## Main

Agrochemicals such as synthetic fertilizers and pesticides have together made a remarkable contribution to food security in the last 50 years1. Notwithstanding the increased food availability2, the unpreventable ubiquity of agrochemicals throughout the environment has resulted in pollution and has negatively impacted the ecosystem and human health3,4,5. However, in contrast to the global awareness of the environmental footprint related to fertilizers6,7, the global repercussions of pesticide dispersion in the environment remain largely unknown due to the lack of a comprehensive geographic quantification of active ingredient (AI) use and residues. Studies addressing pesticide threats mostly remain site-specific, and only a minority have targeted regional and global extents8,9,10,11 to assess the risks associated with a specific pesticide class (for example, insecticides or organochlorine pesticides) or within a certain environmental compartment (such as surface water8,10 and the atmosphere12,13). Given the expected population growth, the use of agricultural pesticides will probably continue to increase in the future5; yet, in the age of globalization, a global outlook on environmental pollution by pesticides and its relation to ecosystem vulnerability is still missing.

To contribute to filling in this gap, we propose global mapping of the environmental risks posed by the 92 most used AIs (comprising 59 herbicides, 21 insecticides and 19 fungicides) at 5 arcmin resolution (about 10 km × 10 km at the Equator), which we next juxtaposed with water scarcity14, biodiversity15,16,17,18 and national income2. Our assessment targets the ecological risks in four environmental compartments (namely soil, surface water, groundwater and atmosphere), noting that we did not include pesticide impacts on human health and not all living organisms in an environmental compartment are considered. On the basis of these analyses, we ultimately identified susceptible regions that may require tailored strategies for the sustainable use of pesticides in agriculture.

## Pesticide risk in global agricultural land

To quantify pesticide risk in each geographic grid cell, we calculated the non-cumulative predicted environmental concentration (PEC) of each targeted AI in the four environmental compartments mentioned above using a spatially explicit model19 fed with georeferenced environmental datasets and AI physicochemical properties as inputs (Methods and Supplementary Tables 1 and 2). We sourced the geographic and crop-specific AI application rates from our recently developed PEST-CHEMGRIDSv120 global database gridded at 5 arcmin resolution (Methods). In each grid cell, the agricultural land consists of multiple crop types21 that receive applications of multiple AIs20. Hence, we adopted the hierarchical approach of the Pesticide Use Risk Evaluation decision-support system22, which sums the risk quotient of all AIs within an environmental compartment. The risk quotient was determined as the ratio between the PEC and the predicted no-effect concentration (PNEC) derived from each AI’s ecotoxicities (Methods and Supplementary Table 2). The ‘risk point’ of each environmental compartment was then evaluated as the log-transformed sum of all risk quotients. Finally, the overall ‘risk score’ in a grid cell (RS) was calculated as the maximum risk point across the four environmental compartments. On the basis of the species sensitivity distribution curve (Methods and Supplementary Fig. 1), we classified RS into negligible (RS ≤ 0), low (0 < RS ≤ 1), medium (1 < RS ≤ 3) and high (RS > 3) risk. This procedure allowed us to draw a global picture of environmental susceptibility to pesticide pollution.

Specifically, we found that 74.8% of the global agricultural land (approximately 28.8 million km2) is at some risk of pesticide pollution (RS > 0, Fig. 1); remarkably, 31.4% (approximately 12.1 million km2) falls within the high-risk class (RS > 3). Regional analysis showed that 61.7% (2.3 million km2) of the European agricultural land is at high risk of pesticide pollution. The three European countries with the largest land area at high risk are located in Eastern and Southern Europe, namely, Russia (0.91 million km2; Supplementary Table 4), Ukraine (0.35 million km2; Supplementary Table 4) and Spain (0.19 million km2; Supplementary Table 4), which are among the largest crop producers in Europe21. Among all regions, Asia has the largest land area at high risk (4.9 million km2), with 2.9 million km2 in China and 0.35 million km2 in Kazakhstan (Supplementary Table 4). The agricultural land in Oceania shows the lowest pesticide pollution risk.

Our RS map in Fig. 1 complements and expands earlier assessments such as the insecticide runoff potential analysis in Ippolito et al.8, which identified similar high-risk regions in Asia, America and South Europe. However, accounting for a wider range of pesticide AIs and environmental compartments in this work reveals additional geographic regions under high pollution risk, for example, areas across Eastern Europe and parts of Africa where the earlier assessment reports medium-to-very-low runoff potential8.

Pollution by pesticide mixtures is an emerging global issue because mixtures can elicit synergistic toxicity in non-target organisms under both acute and chronic exposures23,24. The risk map in Fig. 1 considers their additive effects, but excludes synergistic effects; hence, to better illustrate the global extent of pollution by pesticide mixtures, we counted the AIs that pose risks to the environment in each grid cell. An AI is considered to pose a risk when its PEC in any environmental compartment exceeds the PNEC. Globally, 63.7% of the agricultural land is at risk of pollution by more than one AI and 20.9% by more than ten AIs (Fig. 2). We found that 93.7%, 73.4% and 69.4% of the agricultural land in Europe, North America and South America, respectively, is contaminated by more than one AI. China is at risk of pollution by the greatest number of AIs, with 8.4% of the agricultural land (0.34 million km2; Supplementary Table 5) contaminated by more than 20 AIs.

## Pesticide risk in vulnerable regions

Pesticides can be transported to surface waters and groundwater through runoff and infiltration, causing pollution to water bodies and thereby reducing the usability of water resources. By mapping the pesticide risk and AI count over the water-risk database in AQUEDUCT-v2.1 (ref. 14), we found that, globally, 0.62 million km2 of agricultural land in regions suffering from highly variable and scarce water supply is facing high pollution risk by pesticide mixtures, among which 20.1% is located in low- and lower-middle-income countries (Extended Data Fig. 1a). Nation-wise, China has the most extensive land area subject to water scarcity and high pesticide pollution risk (0.27 million km2, about 3% of China’s total land surface; Extended Data Fig. 1a and Supplementary Fig. 2a), with surface water seeming to be the most susceptible environmental compartment (Extended Data Fig. 2). In contrast, groundwater is relatively protected from pesticide pollution (Extended Data Fig. 2) due to low aquifer net recharge.

To assess whether pesticide use constitutes a threat to biodiversity, we analysed the pesticide risk and AI count maps against geographically gridded species richness for tetrapods, which include mammals16, birds15, amphibians17 and reptiles18. We found that 34.1% of the global high pesticide pollution risk areas (approximately 4.18 million km2) are located in regions bearing high biodiversity (that is, ≥323 tetrapod species, the 75th percentile of global value), with 1.25 million km2 being in low- and lower-middle-income countries (Extended Data Fig. 1b). As the decline in amphibians has been closely linked to pesticide contamination25, we expanded our analysis to highlight the exposure of vulnerable amphibian species to pesticide pollution risk. We found that 0.37 million km2 of areas at risk of pesticide mixture pollution (that is, RS > 0 and AI count > 1) intersect the habitat of at least one of either endangered or critically endangered amphibian species (Extended Data Fig. 1c), with major hotspots located in China, Australia, Guatemala and Chile. Along with many studies underlining the toxicity of pesticides to wildlife26, the biodiversity loss earlier associated with the export of agricultural products that led to deforestation and habitat loss27 received in our analysis an additional element of attention: pesticide dispersion in intensive agriculture is an additional stressor that can exacerbate the loss of biodiversity.

## Regions of concern

To provide a synthesis of our work, we integrated the indicators for pesticide pollution risk, water scarcity and biodiversity into a map that locates regions of concern where tailored strategies for the sustainable use of pesticides may be needed (Fig. 3). In this map, level 1 identifies regions of high pollution risk, high water scarcity and high biodiversity. We identified the top five watersheds perceiving a level 1 concern as Orange in South Africa, Huang He in China, Indus in India, Murray in Australia, and Parana in Argentina. Surprisingly, four out of the five countries with level 1 concern are within high- and upper-middle-income economies. Although the level 1 concern regions cover less than 30,000 km2 of the land surface, we found that 5.20 million km2 are classed as level 2 and spread mainly across Asia and South America, with 1.72 million km2 located in low- and lower-middle-income countries.

The results of our study report a widespread global pesticide pollution risk with vast risk areas located in vulnerable regions that bear high biodiversity and suffer from low availability of freshwater. Our results expand and complement earlier regional-scale studies that report the detection of pesticide residues in freshwater bodies in South Africa28 and the Yellow River (Huang He) in China29. Besides impacting ecosystem health, the leaching of pesticides into water bodies used as sources of drinking water can pose risks to human health. Our analysis reinforces the need for a more detailed global assessment of pesticide contamination levels in major rivers, estuaries and lakes and to account for pollutant levels when assessing water scarcity and quality30.

In a warmer climate with a growing population, the use of pesticides is anticipated to increase to combat the possible rise in pest invasions and feed the planet31; thus, the threats estimated in our study may escalate further. Although protecting food production is essential for human development, reducing pesticide pollution is equivalently crucial to protect the biodiversity that maintains soil health and functions, contributing towards food security32. The increasing public awareness of the adverse impact of pesticides in recent years has called for the establishment of pesticide policies to reduce pesticide use. Within the context of policy-making, the spatially explicit RS estimated in this study can provide an indicator that quantifies pesticide risk in different agricultural settings (that is, not merely the quantity of AIs used), which is lacking in most of the current pesticide policy frameworks33. The RS defined here align with the pesticide load indicators used in Denmark34, although we did not account for pesticide impacts on human health. As our estimates extend globally across 168 nations, the proposed RS, AI counts and assessment of regions of concern can be incorporated into the Environmental Performance Index framework, which provides global metrics to rank the performance of countries on sustainability issues35.

Although this study was solely focused on environmental health, the effect of pesticides on human health is also an important aspect that requires comprehensive assessment. This assessment at a global scale would, however, be highly intricate as it would involve the quantification of human exposure to pesticides resulting from agricultural production and possible intake via diverse pathways including air, water and food, where the latter intake pathway involves food distribution and international food trading. Hence, pesticide use can affect not only the health of local communities but also the consumers in other importing countries. We urgently recommend that a global strategy is established to transition towards sustainable agriculture and sustainable living with low pesticide inputs and reduced food loss and food waste to achieve responsible production and consumption in an acceptable, profitable system.

## Methods

All modelling and analyses were conducted using Mathworks MATLAB (version R2017a).

### Application rates of active ingredients

To determine pesticide pollution risks, we first predicted the pesticide concentrations in all environmental compartments, which implied knowledge of pesticide application rates. For this, we used our previous work (PEST-CHEMGRIDS20) to obtain the global georeferenced crop-specific AI annual application rates in year 2015, which were estimated on the basis of the data provided by the USGS Pesticide National Synthesis Project36 and constrained against the country-specific pesticide use data reported by FAOSTAT2. PEST-CHEMGRIDS provided the high and low estimates of the top 20 AIs used on 175 crops, classified into six dominant crops (alfalfa, corn, cotton, rice, soybean and wheat) and four aggregated crop classes (vegetables and fruit, orchards and grapes, pasture and hay and other crops), totalling 95 different AIs that represent about 84% of the pesticide mass used in 2015. Crops were aggregated on the basis of the classification in the USGS Pesticide National Synthesis Project36 and were described in detail in table 2 in ref. 20. In this study, we excluded three AIs (Bacillusamyloliquefaciens, calcium polysulfide and petroleum oil) from PEST-CHEMGRIDS due to insufficient input data relative to their physicochemical properties and ecotoxicities. Hence, in the assessment of pesticide pollution risks, we accounted for the application of 92 AIs in total (listed in Supplementary Table 2) on 10 crop classes at median annual rates.

### PECs

Because the AI application history at a specific location was not known, we calculated the non-cumulative PEC of each AI in groundwater, surface water, soil and the atmosphere using the spatially explicit approach of the Environmental Potential Risk Indicator for Pesticide version 2.1 (EPRIP 2.1)19 with the assumption that all AIs were applied once a year at the annual application rates of 2015 obtained from PEST-CHEMGRIDS. The estimated PECs refer to those observed following an application and were not calculated as accumulating over time.

The non-cumulative PEC in groundwater ($${\mathrm{PEC}}_{i,j}^{{\mathrm{GW}}}$$) of active ingredient i on crop j was calculated as a function of application rate Ri.j, soil properties (porosity, bulk density, field capacity and organic carbon content), groundwater characteristics (water table depth, groundwater thickness and net recharge rate) and AI physicochemical properties (degradation rate, volatility and adsorption capacity). In surface water, $${\mathrm{PEC}}_{i,j}^{{\mathrm{SW}}}$$ was calculated using the empirical approach in the SYNOPS37 and DRIPS38 models to account for Ri.j, topography (slope angle), rainfall depth and the AI fraction available for transport via runoff determined by AI degradation rate and its adsorption to soil organic carbon. The PEC in soil, $${\mathrm{PEC}}_{i,j}^{{\mathrm{SL}}}$$, in the top 2 cm, was calculated as a function of Ri.j and soil bulk density, and used to determine the AI PEC in the atmosphere $${\mathrm{PEC}}_{i,j}^{{\mathrm{AT}}}$$. Using the approach taken in the VOLASOIL39 model, we calculated $${\mathrm{PEC}}_{i,j}^{{\mathrm{AT}}}$$ as a function of $${\mathrm{PEC}}_{i,j}^{{\mathrm{SL}}}$$, soil properties (including porosity, bulk density, field capacity and organic carbon content), AI physicochemical properties (including water solubility, volatility and adsorption) and atmospheric temperature.

### PNECs

We defined the predicted PNEC of the 92 selected AIs in each of the four environmental compartments using an assessment factor approach40 with acute toxicity data sourced from the Pesticide Properties DataBase41 (Supplementary Table 2). The PNECs in surface water and soil were determined using the median lethal concentration (LC50) of fishes and earthworms, respectively, with an assessment factor of 1,000; that is, $${\mathrm{PNEC}}_i^{{\mathrm{SW}}} = {\mathrm{LC}}50_i^{{\mathrm{fishes}}}/1,000$$ and $${\mathrm{PNEC}}_i^{{\mathrm{SL}}} = {\mathrm{LC}}50_i^{{\mathrm{earthworms}}}/1,000$$. For the atmosphere, we defined PNEC as the inhalation LC50 of rats with an assessment factor of 1,000. Following the European Commission guidelines42, we defined the PNEC for groundwater as 0.1 µg l−1 for all AIs with no assessment factor applied.

### Pesticide pollution risks

For each environmental compartment k, we calculated the crop-specific risk quotient (RQ) of each AI as the ratio between PEC and PNEC (that is, $${\mathrm{RQ}}_{i,j}^k = {\mathrm{PEC}}_{i,j}^k/{\mathrm{PNEC}}_i^k$$). Because a specific AI can be used across multiple crop classes within a grid cell, we calculated the overall RQ of each AI by weight averaging the crop-specific RQs with the crop harvested areas A (that is, $${\mathrm{RQ}}_i^k = \mathop {\sum}\nolimits_j {({\mathrm{RQ}}_{i,j}^k \times A_j)} /\mathop {\sum}\nolimits_j {A_j}$$). By adopting the hierarchical approach of the Pesticide Use Risk Evaluation decision-support system22, we determined the risk point (RPk) in an environmental compartment k as the log-transformed sum of all RQs in that compartment (that is, $${\mathrm{RP}}^k = \log \mathop {\sum}\nolimits_i {{\mathrm{RQ}}_i^k}$$).

The overall RS in a grid cell was then calculated as the maximum of the RPs across the four environmental compartments (that is, RS = max{RPk}). We classified RS into four risk classes: negligible (RS ≤ 0), low (0 < RS ≤ 1), medium (1 < RS ≤ 3) and high (RS > 3) on the basis of the average species sensitivity distribution curve for pesticides (Supplementary Fig. 1) determined using the parameters reported in ref. 43. Specifically, RS ≤ 0 corresponds to less than 5% probability for any of the species to experience an effect, whereas RS > 3 signifies that the probability for a random species to be affected by the pesticides is equal to 90%.

### Model input data

The model input variables were determined from spatially explicit global datasets (Supplementary Table 1). We sourced the soil bulk density, porosity and organic carbon content from the SoilGrids44, which consists of globally gridded soil profiles to 2 m depth. We estimated the soil water content at field capacity using the soil porosity, the globally gridded soil field capacity obtained from the IGBP-DIS dataset45, air entry suction and pore-volume distribution index λ obtained from ref. 46, following the model from Brooks and Corey47 (that is, soil water content = [field capacity/air entry suction]λ × porosity). The soil properties used in this work were the averages along the top 2 m soil depth.

We acquired the equilibrium groundwater table depth from ref. 48 and we estimated the groundwater thickness by subtracting the groundwater table depth from the soil thickness (distance to bedrock), which was sourced from the Distributed Active Archive Centre for Biogeochemical Dynamics of the Oak Ridge National Laboratory49. The net groundwater recharge was estimated as the balance between annual rainfall and evapotranspiration. We sourced globally gridded daily rainfall data from the CPC Global Unified Precipitation data provided by the NOAA/OAR/ESRL PSD50 and the monthly actual evapotranspiration from ref. 51; the atmospheric temperature was sourced from the Global Historical Climatology Network-Daily dataset52. We obtained the globally gridded terrain slope maps from the Harmonized World Soil Database v1.2 (ref. 53).

The AI physicochemical and ecotoxicological properties were obtained from the Pesticide Properties DataBase41 database and the literature54,55,56,57,58,59 (see Supplementary Table 2 for details).

### Output maps and data analyses

We ultimately produced three output maps60 gridded at 5 arcmin resolution (approximately 10 km at the Equator): the first is the RS map showing the exposure of agricultural land to pesticide pollution (Fig. 1); the second is the AI count map quantifying the number of AIs posing pollution risk to agricultural land and showing the exposure of the environment to pesticide mixtures (Fig. 2); and the third is the regions of concern map identifying locations susceptible to pesticide pollution upon meeting the selected criteria described below (Fig. 3). To produce these maps, we selected 1,199,195 grid cells with agricultural land using the harvested area maps of the 10 crop classes distributed along with PEST-CHEMGRIDS20, which were originally produced by ref. 21 and ref. 61. Among the selected grid cells, 2,408 cells (~0.2%) were neglected due to insufficient input data for computing the RS values and we thus modelled 38.54 million km2 of agricultural land in total. For the AI count map, we considered an AI to pose a pollution risk if any of its $${\mathrm{RQ}}_i^k$$ values were greater than 1, whereas the regions of concern were identified against water scarcity and biodiversity indicators.

We used the physical quantity risk indicator reported in AQUEDUCT-v2.1 (ref. 14) to locate areas suffering from high water risk. The physical quantity risks measure the risks related with the availability and variability of water supply; higher values indicate higher water risks. A grid cell is considered to be at high water risk when its physical quantity risk exceeded 4. To identify areas bearing high biodiversity, we used the geographically gridded species richness maps for tetrapods, which include mammals16, birds15, reptiles18 and amphibians17. We considered a grid cell to have high biodiversity when the total number of species in that grid cell was greater than the 75th percentile of global values (that is, 323 species). We classified countries into different income groups according to the definitions in FAOSTAT2 (Supplementary Table 3).

Finally, we integrated the pesticide pollution risk, water scarcity and biodiversity indicators to identify regions of concern. We assigned ‘no concern’ to all grid cells with RS ≤ 0 and ‘concern level (4 – N)’ to grid cells with RS > 0 that satisfied N criteria, which are (1) high pesticide pollution risk (that is, RS > 3); (2) high water risk (the physical quantity risk > 4); and (3) high biodiversity (total number of species >75th percentile of global values).

### Uncertainty and data quality

We quantified the reliability of our estimates by performing a global sensitivity analysis for 11 selected input variables that include AI application rates, soil properties (bulk density, porosity, water content and organic carbon content), groundwater characteristics (water table depth, groundwater thickness and net recharge rate), slope angle and hydroclimatic variables (rainfall and atmospheric temperature). We assumed that all variables could span ±50% of the reference values obtained from global datasets. For AI application rates, we tested ranges that span +50% of the high estimates and −50% of the low estimates provided in PEST-CHEMGRIDS. We sampled randomly across the variable space using a uniform distribution, and conducted a total of 50,000 model realizations per grid cell (5.98 × 1010 realizations in total).

Within the tested variable space, we determined the certainty index (CI)60 of a grid cell as the probability of that grid cell falling into the risk class estimated in the RS map in Fig. 1. Hence, CI = 0 indicates low certainty and CI = 1 indicates high certainty. We found that the estimated risks (Fig. 1) in approximately 22% of grid cells were highly certain (that is, CI = 1; Supplementary Fig. 3a) and fewer than 9% of grid cells had low certainty (CI < 0.6).

For grid cells with CI < 1, we determined the variable that had the highest contribution to the uncertainty by using AMAE and AMAV indices62, which measure the relative contribution of variables to the mean and variance of the model output, respectively. Among all tested variables, AI application rates had the greatest control over uncertainties in more than 42% of grid cells (Supplementary Fig. 4). Hence, to compute the quality of our estimates (QI)60, we combined CI with the data quality of PEST-CHEMGRIDS (QIAPR), that is, QI = (CI + QIAPR)/2. PEST-CHEMGRIDS provides AI- and crop-specific quality indices, and hence we compute the overall QIAPR as the average quality weighted by the application rates. In this work, our estimates had mid-to-high quality in 93% of grid cells (that is, QI ≥ 0.6; Supplementary Fig. 3b).

### Assumptions and limitations

The pesticide pollution risk presented in this study may be overestimated because: (1) it assumes a single application at an annual rate, (2) it assumes that all fields are adjacent to surface water bodies, (3) it assumes maximum exposure of non-target organisms in time and in space, and (4) it assumes no loss due to drift and interception by crops. In this study, pesticides were assumed to reach the soil as a result of direct deposition, rainfall washing of crop leaves and crop debris fall, regardless of the application methods. We presume that common practices such as spraying may lead to pesticide drift and potentially diluting the concentration and delaying the time taken for pesticides to eventually reach the soil after spraying. We also identified limitations that can lead to underestimating the risks. First, our assessment did not consider legacy pollution from AIs that were banned before 2015. For example, atrazine was not included in the calculation of RS in European Union countries that banned its use before 2015. However, many field studies have reported a high detection frequency of atrazine and its degradation products in European soils despite its ban about a decade ago63. Second, we did not account for the pollution risks of pesticide degradation products, which may still be toxic and be more persistent than the parent molecules64. Third, the calculated PECs were non-cumulative and not dynamic in time—that is, we did not consider the effect of accumulation of pesticides and their degradation products over time, and thus may not fully capture the pervasiveness of certain AIs. Fourth, we did not account for the synergistic effects of pesticide mixtures65 as there is very limited data on the ecotoxicity of pesticide mixtures and only a small number of organisms have been tested for PNECs.