Optimizing pesticide water monitoring: balancing accuracy and efficiency using composite sampling and SEAWAVE-QEX Authors: Andy Jacobson, Zechariah Stone, Russell Krueger, Richard Brain Published: 2026 ABSTRACT Surface water monitoring is undertaken when characterization of empirically measured levels of pesticides in the environment is necessary and/or desired. While monitoring data can be used to validate exposure models, it is seldom used quantitatively to directly estimate pesticide exposure in regulatory risk assessments due to insufficient sampling frequency. Although daily monitoring is ideal for estimating pesticide exposure in surface water, the costs, resources, and logistics of daily sample collection are prohibitive for most monitoring programs. With the development of statistical tools, such as the United States Geological Survey’s Seasonal Wave with Streamflow Adjustment with Extended Capability (SEAWAVE-QEX) model, which can assist with estimating pesticide exposure in surface water using less frequent monitoring data, it is important for monitoring programs to properly evaluate which sampling approach (e.g., sample type and frequency) is most appropriate to satisfy their objectives. To better inform monitoring program design, this study leveraged high-frequency (daily or near-daily) monitoring datasets to assess the performance of extended (weekly and biweekly) composite sampling, as well as weekly and biweekly sampling used in conjunction with SEAWAVE-QEX, to estimate both short-term (acute) and long-term (chronic) pesticide exposure in surface water. Both weekly and biweekly composite sampling were shown to accurately estimate chronic pesticide exposure, whereas a proof-of-concept application of SEAWAVE-QEX using weekly composite sampling reasonably estimated acute exposure at one study site. Precipitation data also proved a viable, cost-saving substitute for streamflow data as a covariate in SEAWAVE-QEX. These findings offer practical guidance for optimizing pesticide water monitoring programs, balancing accuracy and efficiency.Supplementary InformationThe online version contains supplementary material available at 10.1007/s10661-026-15550-4. Introduction Pesticides are chemicals designed to control or eliminate pests, including weeds (herbicides), insects (insecticides), and fungi (fungicides). They are essential in agriculture for protecting crops, increasing yields, and ultimately ensuring food security for the world’s population, which is increasingly consequential given continued population growth under declining growth rates (United Nations Department of Economic and Social Affairs, Population Division [UN DESA], 2022) and shrinking agricultural land (Brain et al., 2023). While pesticides are vital for agriculture, they can also pose potential risks to human and environmental health. In the United States (US), the Federal Insecticide, Fungicide, and Rodenticide Act (FIFRA) authorizes the Environmental Protection Agency (EPA or the Agency) to regulate the registration, distribution, sale, and use of pesticides. As a result of FIFRA, the EPA is responsible for assessing the potential risks and benefits of pesticides, as well as ensuring the protection of public health, the environment, and federally threatened or endangered species protected by the Endangered Species Act (ESA). To evaluate the potential risks of pesticides, the EPA conducts human health and ecological risk assessments, which consider both the toxicity and exposure of pesticides. One key component of these assessments is the evaluation of exposure from pesticide runoff and spray drift to surface water, including both aquatic and drinking water. As part of a tiered approach, the EPA utilizes aquatic exposure model estimates and measured pesticide concentrations from surface water monitoring programs, when available. While the EPA utilizes pesticide surface water monitoring data to characterize empirically measured levels of pesticides and validate exposure model estimates, quantitative use of this data has been limited since water monitoring programs may underestimate pesticide exposure concentrations due to insufficient sampling frequency (USEPA, 2019a). While daily water monitoring for pesticides is widely acknowledged as the ideal approach, especially for capturing short-term (acute) exposure durations, weekly to monthly monitoring programs are more common due to the cost of sample collection and residue analysis (Vecchia, 2018). However, in 2019, the EPA published a white paper which evaluated and presented new statistical tools to quantitatively assess surface water monitoring data for pesticide risk assessment (USEPA, 2019a). Among these tools was the Seasonal Wave with Streamflow Adjustment with Extended Capability (SEAWAVE-QEX) model, which is a time-series regression model developed by the United States Geological Survey (USGS) to simulate daily pesticide concentrations in surface water using existing, non-daily monitoring data, seasonal patterns, and a covariate which is typically daily streamflow (Vecchia, 2018). Tools like SEAWAVE-QEX, or alternative approaches like universal kriging models (Mosquin et al., 2016, 2017, 2023b), help to address the need for estimating accurate pesticide exposure concentrations at sites with infrequent monitoring and enable additional monitoring data to be utilized quantitatively in pesticide risk assessments. While the sampling frequency of pesticide monitoring programs has the largest influence over accurately estimating pesticide exposures in surface water, the sample type collected by water monitoring programs also has a substantial impact. Historically, surface water monitoring data has been collected using grab samples; however, composite samples have become more prevalent due to the increased use of autosamplers (USEPA, 2019a). A grab sample is a sampling technique in which a single, instantaneous sample or measurement is taken at a specific point in time. While grab samples are cost effective and relatively simple to collect, they do not reflect the potential range of pesticide concentrations at a given site, which can lead to the under- or overestimation of pesticide exposure, especially if the timing of sampling either captures or misses a peak concentration. Comparatively, a composite sample is the average of a collection of individual samples obtained over defined time intervals (e.g., weekly or biweekly), often using autosamplers. Passive sampling is an alternative monitoring technique that uses specialized devices (e.g., Polar Organic Chemical Integrative Samplers (POCIS)) containing sorbent materials to accumulate dissolved chemicals over a defined deployment period, thereby providing time-weighted average (TWA) concentrations (Valenzuela et al., 2020). Due to its cost-effectiveness and simple logistics (Davies et al., 2024), passive sampling has been widely adopted in low- and middle-income countries and in remote or resource-limited settings where maintaining active sampling programs is challenging (Oltramare et al., 2023). Unlike grab samples, both composite and passive samples can provide more representative average concentrations over time, which is useful for assessing long-term (chronic) pesticide exposure. However, passive sampling is generally unable to determine the magnitude and timing of peak pesticide concentrations, especially during extended deployment periods. Similarly, composite samples may dilute peak concentrations with aliquots collected under baseflow or low-flow conditions when pesticide concentrations are minimal or absent, thereby limiting their ability to capture short-term exposure (Norman et al., 2020). To overcome this limitation, flow-triggered (event-based) sampling strategies, in which samples are collected during specific hydrological conditions rather than at fixed intervals, are designed to target acute exposures during runoff events; however, peak concentrations may still be missed if flow thresholds are not optimally set (Bundschuh et al., 2014). Considering that sample type and frequency significantly influence the accuracy of pesticide exposure estimates in surface water, it is crucial for monitoring programs to properly evaluate which sampling approach is most appropriate for their monitoring objectives. Furthermore, monitoring data can represent significantly different measured quantities depending on the chosen sampling strategy employed (Chow et al., 2020; Mosquin et al., 2023a). Monitoring program considerations for sample type and frequency include the duration(s) of toxicological concern (e.g., acute and/or chronic), the properties of the pesticide(s) being targeted (e.g., degradation, persistence, and/or mobility), the use pattern of the pesticide(s) (e.g., seasonal or year-round), the hydrological conditions of the monitoring location(s) (e.g., flow rates, rainfall amounts, and/or seasonal variability), the watershed characteristics of the monitoring location (e.g., size, soils, and/or slope), sampling technique options (e.g., systematic or stratified), and the monetary costs and resources required for sample collection and residue analysis. Furthermore, monitoring programs should also consider the sample frequency needed to use statistical tools like SEAWAVE-QEX, which can assist with estimating pesticide exposures in surface water with less frequent data. Previous studies have evaluated various sampling strategies for pesticide water monitoring. Bundschuh et al. (2014) evaluated and compared grab and composite sampling types for estimating pesticide exposure in surface water and observed that both grab and weekly composite samples significantly underestimated maximum pesticide concentrations, particularly during runoff events, but found that weekly composite samples were suitable for estimating chronic pesticide exposure. Similarly, Norman et al. (2020) discovered that weekly grab and composite samples both frequently failed to detect short-term peak pesticide concentrations in surface water, especially for insecticides, but noted that weekly composite samples were more accurate characterizations of pesticide concentrations when reporting average metrics versus extremes. Furthermore, Spycher et al. (2018) found that 14-day composite samples were adequate to assess chronic pesticide exposure in surface water, whereas composite samples of 3.5 days or less would be needed to assess acute exposure. Building on this, La Cecilia et al. (2021) presented the autonomous MS2Field platform, which performed continuous on-site sampling of 60 compounds at 20-min intervals using a high-resolution mass spectrometer. Their study demonstrated that 3.5-day composite samples underestimated concentration peaks by factors ranging from eight to more than 32. While advanced platforms like MS2Field provide much higher temporal resolution to capture concentration peaks, they involve substantial investment and technical expertise, which can limit their routine or large-scale deployment (Kizgin et al., 2025). While numerous past studies have evaluated sampling strategies for pesticide water monitoring, very few studies have evaluated the performance of sampling strategies utilized in conjunction with SEAWAVE-QEX. Vecchia (2018) described the SEAWAVE-QEX modeling methodology and data requirements, as well as tested the model’s robustness using scenario experiments that evaluated the bias and uncertainty of estimated annual maximum concentrations and concluded that model bias and uncertainty varied across the four pesticides analyzed. Additionally, the USEPA (2019a) quantitatively assessed the accuracy of SEAWAVE–QEX estimations using 7-day and 14-day stratified subsamples from monitoring datasets collected by the Atrazine Ecological Monitoring Program (AEMP) and the National Center for Water Quality Research (NCWQR) at Heidelberg University. The results of this study found that SEAWAVE-QEX produces summary statistics of pesticide concentrations that are within reasonable accuracy relative to measured concentrations. However, the study was primarily focused on the potential underestimation of annual maximum values by SEAWAVE-QEX, instead of overall model accuracy, and was restricted by the number of sites and compounds which had adequate data to support robust evaluation of SEAWAVE-QEX, thereby limiting confidence in the model’s broader application. Ryberg and Gilliom (2015) noted that SEAWAVE-Q (the predecessor to SEAWAVE-QEX) requires adequate data quality and sufficient sampling frequency, indicating that sparsely sampled sites and compounds may not be suitable for trend estimation. Rathjens et al. (2022) further found that weekly and monthly sampling utilized in conjunction with SEAWAVE-QEX was not suitable for estimating annual maximum concentrations for three herbicides in a small Belgian catchment and that the appropriateness of using the model likely depends on pesticide characteristics and application methods. However, the study noted that further research is needed to determine whether findings were applicable to other watersheds with different characteristics and/or monitoring data availability. The USEPA (2019a) also evaluated precipitation as a surrogate covariate in SEAWAVE-QEX and suggested it may be appropriate in specific cases where streamflow data are absent or not representative of natural hydrology. However, the USEPA Scientific Advisory Panel (Scientific Advisory Panel, 2020) cautioned that these analyses were limited in watershed type and geographic scope, and that precipitation’s relationship to streamflow is strongly influenced by landscape characteristics, runoff, and infiltration. Studies of streamflow dynamics in complex systems highlight why precipitation alone may not adequately capture actual streamflow. For example, Tursun et al. (2024) found that precipitation alone does not reliably capture the complex hydrologic responses that drive streamflow patterns in human-regulated watersheds, such as those with dams, diversions, or controlled releases. As an additional complicating factor, precipitation can have dual effects on pesticide concentrations: it can increase concentrations when runoff mobilizes recently applied pesticides or it can decrease concentrations through dilution when rainfall occurs over areas without recent pesticide applications. In watersheds with complex hydrology, where streamflow and precipitation may be poor covariates, process-based models may be needed to more accurately estimate pesticide concentrations in surface water. Unlike statistical approaches such as SEAWAVE-QEX, which rely on empirical relationships, these models explicitly represent hydrological processes (e.g., runoff/infiltration, preferential flow, erosion, and landscape connectivity), which are critical for pesticide transport. For example, Ammann et al. (2020) showed that conceptual, process-based hydrological models, informed by knowledge of watershed processes, can represent herbicide transport pathways in small agricultural watersheds, which would otherwise be difficult to represent without a mechanistic framework. Similarly, Commelin et al. (2022) showed that a substantial fraction of pesticide transport occurs in the particulate phase, highlighting the importance of properly accounting for particle-associated transport processes (e.g., erosion) linked to land use, rainfall characteristics, and pesticide properties. While process-based models can represent key pesticide transport processes, their performance depends on appropriate conceptualization, data availability, and watershed-specific knowledge, thereby limiting their straightforward or large-scale application (Ammann et al., 2020). Based on existing literature, as well as the considerable monetary costs, necessary resources, and logistical challenges of water monitoring programs, additional research is needed to evaluate the appropriate sampling frequency necessary to accurately estimate both short-term and long-term pesticide exposures in flowing surface water. Therefore, the primary objectives of this study were to utilize high-frequency (daily or near-daily) atrazine monitoring datasets from the AEMP and the NCWQR to evaluate the statistical performance of weekly and biweekly composite sampling, in addition to weekly and biweekly sampling used in conjunction with SEAWAVE-QEX, towards estimating short-term (acute) and long-term (chronic) pesticide exposures in surface water. Additional objectives were to compare the statistical performance of using precipitation versus streamflow data as a covariate in SEAWAVE-QEX, and to develop and evaluate a proof-of-concept approach for applying SEAWAVE-QEX to composite samples to improve acute exposure estimates. Although the AEMP and NCWQR datasets were previously used to evaluate the performance of SEAWAVE-QEX (USEPA, 2019a), this study employed substantially more comprehensive data from these monitoring programs, incorporating several additional sites and more than 100 additional site-years. In addition, the study evaluated all possible subsampling permutations for systematic weekly and biweekly sampling intervals and assessed overall model performance using additional statistical metrics. Overall, the goal of this study was to provide guidance for optimizing pesticide water monitoring programs, balancing accuracy and cost-effectiveness. Materials and methods Study area The study area includes nine watersheds from the AEMP, and four watersheds monitored by the NCWQR. The AEMP is a monitoring program with two decades of daily or near-daily sampling across several states in the Midwest and Southern USA, targeting headwater watersheds in the upper 20th percentile (or higher) of runoff vulnerability based on the USGS Watershed Regressions for Pesticides (WARP) model (Stone et al., 2025). Meanwhile, the NCWQR has one of the most long-term, intensive pesticide monitoring programs in the USA, with sample frequencies ranging from daily to monthly in the Lake Erie watershed. These watersheds are ideal for this study due to multiple years of daily or near daily atrazine sampling and flow data, which provide an ideal reference dataset. The thirteen study watersheds are located throughout six states in the Midwest and Southern USA (Fig. 1), range in size from 26 to 16,388 square kilometers, and are predominantly agricultural (> 72%) (Table 1). Additional details about the watersheds and monitoring years utilized in the study are provided in Table 1.Fig. 1Study areaTable 1Study area watershed details and monitoring years utilized for studyProgramSite NameStateYearsArea (km2)Agricultural Land (%)Atrazine Ecological Monitoring Program (AEMP)IA-04Iowa2010–20225890IA-05Iowa2010–20222684LA-04Louisiana2010–20228872MO-01Missouri2004–20223081MO-02/MO-02TMissouri2004–20227383MO-07NMissouri2010–20229083MO-08Missouri2010–202212690NE-04Nebraska2005–20225694TX-01Texas2010–202210690National Center for Water Quality Research (NCWQR)Honey CreekOhio2001–201738681Maumee RiverOhio2001–201716,38878Rock CreekOhio2001–20179076Sandusky RiverOhio2002–2017323980The data utilized from each site had a similar sampling frequency and type (daily or near-daily composite sampling). Two-day and weekly composite samples were also collected at the MO-02/MO-02T site from 2013 to 2017 Study area Table 1Study area watershed details and monitoring years utilized for studyProgramSite NameStateYearsArea (km2)Agricultural Land (%)Atrazine Ecological Monitoring Program (AEMP)IA-04Iowa2010–20225890IA-05Iowa2010–20222684LA-04Louisiana2010–20228872MO-01Missouri2004–20223081MO-02/MO-02TMissouri2004–20227383MO-07NMissouri2010–20229083MO-08Missouri2010–202212690NE-04Nebraska2005–20225694TX-01Texas2010–202210690National Center for Water Quality Research (NCWQR)Honey CreekOhio2001–201738681Maumee RiverOhio2001–201716,38878Rock CreekOhio2001–20179076Sandusky RiverOhio2002–2017323980The data utilized from each site had a similar sampling frequency and type (daily or near-daily composite sampling). Two-day and weekly composite samples were also collected at the MO-02/MO-02T site from 2013 to 2017 Study area watershed details and monitoring years utilized for study The data utilized from each site had a similar sampling frequency and type (daily or near-daily composite sampling). Two-day and weekly composite samples were also collected at the MO-02/MO-02T site from 2013 to 2017 Sampling frequency using extended composite sampling To evaluate the effectiveness of extended (e.g., weekly or biweekly) composite sampling in capturing acute and chronic pesticide concentrations in surface water, an experiment was conducted at the MO-02 AEMP sampling site, located in the headwater of Goodwater Creek in Missouri. This experiment was conducted over the growing season (generally April through July) for multiple years (2013 to 2017) to collect composite water samples at 2-day and weekly (7-day) durations, in addition to the routine daily (1-day) composite samples which were already being collected at the site as part of the AEMP. Routine daily composite samples were collected using an ISCO autosampler which collected eight ~ 50 mL aliquots throughout each day (i.e., one aliquot every 3 h, totaling ~ 400 mL per daily sample). To collect the 2-day and weekly composite samples, an additional ISCO autosampler was installed at the site (MO-02T), which collected ~ 50 mL aliquots every 6 h for both 2-day and weekly composites (i.e., totaling 400 mL and 1,400 mL per 2-day and weekly composite sample, respectively). All water samples were preserved in the field with the antimicrobial agent sodium omadine. Daily and weekly composite samples were retrieved from the autosampler approximately once a week, while the 2-day composite samples were retrieved every two weeks, and then transported to a laboratory for residue analysis. Laboratory methodologies had a limit of quantification (LOQ) for atrazine of 0.05 µg/L and concentrations below this LOQ were replaced with one half of the LOQ. Although there are more rigorous techniques for addressing concentrations below the LOQ (Helsel, 2005), the simple substitution method utilized in this study is consistent with procedures followed by the AEMP (Stone et al., 2025), which followed guidance issued by USEPA (2000). Moreover, concentrations below the LOQ only constituted approximately two percent of the samples analyzed. Once analyzed, the measured atrazine concentrations in the 2-day composite samples were aggregated to create biweekly (14-day) composite concentrations. Following laboratory analysis, atrazine concentration statistics (e.g., mean, median, maximum, and standard deviation) for daily, weekly, and biweekly composite sampling frequencies were calculated using Microsoft Excel and R software (R Core Team, 2021). To assess the stability of atrazine in the 2-day composite samples that were utilized to create biweekly composites, residue analysis for three metabolites was also performed. Kruskal Wallis non-parametric statistical tests were then applied to these metabolite data to confirm that atrazine remained stable in the 2-day composite samples despite being in the field for up to two weeks (Supplementary Information). To assess the performance of extended composite sampling frequencies, the individual weekly and biweekly composite sample concentrations were compared to the mean of the corresponding daily composite samples. For example, a weekly composite sample concentration collected between June 1 st and June 6th was compared against the average of the seven corresponding daily composite sample concentrations from that same period. In addition, maximum annual 4-day (acute) and 60-day rolling average (chronic) values were calculated for daily, weekly, and biweekly composite sampling frequencies, using the RcppRoll package (Ushey, 2024) in R software (R Core Team, 2021), and the results were compared to evaluate the performance of extended composite sampling frequencies in capturing acute and chronic atrazine concentrations in surface water. To calculate these rolling average values, any day without a measured concentration in the daily, weekly, and biweekly composite sampling datasets was assigned an estimated concentration value using stair-step interpolation, where non-sampled days were assigned the concentration of the nearest preceding date on which a pesticide concentration was measured. Although most interpolation occurred between weekly and biweekly composite measurements, interpolation was also occasionally used to estimate concentrations in the daily composite dataset if a sample was missed (e.g., due to an autosampler error). Sampling frequency using SEAWAVE-QEX To assess the performance of the SEAWAVE-QEX model in estimating both short-term and long-term pesticide exposure from less frequent water sampling data, the study area’s high-frequency (daily or near-daily) atrazine monitoring datasets were subsampled. This subsampling approach, which generates various low-frequency data series from high-frequency data, allows for a systematic evaluation of the impact sampling frequency has on the estimation accuracy of water quality indicators (Gao et al., 2020; Jones et al., 2012). The two subsampling intervals selected for study, weekly (7-day) and biweekly (14-day), were chosen based on their representativeness of potential water monitoring program frequencies. For each study area watershed, the near daily monitoring data was subsampled by creating multiple distinct datasets for each interval, where the number of subsamples generated for each sampling interval equals the number of days in that interval (e.g., biweekly subsampling yields 14 distinct subsamples; Fig. 2).Fig. 2Illustration of study’s subsampling methodology Illustration of study’s subsampling methodology Although extended composite samples are generally effective for estimating chronic pesticide exposure, they are typically ineffective in determining acute exposure. To help address this, a proof-of-concept approach utilizing SEAWAVE-QEX with composite samples was developed, using data from MO-02T, to improve composite sampling estimates of acute pesticide exposure in surface water. Since SEAWAVE-QEX is designed for sparse concentration data and carefully matched flow-concentration pairs, a new methodology was developed where composite sample concentrations were assigned to a single day within each sample’s duration interval. To avoid flattening the response curve or misrepresenting peak concentrations by assigning them to dates with maximum flow, composite concentrations were assigned to the date when the flow rate was nearest to the mean flow rate for each composite sampling interval. This methodology aligns the average concentration of each composite sample with a day exhibiting representative flow conditions within the same collection period, thereby enhancing SEAWAVE-QEX’s ability to interpolate daily concentrations from daily flow data and more effectively capture peak concentration events. Since SEAWAVE-QEX requires a year-round covariate, which is typically streamflow, to simulate daily pesticide concentrations in surface water, streamflow data collected from the study area watersheds were prepared for model input. While year-round streamflow data was available for the NCWQR watersheds, the AEMP watersheds only collected streamflow data during the sampling season (generally April through July) of most years. Therefore, monthly or annual average flow data was utilized to gap-fill the off-season streamflow datasets for all AEMP watersheds. Because the SEAWAVE-QEX model was implemented with the samcov parameter set to “partial” (indicating sampling occurs during only part of the year) for the AEMP watersheds, off-season covariate data are expected to have limited influence on model fitting, as concentration observations are restricted to the growing season. In addition, atrazine concentrations are typically negligible outside this period, and many headwater streams in the AEMP experience low-flow, no-flow, or frozen conditions at that time. Nonetheless, this assumption may introduce uncertainty in systems with more variable off-season hydrology or for compounds that persist beyond the growing season. To evaluate the performance of an alternative covariate in SEAWAVE-QEX, the model was also simulated using precipitation data as a covariate. Precipitation for the AEMP watersheds was obtained from nearby weather stations, while precipitation for the NCWQR watersheds was sourced from the Next Generation Weather Radar (NEXRAD) system (NOAA, 1992). Following SEAWAVE-QEX standard operating procedures (USEPA, 2019b), a transformation was performed on the precipitation data before being used as a covariate. This transformation was achieved by first taking the cube root of each value, followed by the antilog of the result, as shown in Eq. 1, where x is the original value:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$transformed\;precipitation=10^{\sqrt[3]x}$$\end{document}transformedprecipitation=10x3 Once monitoring and covariate data were prepared for model input, the SEAWAVE-QEX model was set up using default input values, except for the samcov variable. This variable, which specifies whether samples are collected year-round or for a shorter sampling period, was set based on watershed (USEPA, 2019b), resulting in a value of “partial” for AEMP watersheds and a value of “full” for NCWQR watersheds. After setup, the SEAWAVE-QEX model was run to generate 100 equally probable simulations, or “traces”, for each dataset. Using the simulation output from SEAWAVE-QEX, multiple centiles from the range of predictions were evaluated for model fit and outliers were systematically checked for and removed before calculating summary statistics. After calculating the root mean square error (RMSE), using R Statistical Software (Hamner & Frasco, 2018; R Core Team, 2021), for each model run (unique site, subsampling interval, set, and centile) a two-part filtering method considering both model performance and prediction magnitude was applied to the dataset. First, after calculating the maximum measured daily concentration for each rolling average period, a threshold for implausible model predictions was defined as three times the maximum measured value. Second, using the R software Metrics library, the z-score of the RMSE was computed within each group, defined by rolling average period, subsampling interval, and centile. The z-score quantified the deviation of a given RMSE from the mean of the group in terms of standard deviation, as shown in Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}= \frac{{RMSE}_{i}-{\mu}_{RMSE}}{{\sigma}_{RMSE}}$$\end{document}zi=RMSEi-μRMSEσRMSEwhere RMSEi is the RMSE of model run i, μRMSE is the mean RMSE across all runs in the group, and σRMSE is the standard deviation of RMSE in the group. Model runs with |zi|> 3 were considered outliers, since values beyond 3 standard deviations likely represent abnormal behavior. A model run was marked as an outlier and excluded from summary statistics if it met both criteria: predicted concentration > 3 times the measured maximum, and z-score > 3. This combination of criteria ensured that model runs with both an anomalously poor fit and unrealistic predictions were removed, thereby preserving valid yet variable model predictions. Once outliers were removed, daily concentration predictions from SEAWAVE-QEX were used to calculate maximum annual 4-day (acute) and 60-day rolling average (chronic) values. These values were compared to corresponding values calculated from the original high-frequency monitoring datasets to evaluate model accuracy, which was determined using the coefficient of determination (R2), RMSE, mean absolute error (MAE), mean error (ME), and Nash–Sutcliffe efficiency (NSE) calculated using R Software (R Core Team, 2021; Zambrano-Bigiarini, 2024). Results and discussion Sampling frequency using extended composite sampling Summary statistics and distributions of measured atrazine concentrations from daily (1-day), weekly (7-day), and biweekly (14-day) composite samples collected at MO-02/MO-02T from 2013 to 2017 are presented in Table 2 and Fig. 3. Generally, mean, median, and maximum atrazine concentrations decreased as the composite interval increased. Likewise, standard deviation declined with increasing composite intervals, further illustrating how temporal averaging reduces variability. In addition, concentration differences between daily and weekly/biweekly composite samples were substantially smaller for mean and median estimates than for maximum concentration estimates. This pattern suggests that weekly and biweekly composite samples perform better at estimating mean and median concentrations, which are useful for assessing long-term (chronic) pesticide exposure, than maximum (acute) concentrations. This aligns with previous findings by Norman et al. (2020), which noted that weekly composite samples were more accurate characterizations of pesticide concentrations when reporting average metrics versus extremes. Table 2Summary statistics of atrazine concentrations from daily, weekly, and biweekly composite samples collected at MO-02/MO-02T during the growing season (generally April through July) from 2013 to 2017. Concentrations below LOQ (0.05 µg/L) were replaced with one half of the LOQ (0.025 µg/L)YearComposite Sample Interval (Days)Samples AnalyzedMean (µg/L)Median (µg/L)Maximum (µg/L)Standard Deviation (µg/L)201311031.790.4864.086.877201.02 (− 0.77)0.38 (− 0.1)9.97 (− 54.11)2.23 (− 4.64)14120.85 (− 0.94)0.34 (− 0.14)5.71 (− 58.37)1.58 (− 5.29)20141994.622.4039.637.157164.35 (− 0.27)2.37 (− 0.03)18.19 (− 21.44)5.09 (− 2.06)1493.63 (− 0.99)1.97 (− 0.43)13.60 (− 26.03)4.27 (− 2.88)20151964.211.4619.054.917183.53 (− 0.68)1.41 (− 0.05)12.91 (− 6.14)3.94 (− 0.97)14122.71 (− 1.5)1.19 (− 0.27)11.54 (− 7.51)3.28 (− 1.63)2016110210.262.64149.9820.87197.36 (− 2.9)1.50 (− 1.14)58.59 (− 91.39)14.0 (− 6.8)14106.34 (− 3.92)1.51 (− 1.13)35.75 (− 114.23)11.1 (− 9.7)201711027.353.2649.078.677195.55 (− 1.8)2.27 (− 0.99)18.35 (− 30.72)6.15 (− 2.52)14104.44 (− 2.91)2.29 (− 0.97)19.44 (− 29.63)5.96 (− 2.71)Values in parentheses represent the concentration difference (µg/L) relative to the 1-day composite sample intervalFig. 3Atrazine concentration distributions for daily, weekly, and biweekly composite samples collected at MO-02/MO-02T from 2013 to 2017. Concentrations below LOQ (0.05 µg/L) were replaced with one half of the LOQ (0.025 µg/L) Table 2Summary statistics of atrazine concentrations from daily, weekly, and biweekly composite samples collected at MO-02/MO-02T during the growing season (generally April through July) from 2013 to 2017. Concentrations below LOQ (0.05 µg/L) were replaced with one half of the LOQ (0.025 µg/L)YearComposite Sample Interval (Days)Samples AnalyzedMean (µg/L)Median (µg/L)Maximum (µg/L)Standard Deviation (µg/L)201311031.790.4864.086.877201.02 (− 0.77)0.38 (− 0.1)9.97 (− 54.11)2.23 (− 4.64)14120.85 (− 0.94)0.34 (− 0.14)5.71 (− 58.37)1.58 (− 5.29)20141994.622.4039.637.157164.35 (− 0.27)2.37 (− 0.03)18.19 (− 21.44)5.09 (− 2.06)1493.63 (− 0.99)1.97 (− 0.43)13.60 (− 26.03)4.27 (− 2.88)20151964.211.4619.054.917183.53 (− 0.68)1.41 (− 0.05)12.91 (− 6.14)3.94 (− 0.97)14122.71 (− 1.5)1.19 (− 0.27)11.54 (− 7.51)3.28 (− 1.63)2016110210.262.64149.9820.87197.36 (− 2.9)1.50 (− 1.14)58.59 (− 91.39)14.0 (− 6.8)14106.34 (− 3.92)1.51 (− 1.13)35.75 (− 114.23)11.1 (− 9.7)201711027.353.2649.078.677195.55 (− 1.8)2.27 (− 0.99)18.35 (− 30.72)6.15 (− 2.52)14104.44 (− 2.91)2.29 (− 0.97)19.44 (− 29.63)5.96 (− 2.71)Values in parentheses represent the concentration difference (µg/L) relative to the 1-day composite sample interval Summary statistics of atrazine concentrations from daily, weekly, and biweekly composite samples collected at MO-02/MO-02T during the growing season (generally April through July) from 2013 to 2017. Concentrations below LOQ (0.05 µg/L) were replaced with one half of the LOQ (0.025 µg/L) Values in parentheses represent the concentration difference (µg/L) relative to the 1-day composite sample interval Atrazine concentration distributions for daily, weekly, and biweekly composite samples collected at MO-02/MO-02T from 2013 to 2017. Concentrations below LOQ (0.05 µg/L) were replaced with one half of the LOQ (0.025 µg/L) Extended composite concentrations vs. corresponding daily means Mean atrazine concentrations calculated from daily composite samples were compared to corresponding weekly and biweekly composites to evaluate their correlation over time. Figures 4 and 5 illustrate the atrazine concentration time-series for weekly and biweekly composite samples against the corresponding daily composite means monitored at MO-02/MO-02T from 2013 to 2017. The weekly and biweekly composite samples generally reflect the trends observed in corresponding daily composite means, with rises and falls in concentration occurring on similar time frames. These trends corroborate work by Norman et al. (2020), which reported that weekly composite samples effectively capture the dynamics of daily composite means. The largest differences between the weekly and biweekly composite concentrations and the corresponding daily composite means occurred during peak concentration periods, with more frequent and larger differences occurring between biweekly composite samples and daily composite means.Fig. 4Comparison of weekly composite sample atrazine concentration time-series to corresponding daily composite means monitored at MO-02/MO-02T from 2013 to 2017Fig. 5Comparison of biweekly composite sample atrazine concentration time-series to corresponding daily composite means monitored at MO-02/MO-02T from 2013 to 2017 Comparison of weekly composite sample atrazine concentration time-series to corresponding daily composite means monitored at MO-02/MO-02T from 2013 to 2017 Comparison of biweekly composite sample atrazine concentration time-series to corresponding daily composite means monitored at MO-02/MO-02T from 2013 to 2017 Figure 6 presents scatter plots comparing weekly and biweekly composite concentrations to their corresponding daily composite means monitored at MO-02/MO-02T, overlaid with a least-squares linear regression line (solid line) and a 1:1 reference line (dotted). For both composite lengths (weekly and biweekly), the shaded prediction interval encompasses 95% of the data.Fig. 6Comparison of weekly (left) and biweekly (right) composite sample atrazine concentrations to corresponding daily composite means monitored at MO-02/MO-02T Comparison of weekly (left) and biweekly (right) composite sample atrazine concentrations to corresponding daily composite means monitored at MO-02/MO-02T Figure 6 illustrates that weekly and biweekly composite samples yielded concentrations comparable to the corresponding daily composite means, as evidenced by the clustering of data points along the 1:1 line, although weekly composite samples were slightly more analogous. For both composite lengths, the regression line is slightly above the 1:1 line at lower concentrations, while slightly below the 1:1 line at higher concentrations, which suggests that composite sampling tends to reduce the measured range of concentrations. Overall, most composite sample concentrations were within 1 µg/L of the corresponding daily composite means, with 87% for the weekly composites and 81% for the biweekly composites. The largest variance between the composite sample concentrations and the corresponding daily composite means was 7.71 µg/L for weekly composites and 6.71 µg/L for biweekly composites. For both composite lengths, over 92% of the differences with corresponding daily composite means were within 3 µg/L. Acute and chronic exposure Figure 7 illustrates the maximum annual 4-day (acute) and 60-day (chronic) rolling average atrazine concentrations for daily, weekly, and biweekly composite samples collected at MO-02/MO-02T. Rolling average atrazine concentrations varied by composite length and year, with daily composite concentrations generally the highest, followed by weekly and then biweekly composite concentrations. The contrast in 4-day rolling average atrazine concentrations between the three composite lengths is stark, with the difference between concentrations as high as 29.41 µg/L between daily and weekly composites, and as high as 52.25 µg/L between daily and biweekly composites. Conversely, 60-day rolling average atrazine concentrations for the three composite lengths were very similar. The daily and weekly composite concentrations differed by a maximum of only 1.54 µg/L, while the daily and biweekly results differed by as much as 3.16 µg/L. These results indicate that weekly and biweekly composite sampling for atrazine can yield realistic estimates for chronic exposure, but do not accurately estimate acute exposure. This aligns with previous studies (Bundschuh et al., 2014; Norman et al., 2020; Spycher et al., 2018) which also found that extended composite samples are more suitable for chronic than acute exposure assessment.Fig. 7Maximum annual 4-day and 60-day rolling average atrazine concentrations for daily, weekly, and biweekly composite samples collected at MO-02/MO-02T Maximum annual 4-day and 60-day rolling average atrazine concentrations for daily, weekly, and biweekly composite samples collected at MO-02/MO-02T Sampling frequency using SEAWAVE-QEX To evaluate the accuracy of SEAWAVE-QEX using subsampled concentration data, maximum annual rolling average values for an acute (4-days) and chronic exposure period (60 days) were examined. The 50th, 75th, and 95th centile predictions from SEAWAVE-QEX were evaluated for simulations featuring systematic weekly (7-day) and biweekly (14-day) subsampling (Figs. 8 and 9). The horizontal clustering in these figures results from evaluating all possible subsampling permutations, producing 7 or 14 distinct model predictions per monitored site-year for weekly and biweekly intervals, respectively. All permutations were included because each has an equal likelihood of occurring.Fig. 8Evaluation of 4-day rolling averages derived from SEAWAVE-QEX predictions using weekly and biweekly subsamplingFig. 9Evaluation of 60-day rolling averages derived from SEAWAVE-QEX predictions using weekly and biweekly subsampling Evaluation of 4-day rolling averages derived from SEAWAVE-QEX predictions using weekly and biweekly subsampling Evaluation of 60-day rolling averages derived from SEAWAVE-QEX predictions using weekly and biweekly subsampling For acute exposure periods, predictions from weekly subsampling generally outperform those from biweekly subsampling (Fig. 8). This was evidenced by a stronger correlation of predicted values to measured values (higher R2 values), as well as a lower magnitude of relative error (lower RMSE and MAE values). While the 50th centile predictions from weekly subsampling runs showed a better overall fit than those of the 75th centile, they tended to underpredict the measured results in 64.5% of the model simulations for AEMP watersheds, with a mean error (ME) of − 4.82 µg/L. The 75th centile predictions, which are the suggested centile from the USEPA (USEPA, 2019a), showed a similar correlation to the measured values, albeit higher levels of error. While slightly more error prone, the 75th centile predictions underpredicted less frequently (44.1%) and erred on the side of conservatism (ME = 4.81 µg/L). For the 50th and 75 centile predictions, NSE values greater than 0 demonstrate that the model produces maximum annual values more accurately than a baseline prediction equal to the mean of measured values. Maximum annual 60-day rolling average concentrations calculated from the SEAWAVE-QEX simulations exhibit a similar trend to those from the 4-day rolling averages (Fig. 9). The results from simulations using weekly subsampling statistically outperformed biweekly subsampling for chronic exposure, indicated by higher NSE values, providing more representative estimates of measured values. Again, the 50th centile results were statistically closer to measured values than those from the 75th centile, albeit 50th centile results more underpredicted measured values (71.7% vs 58.0%). Comparatively, the 95th centile predictions showed poor overall correlation to measured values, and the magnitude of overpredicted values was higher. Collectively, these results suggest that collecting a weekly sample and utilizing 75th centile predictions from SEAWAVE-QEX may be a suitable replacement for a daily sampling regimen for estimating chronic pesticide exposure. Nonetheless, given the slight underprediction associated with the 75th percentile predictions (ME = 1.24 µg/L), monitoring programs may opt to use alternative percentiles (e.g., 50th or 95th) to generate exposure estimates that are either more accurate or more conservative, depending on program objectives. Overall, SEAWAVE-QEX predictions were found to yield more accurate estimates of chronic atrazine exposure than estimates of acute exposure, irrespective of the subsampling schema or centile evaluated. This highlights the model’s strength in estimating long-term average concentrations from sparse data. Evaluation of precipitation as an alternative covariate To leverage data availability and explore cost-reduction opportunities for monitoring programs, the performance of SEAWAVE-QEX using precipitation data as an alternative covariate to streamflow was evaluated. Summary statistics comparing flow versus precipitation as a covariate were compared, assuming 75th centile predictions for a weekly sampling schema (Fig. 10). NSE values for both exposure durations demonstrate that the model more skillfully predicts annual maximums than a naïve prediction using the mean of the observations, with the precipitation covariate performing better than the flow covariate for acute exposure and slightly worse for chronic exposure. While the correlations of maximum predicted values versus maximum measured values (R2) were slightly better using flow as a covariate, the magnitudes of the relative errors (RMSE and MAE) were similar. This finding is significant because precipitation data is more publicly available and typically less expensive to measure than continuous streamflow data. This also suggests that precipitation can serve as a viable alternative covariate in SEAWAVE-QEX for estimating maximum rolling average concentrations without sacrificing substantial loss of accuracy. However, these findings may be less applicable to managed watersheds, where human interventions weaken the natural coupling between precipitation and streamflow.Fig. 10Comparison of SEAWAVE-QEX model performance for acute (left) and chronic (right) atrazine exposure using flow as a covariate versus precipitation as a covariate Comparison of SEAWAVE-QEX model performance for acute (left) and chronic (right) atrazine exposure using flow as a covariate versus precipitation as a covariate Proof-of-concept application of SEAWAVE-QEX with composite samples for acute exposure assessment To evaluate a proof-of-concept application of SEAWAVE-QEX for improving composite sampling estimates of acute pesticide exposure in surface water, performance statistics for estimating maximum annual 4-day (acute) rolling average concentrations were compared using weekly and biweekly composite sampling in conjunction with SEAWAVE-QEX (Table 3). This comparison showed that using both streamflow and precipitation data as covariates, the correlation of model results and relative error were improved when utilizing weekly composite samples. For weekly composite samples, the 75th centile and 95th centile predictions have a very similar R2, with 95th centile predictions yielding better concentration predictions (smaller RMSE and MAE; Table 3). Overall, the 95th centile SEAWAVE-QEX predictions using weekly composite samples and streamflow as a covariate yielded the best concentration predictions. These results suggest that this proof-of-concept approach can improve predictions of acute exposure when using composite samples, addressing a key limitation of composite sampling alone. However, because this analysis was conducted at a single, geographically specific study site and under a limited set of conditions, the extent to which these findings generalize to other pesticides, watershed types, or hydrologic settings remains uncertain, and broader validation is needed before wider application. Table 3Performance statistics for estimating maximum annual 4-day (acute) rolling average concentrations at MO-02T using weekly/biweekly composite samples and streamflow/precipitation covariates in SEAWAVE-QEXComposite Length (days)CovariateCentileRolling Average Length (days)R2RMSEMAE7Flow0.7540.97212.02810.3027Precipitation0.7540.9818.3487.55614Flow0.7540.29522.74219.79714Precipitation0.7540.77712.46111.987Flow0.9540.9695.3644.3927Precipitation0.9540.9799.5647.27514Flow0.9540.23833.50222.91814Precipitation0.9540.82345.06836.0657Raw MO-02T Composite40.9817.42415.30514Raw MO-02T Composite40.91726.94821.699 Table 3Performance statistics for estimating maximum annual 4-day (acute) rolling average concentrations at MO-02T using weekly/biweekly composite samples and streamflow/precipitation covariates in SEAWAVE-QEXComposite Length (days)CovariateCentileRolling Average Length (days)R2RMSEMAE7Flow0.7540.97212.02810.3027Precipitation0.7540.9818.3487.55614Flow0.7540.29522.74219.79714Precipitation0.7540.77712.46111.987Flow0.9540.9695.3644.3927Precipitation0.9540.9799.5647.27514Flow0.9540.23833.50222.91814Precipitation0.9540.82345.06836.0657Raw MO-02T Composite40.9817.42415.30514Raw MO-02T Composite40.91726.94821.699 Performance statistics for estimating maximum annual 4-day (acute) rolling average concentrations at MO-02T using weekly/biweekly composite samples and streamflow/precipitation covariates in SEAWAVE-QEX SEAWAVE-QEX limitations Despite its utility, SEAWAVE-QEX has several important limitations that should be considered when interpreting results. Model performance is sensitive to sampling frequency and data availability, with reduced accuracy observed under sparse or irregular sampling conditions, particularly for estimating short-term (acute) exposure, consistent with Ryberg and Gilliom (2015) and Rathjens et al. (2022). In addition, the model relies on the strength and representativeness of the selected covariate (e.g., streamflow or precipitation), and performance may decline in watersheds where the covariate is weakly correlated with pesticide transport, such as highly managed or hydrologically complex systems (USEPA, 2019a; Tursun et al., 2024). The model also assumes consistent seasonal patterns and may be less reliable for pesticides with irregular application timing or highly variable transport pathways. A further limitation is that the model can generate unrealistic extreme predictions, requiring post-processing through outlier filtering. Although conservative criteria were used to remove clearly implausible values, some unrealistic predictions may remain, while stricter filtering could exclude valid extreme events. As a result, summary statistics and estimates of high-end exposure may be sensitive to the choice of filtering approach. Collectively, these limitations indicate that SEAWAVE-QEX is most appropriate for well-sampled systems with strong covariate relationships and should be applied cautiously in data-limited or hydrologically complex watershed settings. Conclusions To evaluate the appropriate sampling frequency necessary to accurately estimate both short-term (acute) and long-term (chronic) pesticide exposure in flowing surface water, this study utilized high-frequency (daily or near-daily) atrazine monitoring datasets to assess the statistical performance of weekly and biweekly composite sampling, as well as weekly and biweekly sampling used in conjunction with the SEAWAVE-QEX model. The results of this study found that both weekly and biweekly composite sampling performed well at estimating chronic, maximum 60-day rolling average atrazine concentrations, but considerably underestimated acute, 4-day rolling average concentrations. These composite sampling results suggest that pesticide water monitoring programs primarily interested in chronic exposure could potentially utilize weekly or biweekly composite sampling to reduce costs, while maintaining the accuracy of their exposure estimates. However, if acute exposure is of interest, then composite sampling with intervals of less than one week would likely be required. While this composite sampling experiment was limited to a single watershed and pesticide, previous studies have arrived at similar conclusions using different pesticides and study areas. This study further found that weekly subsampling, used in conjunction with SEAWAVE-QEX 75th percentile predictions, also performed reasonably well at estimating chronic, 60-day rolling average atrazine concentrations (ME = − 1.24 µg/L), although not as accurately as weekly and biweekly composite sampling. The modest observed underprediction is attributable to the use of the 75th percentile predictions from SEAWAVE-QEX. Consequently, monitoring programs may consider using alternative percentile predictions to generate exposure estimates that better align with specific program objectives, whether emphasizing improved accuracy or adopting more conservative, protective assessments. Conversely, biweekly subsampling with SEAWAVE-QEX was unable to accurately estimate chronic atrazine exposure, and neither of the subsampling strategies (weekly or biweekly) were suitable for estimating acute, 4-day rolling average atrazine concentrations. Furthermore, precipitation data was found to perform comparably to streamflow as a covariate in SEAWAVE-QEX. This is noteworthy since precipitation data is more publicly available, which reduces the need for continuous streamflow data, as well as the associated cost. However, because the watersheds included in this study are relatively natural and exhibit a tighter coupling between precipitation and streamflow, this finding may be less applicable to managed watersheds, where human interventions (e.g., dams, diversions, and controlled releases) produce more complex flow responses to precipitation. To address the inaccuracies in estimating acute atrazine exposures, a proof-of-concept approach was developed and demonstrated that weekly composite sampling utilized in conjunction with SEAWAVE-QEX can reasonably estimate acute, 4-day rolling average atrazine concentrations at one study site. This addresses a gap, allowing composite sampling data to be used for both chronic and acute exposure assessments with the aid of modeling. While daily sampling is optimal for estimating pesticide exposure in surface water, the costs, resources, and logistics of daily sample collection are prohibitive for most water monitoring programs. Consequently, the key findings in this study are consequential towards reducing the costs of monitoring programs while also preserving the accuracy of exposure estimates. Although this study’s findings are promising, selecting the appropriate sampling frequency and program design for pesticide monitoring in surface water is a complex decision that is dependent on many factors, including pesticide properties, environmental factors, monetary/resource constraints, and monitoring objectives. Therefore, because this study focused on a single pesticide (atrazine) and a limited set of watersheds, the extent to which these findings generalize to other pesticides, watershed types, and hydrologic conditions remains uncertain. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (PDF 259 KB) Supplementary file1 (PDF 259 KB)