Neal M Williams

Pesticide use in agricultural landscapes creates environmental contamination that is heterogenous in space and time. Mobile organisms, such as bees, are exposed to multiple contamination sources when visiting patches that vary in the amount, timing and toxicity of pesticides used. Yet, environmental risk assessments (ERA) typically fail to consider this heterogeneity, in part because of the complexities of estimating exposure to different pesticides, and subsequent risk at organism‐relevant scales. We use pesticide assays of 269 bee‐collected pollen samples to understand the spatiotemporal variability of risk across a network of 41 field sites in southern Sweden. Observed bee pesticide risk is calculated based on compound‐specific residue quantifications in pollen and standardized toxicity data. We then compare the ability of three classes of landscape‐scale variables to predict this risk: (1) landscape composition and configuration metrics, (2) landscape load based on national pesticide use data and (3) predictions from a newly developed bee pesticide exposure model. Based on use data, 10 crops account for 81% of the total risk. We detected 49 pesticide compounds in bee‐collected pollen. Although herbicides and fungicides constitute the bulk of detected pesticides, both in frequency and amount quantified, unsurprisingly, insecticides contribute the most to risk. Landscape composition and configuration metrics did not predict observed pesticide risk, and interactions with bee species indicate taxa‐dependency in predictions. Landscape load predicted observed risk consistently between taxa. Risk estimates from our exposure model were strongly predictive but only when considering realized risk (i.e., risk estimates based on prior pesticide use information). Synthesis and applications . Predicting pesticide risk based on landscape patterns could enable landscape‐scale ERA. However, simple metrics of landscape pattern, such as proportion of agricultural land, are not sufficient. We found that risk observed in bee‐collected pollen was best predicted when integrating spatialized pesticide use in the pesticide exposure model, underscoring the importance of such data for research, monitoring and mitigation. Further, we propose a guidance framework for future landscape ecotoxicological risk analyses that clarifies data needs relative to risk prediction goals.