formulate mitigation measures. There are still some limitations. First, there is an inevitable problem of omitted variables in the current model, which is largely due to the complexity and diversity of factors affecting crop yields and food security. Not only have measures to contain the spread of the epidemic resulted in an unprecedented economic recession, but also other important drivers are behind recent setbacks in agriculture and food security. These include conflict and violence in many countries as well as various climate-related disasters (not only temperature rise and drought) all over the world. However, such drivers as conflicts and some unusual natural disasters exist in individual countries, so they were not added in the global modelling of this study. Second, the research time should be extended. The current round of global food crisis has not been lifted. Since the outbreak of the Russia-Ukraine conflict in February 2022, food production and trade have been severely impacted, and the current level of global hunger is at a new high. Accordingly, future research should take more driving factors into consideration and extend the research time.

CRediT authorship contribution statement

Peiwen Yao: Conceptualization, Data curation, Formal analysis, Writing – review & editing. Qilong Wu: Writing – review & editing. Hong Fan: Funding acquisition, Writing - review & editing. Jiani Ouyang: Writing – review & editing. Kairui Li: Writing – review & editing.

Declaration of competing interest

We acknowledge that our manuscript is original and it is not submitted for review in another journal. We have provided references for the data set used in our study. 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: Jacopo Bacenetti

Data availability

Data will be made available on request.