Pesticide use negatively affects bumble bees across European landscapes

Abstract

Sustainable agriculture requires balancing crop yields with the effects of pesticides on non-target organisms, such as bees and other crop pollinators. Field studies demonstrated that agricultural use of neonicotinoid insecticides can negatively affect wild bee species^1,2, leading to restrictions on these compounds³. However, besides neonicotinoids, field-based evidence of the effects of landscape pesticide exposure on wild bees is lacking. Bees encounter many pesticides in agricultural landscapes^4-9 and the effects of this landscape exposure on colony growth and development of any bee species remains unknown. Here we show that the many pesticides found in bumble bee-collected pollen are associated with reduced colony performance during crop bloom, especially in simplified landscapes with intensive agricultural practices. Our results from 316 Bombus terrestris colonies at 106 agricultural sites across eight European countries confirm that the regulatory system fails to sufficiently prevent pesticide-related impacts on non-target organisms, even for a eusocial pollinator species in which colony size may buffer against such impacts^10,11. These findings support the need for postapproval monitoring of both pesticide exposure and effects to confirm that the regulatory process is sufficiently protective in limiting the collateral environmental damage of agricultural pesticide use.

Fig. 1. Effects of landscape exposure to pesticides on bumble bee colony weight and production.

a, We deployed bumble bee (B. terrestris L.) colonies (n = 316) adjacent to apple (APP, green points) and oilseed rape (OSR, yellow points) across eight European countries. b, Colony production (total number of produced bees estimated by the sum of closed and eclosed cocoons) declined with pesticide risk (log-transformed and centred toxicity-weighted pesticide concentrations in pollen stores; Methods). c,d, Colony weight gain (response ratio ln(g_max/g_initial) and percentage change (exp(lnRR)) also declined with pesticide risk (note double and shared y axes). Focal crop (c) and landscape context (b,d) modified these effects, with stronger declines at apple (c; green line) compared to at oilseed rape sites (c; yellow line) and in landscapes with more cropland (b,d; solid line +1 s.d. proportion of cropland). Point colours (b,d) correspond to country colours (a) and are scaled by their MCR, the factor by which the mixture of compounds in a sample is riskier than the single most risky compound (Methods). Points in c are scaled by the number of pesticide compounds quantified in a sample. Fitted lines are estimated on the basis of generalized (b) and linear (c,d) mixed effects models. Shaded areas represent the regression 95% CI. Results from statistical models are given in Table 1.

Extended Data Fig. 1. A simplified view of landscape exposure and resulting pesticide risk to bees.

Pesticide use creates potential hazard for non-target organisms. For bees in agricultural landscapes, pesticide risk results when their activity exposes them to this hazard (top left panel). Without the co-occurrence of hazard and exposure we expect no risk (remaining panels). Of course, the degree of hazard and exposure will depend on pesticide properties (e.g., toxicity, environmental fate, product formulations, use patterns) and bee traits (e.g., foraging range, sociality, body size, detoxification pathways). Moreover, real-world exposure occurs at landscape scales (see insets), because bees can integrate multiple sources of exposure by visiting spatially separated patches that vary in the identity, amount, timing and toxicity of hazard. We use the colony pollen stores collected by bumble bees (Bombus terrestris) to quantify pesticide risk resulting from this landscape exposure. We quantify exposure as the concentrations (µg/kg) of 267 substances in the pollen while hazard is quantified by the substances’ toxicities (LD_50s). Scaling concentrations by toxicities and summing these toxicity-weighted concentrations provides a relative measure of pesticide risk to bees.

Extended Data Fig. 2. Percent of focal crop pollen and the number of pesticide compounds in pollen stores.

Bumble bee colony stores contained a substantial but variable portion of focal crop pollen types (a). More pesticide compounds were found in pollen collected at apple (n = 50) sites than at oilseed rape (n = 56) sites (b, F_1,104 = 39.59, P < 0.001). For proportion of focal crop pollen (a), large points are means based on raw data and error bars are standard deviations. For the number of unique compounds (b), large points are estimated means from linear models and error bars are 95% confidence intervals. Small points are the individual data points.

Extended Data Fig. 3. Bumble bee colony performance metrics are correlated.

Colony weight gain (response ratio: ln(g_max/g_initial)) is positively related with (a) total colony production (total number of produced bees estimated by the sum of closed and eclosed cocoons; χ² = 354.27, P < 0.001) and (b) queen production (sum of closed and eclosed queen cocoons; χ² = 37.42, P < 0.001). Fitted lines are estimated based on generalized linear mixed effects models with a negative binomial error distribution. Shaded areas represent the regression 95% confidence intervals. Point colours correspond to country colours in Fig. 1a and Extended Data Fig. 5a.

Extended Data Fig. 4. Effects of field exposure to pesticides on bumble bee colony queen production.

Colony queen production (sum of closed and eclosed queen cocoons) declined with pesticide risk (centred toxicity-weighted pesticide concentrations in pollen stores, see Methods) and landscape proportion cropland modified this effect, with stronger declines in landscapes with more cropland (solid line +1 SD proportion of cropland). Points are scaled by pesticide maximum cumulative ratio (MCR), the factor by which the mixture of compounds is riskier than the single most risky compound (see Methods). Fitted lines are estimated based on generalized linear mixed effects models with a negative binomial error distribution. Shaded areas and error bars represent the regression 95% confidence intervals. Results from statistical models are given in Extended Data Table 1. Point colours correspond to country colours in Fig. 1a and Extended Data Fig. 5a.

Extended Data Fig. 5. Overview of study design and set-up.

Our research network (a) included 128 sites in two focal crops (apple: green points, oilseed rape: yellow points) in eight European countries. At each site, three bumble bee (Bombus terrestris) colonies were deployed prior to focal crop bloom and weighed three times: before, during and after focal crop bloom. (b) The interval between first and second weights (circles) and second and third weights (diamonds) varied depending on region- and crop-specific bloom periods. (c) Colony weights at the time of deployment. Crop averages for number of days (a) and initial weight (b) across colonies are given as dashed lines. Site coordinates (a) are randomly jittered to protect farmer confidentiality. Colours in b and c correspond to country colours in a and Fig. 1a.

Extended Data Fig. 6. Pesticide risk effects exceed a suggested Specific Protection Goal (SPG).

To evaluate the magnitude of pesticide risk for the bees (a), we assume that the colonies belonging to a low-risk group (blue points, 25th percentile of risk) can be used to calculate the average maximum weight as a baseline (blue line). Using a suggested SPG for bumble bees of 10% reduction in colony weight (yellow line), 60% of the remaining colonies in our study exceed this. In (b) we compare the risk for colonies that exceed the SPG (yellow points; n = 143 colonies) to those that do not (grey points; n = 94). The SPG is meant to protect 90% of the colony population across Europe, therefore in (c-e) we compare colony performance endpoints between the baseline colony group (blue, 25^th percentile of risk; n = 79) and a high-risk colony group (red, 90th percentile of risk; n = 30) based on (a). Points and error bars (b–e) depict estimated means and 95% confidence intervals from linear (b,c) and generalized linear mixed effects models (d,e). * P < 0.05, ** P < 0.01 and *** P < 0.001. Exact p-values from statistical models are given in Extended Data Table 2.

Extended Data Fig. 7. Sensitivity analysis of mixed effect models for the effect of pesticide risk, focal crop and proportion cropland.

We specified the same model structure (equation above plots) and simulated different levels of replication through stratified subsampling without replacement of colonies at the focal crop-country level and repeated the analyses for weight gain (a) and production (b) with these rarefied data sets. Levels of replication spanned 5 colonies per focal crop and country (N = 80 colonies) to 12 (N = 192), with this later value being the lowest focal crop-country level of replication in the data. The large point depicts the level of replication (N = 316 colonies) and p-values reported in the main text (See Table 1). Points and error bars (a,b) are means and 95% confidence intervals calculated over 1000 iterations per replication level.

Extended Data Fig. 8

Bombus terrestris colonies in the field (a) and under dissection in the laboratory (b).

Extended Data Fig. 9. Risk is not correlated with (a) proportion cropland or (b) initial colony weight.

Fitted lines and 95% confidence intervals (a,b) are estimated based on linear mixed effects models. Point colours correspond to country colours in Fig. 1a, S5a and Extended Data Fig. 5a.