Modeling spray drift and runoff-related inputs of pesticides to receiving water

Environmental Pollution 234 (2018) 48e58

Contents lists available at ScienceDirect

Environmental Pollution journal homepage: www.elsevier.com/locate/envpol

Modeling spray drift and runoff-related inputs of pesticides to receiving water* Xuyang Zhang*, Yuzhou Luo, Kean S. Goh California Department of Pesticide Regulation, Sacramento, CA, 95814, United States

a r t i c l e i n f o

a b s t r a c t

Article history: Received 14 March 2017 Received in revised form 6 October 2017 Accepted 8 November 2017

Pesticides move to surface water via various pathways including surface runoff, spray drift and subsurface flow. Little is known about the relative contributions of surface runoff and spray drift in agricultural watersheds. This study develops a modeling framework to address the contribution of spray drift to the total loadings of pesticides in receiving water bodies. The modeling framework consists of a GIS module for identifying drift potential, the AgDRIFT model for simulating spray drift, and the Soil and Water Assessment Tool (SWAT) for simulating various hydrological and landscape processes including surface runoff and transport of pesticides. The modeling framework was applied on the Orestimba Creek Watershed, California. Monitoring data collected from daily samples were used for model evaluation. Pesticide mass deposition on the Orestimba Creek ranged from 0.08 to 6.09% of applied mass. Monitoring data suggests that surface runoff was the major pathway for pesticide entering water bodies, accounting for 76% of the annual loading; the rest 24% from spray drift. The results from the modeling framework showed 81 and 19%, respectively, for runoff and spray drift. Spray drift contributed over half of the mass loading during summer months. The slightly lower spray drift contribution as predicted by the modeling framework was mainly due to SWAT's under-prediction of pesticide mass loading during summer and over-prediction of the loading during winter. Although model simulations were associated with various sources of uncertainties, the overall performance of the modeling framework was satisfactory as evaluated by multiple statistics: for simulation of daily flow, the Nash-Sutcliffe Efficiency Coefficient (NSE) ranged from 0.61 to 0.74 and the percent bias (PBIAS) < 28%; for daily pesticide loading, NSE ¼ 0.18 and PBIAS ¼ 1.6%. This modeling framework will be useful for assessing the relative exposure from pesticides related to spray drift and runoff in receiving waters and the design of management practices for mitigating pesticide exposure within a watershed. Published by Elsevier Ltd.

Keywords: Pesticide Spray drift Runoff Water quality Modeling

1. Introduction Pesticides have been widely detected worldwide in the surface water and have been determined as one of the major pollutants attributed to degradation of the aquatic ecosystems (Stehle and Schulz, 2015). In agricultural areas, the major pathways for pesticides to move from treated fields to the surface waters include surface and subsurface runoff, spray drift, dust and vapor transport. Surface runoff has been recognized as the most prevalent pathway (Reichenberger et al., 2007; Schulz and Matthies, 2007). However, many studies have suggested that spray drift could also be

*

This paper has been recommended for acceptance by Charles Wong. * Corresponding author. E-mail address: [email protected] (X. Zhang).

https://doi.org/10.1016/j.envpol.2017.11.032 0269-7491/Published by Elsevier Ltd.

significant (Cryer et al., 2001; Raupach et al., 2001b). Wauchope et al. (2004) estimated that 40e55% of the amount of pesticides applied could move offsite via spray drift. Such values were confirmed by other studies with values ranging from 20 to 50% of the applied dose (Maybank et al., 1978; Ravier et al., 2005). Understanding the pathways of pesticide transport is essential to the mitigation of the negative impacts of pesticides on aquatic ecosystems (Reichenberger et al., 2007). Resources should be allocated to the right type of management practices targeting different pathways of pesticide transport. For example, in areas where spray drift is significant, windbreaks and buffers could be installed to prevent spray drift for pesticide applications when wind is blowing toward surface waters. In contrast, if surface runoff is the major pathway, a different set of management practices aiming to reduce tailwater runoff such as sediment pond, recycling tailwater, and vegetated ditches could be implemented.

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

Studies dedicated to the understanding of pesticide transport pathways and their relative significance are very limited. Schulz (2001) compared spray drift and runoff-related inputs of azinphose-methyl (AZP) (organic carbon normalized soil adsorption coefficient (Koc) ¼ 1112) and endosulfan (END) (Koc ¼ 11500) from fruit orchards into the Lourens River, South Africa. Water samples were collected at sites located at tributary and downstream receiving water. They compared concentration and mass loadings of AZP and END from runoff and spray drift. Both concentration and loadings were higher in runoff samples than in spray drift samples taken at the receiving water site. In terms of loadings, values following runoff were higher than following spray drift by factors of 41e860 for AZP. For END, the factors were between 14 and 2100. Assuming an average year with 12 spray drift, 1.7 large runoff events, and 3.4 small runoff events, they calculated that the load contribution from spray drift is 0.7 and 0.3% for AZP and END, respectively, with the largest contribution from surface runoff. Another study compared the relative contribution of waterborne (mainly runoff) and airborne (spray drift, vapor, and dust) pathways for END in cotton growing regions in Nanoi River catchment, Australia (Raupach et al., 2001a, 2001b). Their results showed that END concentration from runoff events ranged from 0.6 to 0.23 ppb, spray drift from 0.02 to 0.09 ppb and vapor from 0.01 to 0.05 ppb. Assuming an average year with 26 spray drift, and 11 runoff events, the load contribution from spray drift for END was calculated as 9.9e59%. They concluded that runoff was the dominant pathway when it occurs; however, runoff did not occur frequently, and most of the time, the END detections were due to airborne transport (spray drift and vapor). They also found that dust transport was not significant. Besides the differences in concentration and loadings, runoff and spray drift events also showed different toxicity effects on aquatic organisms. Dabrowski et al. (2005) conducted microcosm experiments to compare the toxicity of cypermethrin to mayfly from spray drift and runoff events. They found that spray-drift related inputs result in increased dissolved phase pesticide levels, which is readily bioavailable. On contrast, runoff-related inputs result in both dissolved and adsorbed phase as well as increased velocity and turbidity. The adsorption to suspended sediment may reduce the bioavailability of the chemical, while increased velocity may enhance toxicity. They concluded that for cypermethrin, a hydrophobic pesticide, spray drift inputs were more toxic compared to those related to runoff. The relative toxicity of spraydrift vs. runoff inputs depends on the chemical and pesticides with high solubility may show a higher significance of runoff regarding its contribution to aquatic toxicity. In the above-mentioned studies, the concentration of pesticides resulting from runoff is much higher than those from spray drift or vapor transport. For END (a highly sorbed pesticide), The factor ranged from 2 to 42 (Schulz, 2001) and 2.6e12 (Raupach et al., 2001a). In terms of mass loading, the factor may even be higher (14e2100, Schulz, 2001). The relative contribution of the spray drift and runoff varies significantly depending on the study area and chemical. Models have been increasingly used to simulate fate and transport of pesticides within a watershed. Most of the watershed models focus on the pathways of runoff (surface and subsurface) and leaching. Spray drift is largely ignored or over-simplified. Mottes et al. (2014) reviewed 16 watershed and field models for pesticide transfer. They found that for most of the models, drifted pesticides may be represented as leaving the system (pure loss of mass) or simply not considered. Only two of the 16 models simulate spray drift as a function of distance from the field. Mottes et al. (2014) highlighted the importance of spray drift as one of the

49

major processes and they recommend that future development of catchment models should consider the integration of landscape drift models. A couple of studies have been devoted to the modeling of the relative contribution of runoff and spray drift within watersheds. Dabrowski and Balderacchi (2013) developed an indicator to assess the relative mobility and risk of pesticides in the Lourens River catchment, South Africa. Their indicators were able to provide relative exposure and risk rating among pesticides but did not represent absolute exposure. In addition, their method does not consider routing of pesticides within watershed or movement of pesticides attached to sediment. Cryer et al. (2001) developed a modeling system incorporating the AgDRIFT®, the Pesticide Root Zone Model version 3 (PRZM3) and the Hydrological Simulation Program-FORTRAN (HSPF) models for simulating spray drift and runoff of chlorpyrifos in the Orestimba Creek Watershed of California, USA. The paper highlighted the importance of identifying the relative contribution of different transport processes for pesticide movement such as spray drift and runoff. Daily runoff from agricultural fields was simulated using the PRZM3 and then downscaled to a sub-daily timescale for input to the HSPF model, which simulated instream hydrology. The results on flow and chlorpyrifos loadings were not calibrated. The model over-estimated chlorpyrifos loading and spray drift contribution. According to the paper, “simulated peaks not seen experimentally are probably due to the resolution of the spray drift event since spray drift is the largest predicted contributor to Orestimba Creek loadings”. . The hour at which a pesticide application was unknown and the wind direct/speed data was at daily resolution. Additional uncertainties were introduced by using the PRZM3 for simulating pesticide runoff as generated by over-irrigation. The irrigation routines in PRZM3 do not allow runoff resulting from overirrigation. As a result, a modified RPZM3 algorithm was used for generating tail-water runoff. Yet this algorithm was not validated. The above mentioned studies represented useful initial steps towards modeling the pesticide inputs via spray drift and runoff in a watershed. The objective of this study is to extend the efforts of simulating watershed behavior by developing a modeling system for evaluating the significance of pesticide spray drift and runoff within watersheds. 2. Methods 2.1. The modeling framework The modeling framework presented here is an integrated system containing models that simulate fate and transport of pesticides within agricultural watersheds. The core components of the modeling framework include a GIS procedure for identifying agricultural fields with drift potential, the AgDRIFT® model (Teske et al., 2002), and the Soil and Water Assessment Tool (SWAT) model (Arnold et al., 1998) (Fig. 1). The three components are interconnected in the following manner: (1) the GIS procedure identified target fields and spray events with potential to drift pesticides to the receiving water, using the location of the applied field, their distance to the receiving water, and the wind directions during the day of application; (2) the outputs of the GIS procedure serve as the starting point for the AgDRIFT modeling. For each drift event, AgDRIFT model predicts the fraction of mass that moves offsite via spray drift according to pesticide application method and downwind distance from the applied field; (3) finally, the amount of pesticides that deposit on the receiving water during each drift events were added to the corresponding SWAT modeling units as point source inputs. This fraction of pesticides together with pesticide runoff from treated fields will be routed throughout the

50

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

Fig. 1. Diagram of the modeling framework.

watershed to the outlet while undergoing various transformation processes. The models were carefully chosen considering their abilities to simulate important processes associated with pesticide transport (e.g. spray drift, surface and subsurface runoff, leaching, soil erosion, and channel routing), transformation (e.g. degradation in soil, water and air, volatilization, adsorption), and agricultural practices (e.g. irrigation, application method, mitigation measures) as well as their validities. The AgDRIFT model was developed under the joint efforts by U.S. EPA, the Spray Drift Task Force (a coalition of agricultural chemical companies) and the US Department of Agriculture's Forest Service. The model was designed to address the assessment of offsite drift of pesticides from agricultural applications (Teske et al., 2002). It contains aerial, ground, and orchard/ airblast modules. The aerial model was derived from the AGricultural DISPersal (AGDISP) model while the ground and orchard airblast models were based on empirical models derived from experimental datasets. The AgDRIFT model has been used in the U.S. and worldwide for evaluating exposure from spray drift. It has been a useful tool for evaluating effectiveness of buffer zones. SWAT is a watershed hydrological model which has been widely applied to simulate spatial and temporal trends of stream flow, sediment yields, nutrients and more recently (Gassman et al., 2007; Luo et al., 2008). The model has also been proven useful for simulating management practices for reducing pesticide concentrations within a watershed (Zhang and Zhang, 2011; Gali et al., 2016). A review on various pesticide models by Mottes et al. (2014) showed that SWAT has strong capabilities in simulating important pesticide processes within a watershed. Regarding spray drift, SWAT considered the drift amount as a loss from the modeling system, which is an oversimplification of reality. Holvoet et al. (2008) had modified the SWAT model to simulate pesticide direct losses (spray drift and point loss) using the Ganzelmeier method (Ganzelmeier and Rautmann, 2000). However, this algorithm has not yet been incorporated in the current version. This limitation of the SWAT model is addressed in this study by integrating the AgDRIFT model

and the GIS procedure with the SWAT model. Studies have indicated that the Ganzelmeier curves for orchard airblast application was similar with the one in the AgDRIFT model (USEPA, 1999; Walker, 2001). They indicated that these two curves were similar for orchard airblast applications. Most of the inputs required by the core models are publically available in California. These inputs include the Pesticide Use Reporting Database (PUR) (CDPR, 2017), California Irrigation Management Information System (CIMIS) database, drainage network from the National Hydrological Dataset (NHD), landuse from the National Landuse Land Cover dataset (USGS, 2001) and the soil properties from the SSURGO database (USDA, 2017). The final outputs of the model framework are the daily concentration and loadings of pesticides in water and sediment, daily discharge and sediment concentration at the outlet of the watershed as well as the outlets of each sub-basin. The pesticide loadings as predicted by the SWAT model will be compared with the observation taken from monitoring studies. By this comparison, models can be calibrated and validated. At the end, pesticide load contributions from spray drift and runoff events can be calculated during the temporal cycle of interest (e.g. per year). The following sections demonstrate the applications of the modeling framework for deriving the relative inputs of pesticides from spray drift and runoff to the receiving water using a case study. 2.2. Study area The Orestimba Creek Watershed (OCW) is a tributary to the San Joaquin River located in the Central Valley of California (Fig. 2). The watershed boundary was adapted from previous modeling work in the watershed based on extensive investigations on the surface drainage network (Zhang and Zhang, 2011). Watershed delineation was further refined based on recent studies on the watershed (Karpuzcu et al., 2013). The watershed covers an area of about 513 square kilometer. The upper part of the watershed is mainly mountains of the coastal range and the lower part of the watershed

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

51

Fig. 2. The Orestimba Creek Watershed. upper left: The location of the Orestimba Creek Watershed (red spot) within the State of California, USA and the discretization of the watershed showing the subbasins and hydrological response units (HRUs) for SWAT modeling; lower left: Major streams and monitoring sites within the Orestmiba Creek Watershed; right: the lower Orestimba Creek Watershed with the identified agricultural fields with drift potential. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

is dominated by agricultural land. The creek is ephemeral, with high flows normally occurring in late winter, and irrigation drainage accounting for low flows during the summer months. The water delivery and drainage system in the lower watershed is highly engineered representing a great challenge for hydrological modeling. The Lower Orestimba Creek flows across a 146 km2 stretch of irrigated agricultural land, which is planted with various crops including alfalfa, walnuts, almonds, irrigated pasture, dry beans, tomatoes and corn. There are three flow gage stations, one at the upper portion of the creek maintained by the US Geological Survey (USGS 11274500), one at the middle portion maintain by California Department of Water Resources (DWRB08735) and the other one at the confluence of the Orestimba Creek and the San Joaquin River also maintained by USGS (USGS 11274538) (Fig. 2). In the 1990s, chlorpyrifos was commonly used in the watershed on orchards and alfalfa fields for treating various insects, resulting in frequent detections of chlorpyrifos in the watershed. In 1996, Dow Agro Science Inc. conducted a monitoring study on chlorpyrifos concentrations in the OCW (Poletika and Robb, 1998). They took daily water samples for one year at locations of the three gage stations noted as L1, L2 and L3 on the map, with L1 being the most downstream site (Fig. 2). The study also installed an onsite weather station to gage the rain amount during the study period (May 1, 1996 to April 30, 1997). In addition, the CIMIS weather stations of Patterson, Kesterson and Modesto provide additional weather data.

Li ¼ Fi  Ci  0:0864

(1)

Where Li is pesticide loading on day i in kilogram, Fi is daily discharge on day i in cubic meter per second (converted to daily total using the unit converting factor), Ci is concentration of pesticides on day Li, and 0.0864 is the unit converting factor. To determine the portion of loading pertaining to spray drift events, daily chemographs were carefully examined to identify the duration of the events, which often lasted for 3e4 days post application (Fig. 3). Pulses of loading begin usually on the day of application or the day after. The chemographs rise and fall until a new event occurs as triggered by a new pesticide input (runoff or another spray drift event). The total loading attributed to a particular spray drift event was then calculated as:

Ldr ¼

n X

Li

(2)

i¼m

Where Ldr is the total loading triggered by a spray drift event; Li is pesticide loading on day i; ?? is the day when concentration rises as following a spray drift event and n is the day when the concentration falls. During the study period of May 1, 1996eApril 30, 1997, the fraction of loading related to spray drift was then determined as the sum of Ldr in Equation (2) divided by the sum of Li in Equation (1). 2.4. Determination of pesticide deposition via spray drift

2.3. Calculation of chlorpyrifos loading from monitoring data Daily pesticide loading in the receiving water was calculated using the following equation:

To facilitate the modeling processes, a database was created containing various inputs on pesticide use location, weather, landuse, soil, digital elevation, stream network, and monitoring results (Fig. 1). To identify fields and spray events with spray drift

52

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

Fig. 3. Measured daily chlorpyrifos concentration from the three monitoring sites along the Orestimba Creek. Measurements started on May 2, 1996 and ended on April 30, 1997. L1 is the most downstream and L3 is the most upstream site.

potential, first, a GIS layer was created containing all identified fields within the watershed that receive pesticide applications during the study period. Second, a buffer of 400 meters around the receiving water was created assuming that fields within 400 meters could potentially drift to the Orestimba Creek. According to the AgDrift model, few pesticide droplets (<0.5%) move beyond 400 m (1312 ft) via spray drift (Fig. 5). Finally, for all the fields identified in step 1 and 2, pesticide applications during the study period were identified as well as the prevailing wind rose on the day of application. If wind blew from the applied fields towards the Orestimba Creek, the applications were then identified as spray drift event. For spray events that were identified with spray drift potential, the AgDRIFT model was used to produce spray drift deposition curves, one for each event. For the orchard airblast applications in walnut orchards, the USEPA Tire I orchard airblast scenario was chosen with the deposition curve for the “orchard” option. This option assumed 20 rows of planting with application efficiency of 98.63%. For the aerial application occurred in alfalfa fields, Tier II aerial spray with a boom length of 65%, boom height at 10 ft, swath width of 60 ft and wind speed at 10 mph. The wind speed was set at a conservative value since the hour at which the applications were made were unknown (the daily average was around 4.5 mph). According to the distance between the applied field and the Orestimba Creek, the mass of chlorpyrifos drifting to the creek was

calculated as:

Md ¼ Rp  A  Fd

(3)

Where Md is the mass of pesticides drifted and re-deposited to the receiving water in kg; Rp is the pesticide application rate in kg/ha, which is obtained from the PUR database for each drifting application. Fd is the fraction of applied rate per unit area derived from the AgDRIFT deposition curve at downwind distance d, and d is the shortest distance between the edge of the applied field and the creek (Fig. 4). A is the surface area of the creek that captures drifted droplets in ha, which is determined using the prevailing wind direction and field location as illustrated in Fig. 4. For the purpose of spray drift calculation, a circle is used to represent the field. The circle has the same area and centroid as the actual applied field. Two tangent lines with the direction of the prevailing wind are drawn for the circle. The intersected areas of the creek by the two lines are considered as the receiving area A (Fig. 4). 2.5. SWAT modeling For watershed simulation, this study uses the ArcSWAT modeling package, which runs the 2012 version of the SWAT model

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

53

1996 and April 1997 were used as the dataset for observation. 3. Results and discussion 3.1. Identification of spray drift potential

Fig. 4. Determination of the deposited area for drifting events.

within the ESRI ArcGIS 10.3 environment. The interface facilitates the preprocessing of the GIS data of elevation, soil, land use, and weather as basic model inputs. A SWAT project database was created with model parameters extracted from the various input datasets: the user-defined sub-basin boundary, stream network (NHD), landuse (NLULC and the PUR crop information), soil (SSURGO) and user input weather data (CIMIS and onsite rain gage). Land use was firstly classified based on the NLULC dataset and further refined for agricultural fields with specific crop information taken from the PUR database. The watershed was delineated into sub-basins and further into hydrological response units (HRUs), which are unique combinations of land use, soil type and slope. HRUs were defined such that the sizes of the HRUs were close to those of the actual fields. The major crops in the watershed included alfalfa, lima beans, orchard, pasture, tomato, rangeland, and strawberry. Management inputs on planting, fertilization, harvest, and irrigation were set up for each HRU according to the common practices of the crop associated with the HRU. Pesticide use settings were taken from the PUR database with the specific date and application rate. Since the sizes of the HRUs were different from the size of actual field, the pesticide use per unit land (application rate) was adjusted such that the total use amount stayed the same as the associated fields. The spray drift fraction of chlorpyrifos was set up as point source inputs. Model simulation period was 10 years from 1990 to 1999, with 1990e1995 as warm up, 1996e1997 as calibration and 1998e1999 as validation period. Calibration of the SWAT model was carried out automatically using the SWAT calibration and uncertainty analysis program (SWAT-CUP) (Abbaspour et al., 2015) and manually. For automatic calibration, the Sequential Uncertainty Fitting (SUFI2) algorithm was used (Abbaspour et al., 2015). Multiple objective functions were used as targets for optimization including the Nash-Sutcliffe efficiency coefficient (NSE), the coefficient of determination (R2), percent of bias (PBIAS) and ratio of the root mean square error to the standard deviation of measured data (RSR). Daily measurements on flow and pesticide concentration taken between May

During May 1996 and April 1997, a total of 3592 kg of chlorpyrifos active ingredient was applied within the OCW mainly on walnut and alfalfa fields. A total of 14 pesticide applications were made within the 400-meter buffer area, among which 13 were identified with spray drift potential given the proximity between applied field and the creek as well as the wind directions on the day of application (Table 1). One of the applications was made on the field (MQ4) located downstream of the L1 sampling site. Since the spray drift impact on this field could not be captured by the measurements made on the three sampling sites (L1, L2, L3) (Fig. 2), this application was excluded from the model input to ensure a fair comparison between predicted and measured values on chlorpyrifos loads. Among the 13 spray drift events, 11 applications occurred on walnut orchards and two on alfalfa fields (Table 1). The applications on walnut orchards were made using orchard airblast. And the applications made on alfalfa field were by aerial spray (Table 1). The spray drift events occurred on 11 different fields, among which 9 fields were located next to the creek (Table 1). For those fields, the distance of zero was assigned for d. The AgDRIFT model was used to generate deposition curves for orchard airblast and aerial spray (Fig. 5). The amount of chlorpyrifos drifted and deposited on the creek ranged from 0.08 to 5.67% of the total amount that was applied on the field (or 0.01e28 kg of chlorpyrifos) for the orchard airblast applications. For the aerial spray application, the spray drift amounts were 0.11 and 6.09% of the applied mass (or 0.01 and 0.83 kg of chlorpyrifos) (Table 1). The spray drift fractions might be over-estimations due to the following three reasons: (1) It was unknown at what time of the day pesticides were applied. As a result, spray events were identified as having spray drift potential as long as wind rose of the day showed that wind had blown from the fields to the creek regardless of the hour; (2) Spray drift could be reduced by vegetated buffers. Since it was unknown whether there were buffers exists between applied field and the creek, we assumed none; (3) The shortest distance between the edge of the field and the creek was used for deriving deposition fractions and applied for the whole water surface. Theoretically, the distance would vary from edge to edge of the water body. Since the creek is narrow (width is mostly between 1 and 5 m with the maximum at around 24 m) and the difference from edge to edge is expected to be relatively small, this method is appropriate as a conservative approach. On the other hand, the spray drift fractions could also be underestimations due to the uncertainties in the spatial referencing of the applied fields in the PUR database. The PUR database contains records of all reported pesticide applications made to production agriculture including their amount, application date and method of application. The spatial resolution is reported at the section level (1  1 mile area) using the Public Land Survey System (PLSS) method. The exact field location was not recorded in the PUR database, even though field location identification, grower permit ID and crop can be used to identify individual fields without spatial reference. This study used this field identification information to spatially reference GIS locations of each individual field. A small portion of fields in the PUR database were not spatially identified as no matching fields were found in the GIS layer. In addition, the PUR database and its error correction algorithms were designed to capture records with high application rates. Errors with low application rates were less likely to be captured. Besides, some applications might not be reported. Therefore, it was possible that a

54

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

Table 1 Identified drift events associated with the application of chlorpyrifos to crops in Orestimba Creek Watershed, California. Field ID LII14

Crop

walnut

NJ13 LII1 LII10 LZ1 MH1 MQ2 LY3 LY3 MR4 LX35 MQ4

alfalfa corn

Spray date

04/30/96 05/19/96 04/30/96 09/05/96 07/26/96 07/26/96 07/03/96 05/29/96 05/29/96 08/03/96 04/22/97 09/19/96 03/11/97 06/30/96

Method

airblast

aerial airblast

Treated area (ha)

28.3 28.3 20.2 14.2 7.3 36.4 16.2 10.9 14.2 12.1 12.1 24.3 16.2 8.1

Wind direction

SSW SSW SSW SSW SSE SSW SSW, WSW ENE SSE, ESE SSW NNW SSW NNW SSW

Applied rate Rp (kg/ha) 2.24 2.24 2.24 1.12 2.24 2.24 1.68 2.24 2.24 2.24 2.24 0.56 0.56 1.12

Distance d

Drift fraction Fd

(m) 0 0 0 0 110 0 0 0 85 0 0 0 0

Deposition area A (ha)

Reach widtha (m)

Drift Mass (kg)

Drift/ total applied (%)

0.222 1.43 1e14 0.71 1.12 0.222 1.43 1e14 0.71 1.12 0.222 3.61 5e24 1.80 3.97 0.222 3.61 5e24 0.90 5.67 0.006 0.94 5e10 0.01 0.08 0.222 2.75 1e5 1.37 1.68 0.222 0.99 1e5 0.37 1.36 0.222 0.65 0.32 1.32 0.009 1.27 1e8 0.03 0.08 0.222 0.78 1e8 0.39 1.42 0.222 0.17 1e8 0.09 0.32 0.524 2.82 1e8 0.83 6.09 0.524 0.04 1e6 0.01 0.11 This application was excluded from model inputs as the field was located downstream of the sampling sites; and the impacts of spray drift from this field were not captured by the measured dataset.

a Reach segments are irregular. Widths of the stream segments depend on the water level, which was unknown. Channel width was estimated using the Google earth image. Many channel segments were covered by dense vegetation and the widths were estimated based on exposed channel width. Therefore, the reach width and deposition areas are likely over-estimates.

Fig. 5. Drift deposition curve of chlorpyrifos applications generated by the AgDRIFT model; the Y axis is the fraction of applied rate per unit area of deposition.

small portion of the applications might be missing or underreported. 3.2. Analysis of monitoring data Daily concentrations of chlorpyrifos were measured along the Orestimba Creek at the three sites located at the upper stream (L3), middle stream (L2) and the downstream site (L1) (Fig. 3). Among the 364 samples, 67e70% of the samples were detections (>0.001 mg/L), with the maximum concentration of 2.282 mg/L from a sample taken at the L3 site (Table 2). Peaks of concentration were observed mainly during spring and summer months. Concentrations rose when runoff or spray drift events occurred, and the peaks

often lasted for 3e4 days. At the downstream sites (L1 and L2), some peaks were associated with inputs from upstream (e.g. July 27, 1996 peak) while the others were triggered by inputs from between the monitoring sites (e.g. August 22, 1996 at L2, May 29, 1996 at L1). A total of 25 independent peaks were observed at the downstream site, among which 8 correspond to spray events within the 400 meter buffer of the creek. Direct linkage was established between the 8 concentration peaks and 11 pesticide spray events out of the total of 13 applications (some occurred on the same day). Two spray drift events occurred but did not result in concentration peaks. This might be due to the following reasons: (1) both pesticide applications occurred on fields (LX35, NJ13) located about 4 km upstream of the sampling sites (Fig. 2) and the

Table 2 Statistics for evaluating model performance. Nash-Sutcliffe efficiency (NSE) Coefficient of determination RMSE-observations standard deviation ratio (RSR) Percent bias (PBIAS) (R2) L1 flow calibration (1996e1997) L1 flow validation (1998e1999) L3 flow calibration (1996e1997) L3 flow validation (1998e1999) Chlorpyrifos loading at L1 (1996e1997)

0.65

0.71

0.60

11%

0.74 0.61

0.77 0.5

0.51 0.70

27.5% 19.7%

0.65 0.18

0.54 0.31

0.67 0.90

23.3% 1.6%

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

spray drift input was diluted as it travelled downstream; (2) the application rate was relatively low (0.63 and 1.26 kg/ha). These 8 peaks often occurred within one or two days post pesticide application. The remaining 17 peaks were identified as runoff events either because the timing matched rainfall events or there were no recent pesticide applications. Pesticide loading per day were calculated as the product of daily concentration (mg/L) and observed daily discharge (cubic meter per second) at the outlet of the watershed (L1). During the one-year monitoring period, a total of 1.3 kg chlorpyrifos was loaded in the Orestimba Creek, accounting for 0.04% of the total applied amount. Spray drift accounted for24% of the mass loading and runoff contributed to the remaining 76%. 3.3. SWAT output The Orestimba Creek watershed was delineated into 9 subbasins representing the drainage network. The sub-basins were then further divided into 66 unique HRUs for SWAT modeling. The outlet of the watershed was located at L1, which was also the outlet of sub-basin 9. Flows at L3 and L1 were calibrated for 1996e1997 and validated with the period of 1998e1999. Pesticide loadings were calibrated using the daily measurement data between May 1996 and April 1997. SWAT simulation on the stream flow was satisfactory for both calibration and validation period. As shown by the statistics on Table 2, the Nash-Sutcliffe Efficiency Coefficient (NSE) values ranged from 0.61 to 0.74. The RMSE-observations standard deviation ratio (RSR) ranged from 0.51 to 0.70. And the Percent bias (PBIAS) was less than 28%. According to Moriasi et al. (2007), for SWAT simulation at monthly time step, the performance would be rated as satisfactory or above if RSR is less than 0.7, NSE higher than 0.5 and PBIAS within ±25%. A guide for daily statistics was not provided in the paper, but in general, model simulations are poorer for shorter time steps than for longer time steps and thus a looser performance rating is warrant (Moriasi et al., 2007). This study simulates stream flow at the daily time step and within a highly managed watershed. The simulation of flow is very good given the

55

challenges related to the engineered surface drainage, relatively low flow year round and the low continuous summer flow contributed by irrigation runoff. The model was able to capture the timing of flow peaks yet tended to underestimate the flow discharge especially during big storm runoff events (Fig. 6). Simulation of the summer flow remains a challenge due to lack of information on irrigation timing and amount by various crops. In SWAT2005 and earlier versions, irrigation from the sources outside of the watershed is problematic. The model would return the excess irrigation water back to the source and this amount of water was not taken into account in the daily soil water balance (Dechmi et al., 2012). SWAT2012 has fixed this issue and thus allows for more realistic simulation on runoff generated by excess irrigation. Irrigation can be scheduled both manually and automatically. For manual irrigation, users can schedule irrigation by date or heat units. For auto-irrigation, users can trigger irrigation by soil water deficient or by crop growth. For both irrigation settings, users can specify the surface runoff ratio, which is the fraction of water applied that leaves the field as surface runoff (Neitch et al., 2011). In this study, both manual and automatic irrigation were used. For alfalfa, irrigation was scheduled manually as information on the scheduling of irrigation during the study period of 1996 and 1997 was available. For other crops, auto-irrigation was triggered by soil water stress. Additional knowledge about the timing, amount and efficiency of irrigation practices for each HRU could further improve the simulation results. Simulation results on flow were very sensitive to the rainfall data. Due to the small amount of rainfall under the Mediterranean climate, measurement errors and location of the rain gauge could have a significant impact on the model performance. At the beginning of the study, the SWAT model was set up using the rainfall data from the closest CIMIS station, which was about 35 km away from the centroid of the watershed. The flow simulation was significantly improved later, with the addition of the rainfall data from the onsite rain gage, which measured rainfall for major storms during the study period. This suggests that model uncertainties could be reduced with accurate rainfall data. Simulation of chlorpyrifos was desirable given the low

Fig. 6. Simulated and observed flow at the watershed outlet (L1).

56

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

Fig. 7. Simulated and observed chlorpyrifos loading at the watershed outlet (L1); red arrows denote drift events. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

concentration of chlorpyrifos and flow values that resulted from fine temporal and spatial scales (Fig. 7). Compared to flow simulation, evaluation statistics for chlorpyrifos showed a lower R2 (0.31), NSE (0.18) and higher RSR (0.90) for chlorpyrifos simulation (Table 2). Yet the PBIAS stayed low at 1.6%. Individual value plots show that the data ranges of observed and simulated overlap except for a few extremely high values. The observed mean chlorpyrifos loading was 3532 mg per day and the total loading per year was 1.29 kg. Simulated values were close to those observed with 3616 mg per day and 1.32 kg per year. 3.4. Significance of spray drift and runoff Pesticide inputs related to spray drift and runoff were calculated based on daily mass loading as simulated by the modeling system. The highest loading peaks occurred during winter and were associated with runoff events. In contrast, the peaks during summer months were mostly related to spray drift events (Fig. 7). During the one-year monitoring period, the total loading associated with spray drift was 0.25 kg or 19% of the total annual loading. The remaining 81% of the annual loading was associated with runoff events. Yet over half (54%) of the loading during May and September were due to spray drift events. In terms of concentration, runoff events often resulted in very high concentrations and drift events often resulted in relatively lower peak concentrations. The model-predicted spray drift input of 19% was lower than the 24% as calculated from the monitoring data, but within the ballpark given the uncertainties in daily chlorpyrifos simulation and measurements. This under-prediction is mainly due to the model's under-predicted mass loading during summer months and overpredicted loading during winter months. What's more, most of the spray drift events occurred during summer months between May and September. Both results from the simulation and monitoring data showed that the 85% drift contribution claimed by Cryer et al. (2001) may be an overestimation. And their conclusion that “chlorpyrifos non-point source loadings into Orestimba Creek arise predominantly from drift (85%) as opposed to overland flow or

irrigation tailwater runoff (15%)” is unsubstantiated by the modeling. Our results showed that in the OCW, runoff was the dominant pathway for chlorpyrifos loading in the creek. Spray drift accounted for about 24% of the annual loading during 1996e1997. Pesticide inputs via spray drifts were more significant during summer months accounting for half of the chlorpyrifos mass loading. Surface runoff created high peaks of chlorpyrifos concentration and were more frequent compared to spray drift. This is similar to the findings in Schulz (2001), who also found runoff being the dominant pathway with higher concentrations and loadings for AZP and END in the Lourens River watershed, an orchard-dominated watershed in South Africa. Spray drift contribution in the Orestimba Creek was about 14% during the calibration period and the contribution was more significant during the summer months. This spray drift fraction is higher than those in the Schulz (2001) study for AZP and END. But it was within the range of findings by Raupach et al. (2001a, b) who concluded that spray drift contributed 9.9e59% of END loading in the Nanoi River catchment, Australia. In California, a modeling study on pesticide loadings to the Sacramento River, San Joaquin River, and Bay-Delta estimated that spray drift accounted for about 4.4% of mass loading (Hoogeweg et al., 2011). This value was based on 10-year simulation of 40 pesticides. Best management practices have been increasingly recommend and implemented in agricultural watersheds to reduce pesticide contamination. The results of this study suggest that BMPs such as filter strips, vegetated ditches and irrigation water management could be very helpful in reducing the chlorpyrifos loadings in the OCW, since runoff is the dominant pathway. Zhang and Zhang (2011) demonstrated that a 20 m filter strip implemented in the applied fields could reduce 56% of sediment load and 89% of chlorpyrifos loading to the Orestimba Creek. Ditches with vegetation cover (Manning's roughness coefficient ¼ 0.075) would resulted in a 20% loading of chlorpyrifos in the creek. Cover crops could be used to reduce storm runoff during winter months. Spray drift would be reduced significantly by setting application buffers between the application site and the creek. Although runoff was the

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

major pathway, when considering toxicity, the significance of spray drift impacts might be higher than indicated solely by mass loading in highly engineered systems such as Orestimba Creek. Studies have indicated that pesticide inputs related to spray drift are at dissolved phase and are readily bioavailable; while those related to runoff are in both dissolved and adsorbed phases with reduced bioavailability (Dabrowski et al., 2005). This phenomenon is probably also dependent on the chemical properties of the pesticide. The differences in bioavailability of inputs from spray drift and runoff routes would be less significant for pesticides that are highly water soluble as runoff inputs would arrive mainly in dissolved phase as well. More studies are needed to evaluate the toxicity of pesticide inputs from spray drift and runoff events. The results from various studies on spray drift contribution indicated that the relative significance of spray drift and runoff vary significantly depending on the study area and chemical. The relative contribution of spray drift and runoff is dependent on two main factors: (1) difference in concentration resulting from runoff and spray drift events, and (2) the number of runoff vs. spray drift events occurring within a temporal cycle (growing season or year). 3.5. Sources of uncertainties Daily simulation of flow and pesticide loading is associated with various sources of uncertainties. SWAT is capable of simulating runoff as triggered by rainfall and irrigation events. The uncertainties in daily flow simulation are mainly associated with accuracies of rainfall data, the availability of information on irrigation amount and scheduling, the runoff flow direction, and the knowledge on surface drainage network. SWAT allows detailed management settings on irrigation including the date, amount and rate of irrigation for each HRU. However, this information is often unavailable. As a result, the automatic irrigation algorithms were used for estimating irrigation timing and water use based on either crop growth or soil water deficit. These algorithms introduce additional uncertainties as they reply on accurate simulations of crop growth or soil water. Therefore, calibration of the algorithms using measured or estimated water use data is needed. Finally, the SWAT simulations of flow also rely on accurate configuration of the drainage network. In agricultural watersheds such as the OCW, drainage network is heavily modified by man-made structures and the tailwater was delivered by drainage ditches and canals as opposed to follow elevation gradients. Knowledge on the drainage network and the flow directions will improve delineation of the watershed and consequently reduce simulation uncertainties. For pesticide simulation, additional errors were introduced by the uncertainties in sediment simulation and errors in measuring chlorpyrifos concentrations at low flow conditions (besides the various sources of uncertainties in flow simulations. Post application, the first runoff events often carry the most amount of pesticides compared to the subsequent ones. The closer the runoff events are to the application, the more pesticides they could carry off-site. The irrigation events triggered by the auto-irrigation algorithms may deviate from the actual timing of irrigation on the field. In addition, sediment simulation at daily time step was not calibrated due to lack of measurement information. However, discrete measurements of suspended sediment concentration were available. These data served as guidance in model calibration such that the simulated monthly results were within the ballpark of the monthly values as estimated from the discrete samples. Chlorpyrifos adsorption and desorption processes could be improved had daily suspended sediment concentrations been available. Finally, the uncertainties in measuring daily chlorpyrifos concentration based upon sampling location within the water column might also be a source of error when comparing the simulated mass loading

57

with the observed. Though model performances were limited by the various sources of uncertainties as mentioned above, the model was able to capture the variations of flow and chlorpyrifos mass loading in the Orestimba Creek with satisfactory results. The percentage of mass related to spray drift was also within the ballpark of measurements. The results proved that the modeling framework can be used to assess the relative exposure from pesticides related to spray drift and runoff in receiving waters. It could also assist in the design and implementation of management practices for mitigating pesticide exposure (and also in the evaluation of the mitigation effectiveness of various management practices). The modeling framework could be further improved with expanded capabilities using probabilistic distributions of the various inputs having large uncertainties. 4. Conclusions The modeling framework simulated daily flow and pesticide loading with satisfactory results. During the study period of 1996e1997, 13 pesticide applications of chlorpyrifos were identified as spray drift events. Mass deposition on the Orestimba Creek ranged from 0.08 to 6.09% of applied mass. Analysis of the monitoring data suggest that surface runoff was the major pathway accounting for 76% of the annual loading in the Orestimba Creek, and spray drift contributed to 24% of the mass loading. The results from the modeling framework further confirmed the finding. SWAT simulation showed that the percent of mass loading associated with surface runoff and spray drift were 81 and 19%, respectively. Spray drift contributed over half of the mass loading during summer months. The lower spray drift contribution as predicted by the modeling framework was mainly due to SWAT's under-prediction of pesticide mass loading during summer and over-prediction of the loading during winter. Although model simulations were associated with various sources of uncertainties, the overall performance of the modeling framework was satisfactory as evaluated by multiple statistics. The modeling framework could assist in the design, implementation and evaluation of management practices for mitigating pesticide exposure within a watershed. In the next phase of the project, we will apply the modeling framework to a second watershed to further validate its applicability. A probabilistic approach will be used to take into account the variabilities in model inputs. Acknowledgement Authors thank Pam Wofford, the anonymous reviewers and the editors for their critical review of this paper. This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Appendix A. Supplementary data Supplementary data related to this article can be found at https://doi.org/10.1016/j.envpol.2017.11.032. References Abbaspour, K.C., Rouholahnejad, E., Vaghefi, S., R, S., Yang, H., Kløve, B., 2015. A continental-scale hydrology and water quality model for Europe: calibration and uncertainty of a high-resolution large-scale SWAT model. J. Hydrol. 524, 733e752. Arnold, J.G., Srinivasan, R., Muttiah, R.S., JR, W., 1998. Large area hydrologic modeling and assessment part I: model development. J. Am. Water Resour. Assoc. 34, 73e89. CDPR, 2017. Pesticide Use Reporting Database. California Department of Pesticide Regulation. Available online via. http://www.cdpr.ca.gov/docs/pur/purmain. htm.

58

X. Zhang et al. / Environmental Pollution 234 (2018) 48e58

Cryer, S.A., Fouch, M.A., Peacock, A.L., Havens, P.L., 2001. Characterizing agrochemical patterns and effective BMPs for surface waters using mechanistic modeling and GIS. Environ. Model. Assess. 6, 195e208. Dabrowski, J.M., Balderacchi, M., 2013. Development and field validation of an indicator to assess the relative mobility and risk of pesticides in the Lourens River catchment, South Africa. Chemosphere 93, 2433e2443. Dabrowski, J.M., Bollen, A., Schulz, R., 2005. Combined effects of discharge, turbidity, and pesticides on mayfly behavior: experimental evaluation of spraydrift and runoff scenarios. Environ. Toxicol. Chem. 24, 1395e1402. Dechmi, F., Burguete, J., Skhiri, A., 2012. SWAT application in intensive irrigation systems: model modification, calibration and validation. J. Hydrol. 470e471, 227e238. Gali, K.G., Cryer, S.A., Poletika, N.N., Dande, P.K., 2016. Modeling pesticide runoff from small watersheds through field-scale management practices: Minnesota watershed case study with chlorpyrifos. Air, Soil Water Res. 9, 113e122. https:// doi.org/10.4137/ASWR.S32777. Ganzelmeier, H., Rautmann, D., 2000. Drift, drift reducing sprayers, and sprayer testing. Asp. Appl. Biol. 57, 1e10. Gassman, P.W., Reyes, M.R., Green, C.H., Arnold, J.G., 2007. The soil and water assessment tool: historical development, applications, and future research directions. Trans. ASABE 50, 1211e1250. https://doi.org/10.13031/2013.23637. Holvoet, K., van Griensven, A., Gevaert, V., Seuntjens, P., Vanrolleghem, P.A., 2008. Modifications to the SWAT code for modelling direct pesticide losses. Environ. Model. Softw. 23, 72e81. Hoogeweg, C., Williams, W., Breuer, R., Denton, D., Rook, B., Watry, C., 2011. Spatial and Temporal Quantification of Pesticide Loadings to the Sacramento River, San Joaquin River, and Bay-Delta to Guide Risk Assessment for Sensitive Species. CALFED Science Grant 1055. Karpuzcu, E., Weissmann, G., Stringfellow, W., Gulati, S., Herr, J., 2013. Orestimba Creek Agricultural Drainage Study Report 5.2.2. University of Pacific Ecological Engineering Research Program. Available online via. http://www.waterboards. ca.gov/rwqcb5/water_issues/tmdl/central_valley_projects/san_joaquin_oxygen/ required_studies/2013_model_peer_review/522_ag_drainage_rpt.pdf. Luo, Y., Zhang, X., Liu, X., Ficklin, D., Zhang, M., 2008. Dynamic modeling of organophosphate pesticide load in surface water in the northern San Joaquin Valley watershed of California. Environ. Pollut. 156, 1171e1181. Maybank, J., Yoshida, K., Grover, R., 1978. Spray drift from agricultural pesticide applications. J. Air Pollut. Control Assoc. 28, 1009e1014. Moriasi, D.N., Arnold, J.G., Van Liew, M.W., Bingner, R.L., Harmel, R.D., Veith, T.L., 2007. Model evaluation guidelines for systematic quantification of accuracy in watershed simulations. Trans. ASABE 50, 885e900. zieux, E., 2014. Pesticide transfer Mottes, C., Lesueur-Jannoyer, M., Bail, M.L., Male models in crop and watershed systems: a review. Agron. Sustain. Dev. 34, 229e250. Neitch, S.L., Arnold, J.G., Kiniry, J.R., Williams, J.R., 2011. Soil and Water Assessment Tool Theoretical Documentation. Blackland Research Center Temple, Texas. Poletika, N.N., Robb, C.K., 1998. A Monitoring Study to Characterize Chlorpyrifos

Concentration Patterns and Ecological Risk in an Agriculturally Dominated Tributary of the San Joaquin River. Dow AgroSciences LLC Study ENV96055. Raupach, M.R., Briggs, P.R., Ahmad, N., Edge, V.E., 2001a. Endosulfan transport: II. Modeling airborne dispersal and deposition by spray and vapor. J. Environ. Qual. 30, 729e740. Raupach, M.R., Briggs, P.R., Ford, P.W., Leys, J.F., Muschal, M., Cooper, B., Edge, V.E., 2001b. Endosulfan transport: I. Integrative assessment of airborne and waterborne pathways. J. Environ. Qual. 30, 714e728. Ravier, I., Haouisee, E., Clement, M., Seux, R., Briand, O., 2005. Field expriments for the evaluation of pesticide spray-drift on arable crops. Pest Manag. Sci. 61, 728e736. Reichenberger, S., Bach, B., Skitschak, A., Frede, H.G., 2007. Mitigation strategies to reduce pesticide inputs into ground- and surface water and their effectiveness; a review. Sci. Total Environ. 384, 35. Schulz, R., 2001. Comparison of spray drift- and runoff-related input of azinphosmethyl and endosulfan from fruit orchards into the Lourens River, South Africa. Chemosphere 45, 543e551. Schulz, M., Matthies, M., 2007. Runoff of Pesticides: Achievements and Limitations of Modelling Agrochemical Dislocation from Non-point Sources at Various Landscape Related Scales. Living Rev Landscape Res 1(1) URL. http://www.livi ngreviews.org/lrlr-2007-1. Stehle, S., Schulz, R., 2015. Agricultural insecticides threaten surface waters at the global scale. Proc. Natl. Acad. Sci. 112, 5750e5755. Teske, M., Bird, S., Esterly, D., Curbishley, T., Ray, S., Perry, S., 2002. AgDRIFT: a model for estimating near-field spray drift from aerial applications. Environ. Toxicol. Chem. 21, 659e671. USEPA, 1999. Background Document for the Scientific Advisory Panel on Orchard Airblast: Downwind Deposition Tolerance Bounds for Orchards. USEPA Archive document. Available online via. https://archive.epa.gov/scipoly/sap/meetings/ web/pdf/airblast.pdf. USDA, 2017. Soil Survey Geographic (SSURGO) Database. United State Department of Agriculture Natural Resources Conservation Service. Available online via. https://www.nrcs.usda.gov/wps/portal/nrcs/detail/soils/survey/?ci d¼nrcs142p2_053627. USGS, 2001. National Land Cover Database (NLCD) 2001. United States Geological Survey. Available online via. https://www.mrlc.gov/. Walker, R., 2001. Ecological Risk Assessment in a Tasmanian Agricultural Catchment. Ph.D. disservatin. University of Tasmania. Available online via. http://epri nts.utas.edu.au/22076/1/whole_WalkerRachel2001_thesis.pdf. Wauchope, R.D., Ahuja, L.R., Arnold, J.G., Bingner, R., Lowrance, R., van Genuchten, M.T., Adams, L.D., 2004. Software for pest-management science: computer models and databases from the United States department for agriculture - agricultural research service. Pest Manag. Sci. 59, 691e198. Zhang, X., Zhang, M., 2011. Modeling effectiveness of agricultural BMPs to reduce sediment load and organophosphate pesticides in surface runoff. Sci. Total Environ. 409, 1949e1958.