Introduction

The global population surge has led to a large increase in food demand, resulting in the large and widespread use of pesticides and fertilizers1. Agricultural diffuse pollutants such as nitrogen, phosphorus, pesticides, and antibiotics, lead to eutrophication, and damage to aquatic ecosystems2. Owing to climate warming, two-thirds of the world’s land is exposed to wetter and more variable precipitation patterns3, leading to increased agricultural diffuse pollution losses and challenging watershed ecological security4. A study showed that interannual variability in nutrient discharge loads within watersheds was up to 2.3 times higher, with 76% attributed to annual precipitation variability5. In extreme precipitation years, annual nitrogen loads are 30% higher than the long-term median6. However, the response of pesticide discharge levels to precipitation variability within watersheds remains unclear7, despite two-thirds of the global agricultural land being at risk of pesticide contamination.

The pesticide transport in watersheds are associated with rainfall events, pesticide application, soil type, and slope8,9. Among these factors, soil type and slope have been relatively stable parameters over a long period, and their changes have a negligible impact on the annual variance in pesticide loss. By contrast, precipitation as a meteorological factor is a major factor driving pesticide transport in watersheds10. Most studies on pesticide diffuse pollution acknowledge that pesticide discharge is driven by precipitation and runoff events; however, due to the random occurrence of precipitation events, trends are often discussed by simply comparing precipitation and pesticide losses in parallel, ignoring the multi-year trends in precipitation and the effects on pesticide discharge loads11. All calculated loads are estimates because the amount of data necessary to calculate the actual load is always insufficient. Current estimation methods include mechanistic models, empirical models, and field observations. Although long-term trends in pesticide emissions can be reliably simulated using mechanistic models, these models require local data homogeneity and complicated parameters, and the accuracy of the simulations in large watersheds (>500,000 km2) is questionable1. The limited scale of field observations remains a challenge for monitoring high spatiotemporal frequencies over large regions12. Current research is deficient in pesticide discharge in large watersheds. Empirical models are commonly used for estimating the agricultural diffuse pollutant losses in large watersheds such as the modified export coefficient model. The modified export coefficient model reasonably estimates the spatiotemporal variances of diffuse pollutants discharge loads by employing physical processes, such as precipitation, runoff, slope, canopy interception, and anthropogenic export parameters13. Although the modified export coefficient model has been widely used in estimating diffuse pollutants, it is still not involved in estimating pesticides loads in large watersheds.

Pesticide application is a dominant factor associated with pesticide discharge loads in watersheds14. Pesticide runoff loss typically accounts for 5.9% (0.56–14%) of the applied amount11. The primary factor in assessing watershed vulnerability is pesticide application intensity15; however, temporal trends between pesticide usage and discharge loads remain unclear. Pesticide application is mainly regulated by the demand even if it is always excessive, and the trends fluctuate; thus, the temporal impact on pesticide discharge loads has largely been overlooked. Because of the increasing concern about diffuse pesticide pollution, pesticide reduction policy under policy intervention are mandatory to reduce the amount of pesticide use. A zero-growth pesticide usage policy implemented by government regulators in China in 2015 led to a 152,000 tons reduction in usage and a 1.6% increase in efficiency in 1 year16. From 2016 to 2019, a further 390,000 tons reduction was observed. Correspondingly, field monitoring campaigns in China have demonstrated decreased concentrations of pesticides in some rivers17,18. Although the pesticide reduction policy mitigates diffuse pollution, ignoring the long-term precipitation effects may overestimate the actual efforts to discharge loads, which should be addressed by conducting long-term exploration to fill in the gaps.

Under the dual impact of increased precipitation variability and pesticide reduction policy, identifying the driving factors of pesticide discharge loads in the watershed is essential. A representative herbicide, atrazine, characterized by high usage, high mobility, high persistence, high detection, and endocrine disruption, was used as the target pesticide19,20. Atrazine is a widely used herbicide in corn, broomcorn, sugarcane and other crops. The annual usage of atrazine in USA and China has been ranked number two among conventional pesticides21. Atrazine is moderately dissolved in water and is highly persistent in the environment, with half-life up to 2 years22. Therefore, atrazine is ubiquitously in aquatic environments with 100% detected frequency, and poses potential risks to aquatic life14, and the diffuse pollution from chemical pesticides will be gradually eliminated. Although unusual increases in precipitation affect human efforts to optimize environmental pollution issues, we are confident in the on-going environmental optimization policies.

Notably, this study has several limitations. First of all, the spatial heterogeneity of atrazine application data affects the accuracy of the estimation37. The mismatch in spatial resolution between the atrazine application database and the watershed’s inflow coefficient increased the uncertainty of the estimation results. We acknowledge that the estimated atrazine discharge loads values remained significant uncertainty; however, this uncertainty did not affect the main points in this study, as the spatial resolution’s uncertainty did not propagate to the temporal trend variations and field observation verified the overall spatial pattern of atrazine. The development of a regular monitoring system is required to provide the basic data on the migration behavior of pollutants in the watershed. In addition, the pesticide reduction policy was launched for 8 years and the available data could only trance for 5 year, which needs continue to follow up the policy impact. Although this study is a regional simulation, the changes in precipitation caused by climate change are a global issue that affects us all3. Reducing the use of conventional chemical pesticides is also an inevitable path for sustainable agriculture14,38. Exploring the process of environmental changes under the dual influence of climate change and human policies is beneficial for adjusting current policies and contemplating future development directions.

The current status determines that pesticide reduction policy requires ongoing effort. Our comparison of the atrazine concentrations calculated from the annual atrazine loads and runoff discharge in the watershed demonstrated that the average atrazine concentration from 2014 to 2017 exceeded the surface water atrazine concentration limit of 3 μg/L (Supplementary Fig. 4). Excess atrazine concentration can damage the ecological environment, especially for algae, fish, and frogs, and has potential neuroendocrine effects on mammals20. Because of the spatial variability in the entire watershed and the high exposure to atrazine in agricultural areas29, we speculated that atrazine levels would exceed the surface water concentration limits in intensively cultivated areas during the first few runoff events after its application39. Such high levels of atrazine after runoff mainly occur in corn cultivation areas, such as the US Corn Belt35, the Great Barrier Reef region of Australia25. These regions in intensively atrazine applied areas should pay more attention to the ecological risks to aquatic atrazine under the climate change effect, and require continuously tactical efforts to reduce the atrazine usage and seek alternatives to green pesticides in the future. Current estimations of the atrazine discharge load are useful for understanding the dynamic patterns of pesticide diffuse pollution in the dual effects of climate change and policy regulations and for providing knowledge for future policy development. The scale gap in refined modeling of diffuse pollutants in watersheds needs to be bridged in the future to better support government decision-making.

Methods

Study area

The Yellow River is the second-longest river in China and extends 5,464 km from its source to the estuary, covering 795,000 km2 within the watershed (Fig. 1A). The Yellow River is one of the largest vulnerable watersheds worldwide, and it surrounds the Loess Plateau, which has the most severe soil erosion rate worldwide40. Tremendous amounts of diffuse pollutants are discharged with soil erosion in watersheds, threatening aquatic environmental quality and ecological health41. Agriculture is one of the most important economic parts in the Yellow River Watershed. The main crop types include corn and wheat. Currently, the corn acreage in the watershed is 15.6 million hectares, and corn production is 98.9 million tons24. The main location of intensive agricultural areas is on both sides of the mainstream. The Hetao Plain is the largest gravity-fed irrigation area in Asia, with the mainstream of the Yellow River serving as an irrigation channel. The Fenwei Plain is a traditional irrigation area, the fourth largest plain in China, and the largest alluvial plain in the middle reaches of the Yellow River. The map of Yellow River Watershed and its sub-basins were delineated using ArcGIS 10.2 software based on a 90 m resolution digital elevation model (DEM). All of the visualization results about the Yellow River Watershed in this study were based on this map.

Export coefficient model

The overall frame work of the research process and the methodology is provided (Fig. 6). The atrazine discharge load in the watershed was calculated for the sub-basins using a modified export coefficient model involving the comprehensive water inflow coefficient. Based on the conventional export coefficient method, the modified model incorporates water inflow coefficients and calculates the atrazine loads in each sub-basin13. The modified equation is as follows:

$$L=\theta \lambda \sum {E}_{i}{A}_{i}$$
(1)

where L represents the diffuse pollution load of atrazine (kg/a), θ represents the ratio between atrazine and nitrogen loads, λ represents the comprehensive water inflow coefficient, E represents the application intensity of atrazine for the i-th sub-basin (kg/(km2·a)), and parameter Ai represents the tillage area in the i-th sub-basin (km2). Detailed calculations of λ are provided in the Supplementary Material (Supplementary Methods).

Fig. 6
figure 6

Overall framework of the research process and the methods.

Full size image

The λ coefficient is determined by precipitation, slope, surface/underground runoff, and vegetation interception factors in the watershed. Among these factors, the terrain index is related to slope and is stable over long time in a large watershed. The retention index is related to the fractional vegetation cover, which were stable in the last two decades in over 95% of the Yellow River Watershed42. However, the precipitation, transport, and leaching index were driven by annual precipitation in the decadal scale (2.09 ± 0.53%)43. We also compared the decadal variance of precipitation in the study area from 1990 to 2019 (Supplementary Fig. 5). The average precipitation from 2010 to 2019 significantly increased compared to the previous 20 years. Based on this background, the precipitation was determined as the independent variable and the λ coefficient was determined as the dependent variable in this study.

Field sampling and quantification of the atrazine concentration

Thirteen sites were sampled in September 2019 in the mean stream and tributaries (Fig. 1A). Sampling site locations are listed in Supplementary Table 1. Sampling sites in main stream were determined with the entry and exit of Hetao Plain (BYGL and SHHK), Fenwei Plain (LM and TG) and other agricultural areas along the upstream, middle-stream and down-stream, which could represent the pesticides discharge level in these areas. In addition, four agricultural tributaries were sampled in sites joining the main stream: the Fen, Wei, Tao, and Huangshui Rivers, which could represent pesticide discharge levels in these tributaries. A 2 L sample of river water from each site was poured into a dark-glass bottle and then filtered with a glass fiber membrane (0.77 μm, Whatman, USA). The samples were stored at 4 °C prior to the next procedure.

Extraction and quantification of atrazine were performed as described by Ouyang et al.12. In sequence, a 6 mL mixture of dichloromethane-methanol (v/v, 80/20, acidized by 5 mM HCl), 2 mL methanol, and 5 mL ultrapure water were permeated via graphitized carbon black solid-phase extraction (SPE) columns to activate the SPE columns. Desethylatrazine-13C (30 ng) was used as the surrogate mixture in filtered water. Next, the water samples were pre-concentrated using SPE cartridges at a flow of 1.0 mL/min in vacuum. Then, 1.5 mL methanol and 6 mL dichloromethane methanol (v/v, 80/20, acidized with 5 mM HCl) were used to elute the extract. When the eluents were concentrated by nitrogen flow to 200 μL, the residual volume was fixed at 1 mL using methanol. Next, the residual was filtered using a 0.22 μm polytetrafluoroethylene membrane for equipment detection. Atrazine was quantified using a UPLC-MS/MS machine loaded with a liquid chromatograph and a mass spectrometer (Supplementary Table 2). Atrazine-13C (30 ng) was used as the internal standard. A 0.1% formic acid water solution was used as eluent A, and a 0.1% formic acid acetonitrile solution was used as eluent B. Extracted ion chromatograms of atrazine, desethylatrazine-13C, and atrazine-13C are shown in Supplementary Fig. 6.

Quality control

Quality control methods, namely duplicate field samples, blank field samples, laboratory blanks, and surrogates, were used to ensure measurement accuracy. The calibration equations, method detection limits, and method quantification limits for atrazine and its surrogates are presented in Supplementary Table 3. A loss control method was used for the extraction process44. The average recovery ± standard deviation of surrogates desethylatrazine-13C in surface water was 87 ± 8.3% (N = 13). The average relative percentage difference of atrazine was 7 ± 13.6% (N = 13).

Statistical analysis

R version 4.0.3 was used for data analysis and visualization45. A Fully Bayesian Monte Carlo sensitivity analysis scheme was used to identify the most sensitive factor for atrazine discharge loads in the watershed based on the Bayesian Gaussian process regression model in R package tgp46. Random Latin hypercube samples were drawn at each Markov chain Monte Carlo iteration to estimate the main effects and the total sensitivity indices. Atrazine application intensity and precipitation were set as input variables and the atrazine discharge loads was set as output responses. In the setting of variables in R program, the size of random Latin hypercube samples was set as 1000, the span was set as 0.3, the 3-vector of Monte-Carlo parameters (B)urn in, (T)otal, and (E)very were set as 4000, 8000, and 10, respectively.

The ANOVA assumptions were tested before comparing the means among sites. The spline model is a reliable tool for describing nonlinear relationships between complex variables47. The discharge load of atrazine in the unusual precipitation years was estimated using a spline model.