The study site (Fig. 1) is located inside a flood detention basin in the 1.8 km2 Loechernbach catchment, southwest Germany. Catchment elevation ranges between 213 m a.s.l. at the outlet and 378 m a.s.l. in the western corner. Mean precipitation was 800 mm a-1 between 2009 and 2018 and mean air temperature 11.3 ∘C. Soils mainly consist of calcaric regosols which are formed on aeolian loess and have a typical grain size distribution of 10 % sand, 80 % silt and 10 % clay. Most of the catchment is dedicated to large artificial vineyard terraces (71 %), while croplands occupy the valley bottoms (20 %). Forest only accounts for a small portion (9 %) and is limited to the most elevated part of the catchment. This partition in land use is reflected in the main application areas of pesticide types. Fungicides are applied on vineyard terraces, while herbicides are mainly applied to the cropland in the flat valleys. Large parts of the catchment are drained by a sub-surface pipe network (Fig. 1) connecting vineyards and paved roads to the main channel in the valley. This drainage network causes fast downstream transport of storm water and suspended sediments (Gassmann et al., 2012). In addition, fields in the valley bottoms are drained by a secondary network of smaller and usually shorter field drains that are either connected to the primary drainage network or directly connected to the stream (Schuetz et al., 2016). A 20 000 m3 detention basin was built at the outlet of the Loechernbach catchment to prevent flooding of the downstream village. Inside the detention basin, a 258 m2 vegetated surface flow constructed wetland and a 105 m2 retention pond (maximum depth 1.5 m) are connected in series parallel to the course of the Loechernbach stream. A small dam diverts all flow through the vegetated treatment systems during base flow conditions but allows water to bypass the VTS during large discharge events. The wetland has been in operation since 2010, and its succession was studied by Schuetz et al. (2012). The pond was added to the system in January 2016. The entire detention basin is sealed by an impermeable clay layer that prevents leakage to groundwater. Water residence times range from less than 1 h during flood events up to several days during extreme low flow conditions.

In this study, we focused on six target compounds, including the fungicides boscalid and penconazole, the herbicides metazachlor and flufenacet, and the TPs metazachlor sulfonic acid (met-ESA) and metazachlor oxalic acid (met-OA). Selected physicochemical properties of the target compounds are listed in Table 1. According to the Pesticide Properties DataBase (Lewis et al., 2016), the contaminants can be classified as low (boscalid) to moderately soluble in water and very mobile (TPs) to slightly mobile (fungicides). The target fungicides are considered moderately fast degradable in the water phase and persistent in soils, while the target herbicides are considered stable in the water phase and non-persistent in soils. TPs of metazachlor are considerably more persistent in soil than their PC. The fungicides are considered stable with respect to hydrolysis but degradable via photolysis, while the herbicides are stable regarding both.

Streamflow was measured every minute between April 2016 and September 2017 at two gauges about 200 m upstream of the treatment system (G1) and at its outlet (G2). Water levels at G1 were recorded inside a 1.37 m standard H flume (Bos, 1989) by means of a pressure transducer (Decagon CTD-10) and related to discharge using a standard rating curve. At G2 water levels were measured in a rectangular cross-section 2 m ahead of the detention basin outlet by a radar gauge (Vegapuls 61). The corresponding rating curve for G2 accounted for complete submergence of the control gate valve (Peter, 2005). Pesticide monitoring at G1 and G2 consisted of five manual sample collections during stationary flow conditions and 10 automated event samplings during discharge events. Event sampling was triggered when a water level increase of more than 3 cm/h was recorded at G1. An automatic sampler (ISCO 3700) started to fill pairs of 900 mL glass bottles at 0, 0.5, 1, 2, 6, and 12 h after activation. A second automatic sampler (ISCO 3700) was launched at G2 following the same sampling scheme but with a time lag of 1 h to account for transit between G1 and G2. All samples were recovered from the study site within 24 h after sampling and cooled until analysis. Sampling was complete except for one case. Due to an accident, we lost the first sample of event 9 (8 October 2017, 03:30 CET). As concentrations in the first samples were usually very low (Fig. 3) and not considered to markedly influence mass calculation, we assumed that all contaminants in this sample had zero concentration and left this event in our dataset.

The following analytical methods were used for determining pesticide levels in the water samples. Analytical standards of boscalid (99.9 %), penconazole (99.1 %), metazachlor (99.6 %), and flufenacet (99.5 %) and the internal standards Diuron-D6 (99 %) and Terbutryn-D5 (98.5 %) already dissolved in acetonitrile (100 µg mL-1) were purchased from Sigma-Aldrich Chemie GmbH (Steinheim, Germany). Met-ESA (95 %) and met-OA (98.8 %) were received from Neochema (Bodenheim, Germany). Acetonitrile (LC-MS grade; VWR International GmbH, Darmstadt, Germany) was used as organic mobile phase in chromatography and for the preparation of stock solutions. Aqueous mobile phase was prepared with ultrapure water (Membra Pure, Germany; Q1:16.6 mΩ and Q2: 18.2 mΩ.

Preparation of environmental samples (approx. 1 L) was done by filtering with a folded filter (type 113 P Cellulose ø 240 mm). The supernatant was spiked with the internal standard Diuron-D6 (10 µL of 10 mg L-1). The extraction procedure was solid-phase extraction (SPE). Cartridges (CHROMABOND^® HR-X 6 mL/200 mg) were conditioned with 10 mL methanol and washed with 10 mL pure water. 90 µL of the extract was spiked with 10 µL of Terbutryn-D5 as an internal standard. Each sample was a double determination. Measurements of environmental samples were conducted with a triple quadrupole (Agilent Technologies, 1200 Infinity LC System and 6430 Triple Quad, Waldbronn, Germany). Mobile phases were 0.01 % formic acid (A) and acetonitrile (B) with a flow of 0.4 mL min-1. The gradient was as follows: 0–1 min (10 % B), 1–11 min (10 %–50 % B), 11–18 min (50 %–85 % B), 18–21 min (85 %–90 % B), 21–24 min (90 % B), 24–26 min (90 %–10 % B), and 26–30 (10 % B). A NUCLEODUR®RP-C18 (125/2; 100–3 µm C18 ec) column (Macherey Nagel, Düren, Germany) was used as stationary phase with a set oven temperature of T= 30 ∘C. Calibration curves were prepared in pure water. The linearity was evaluated by preparing three curves with 10 calibration points in the range 1–500 µg/L. The standard curves were then extracted according to the protocol and analyzed using liquid chromatography with tandem mass spectrometry (LC-MS/MS). The calculated linear regression values (R2) were very good, with R2-values > 0.999. The linearity between the peak area and concentration of substances was obtained in a range of 0–5 µg L-1. Hence limits of detection (LOD) and quantitation (LOQ) were calculated with DINTEST (2003) according to DIN 32645 considering an enrichment factor of 5000. LOD and LOQ amounted to 0.4 and 1.3 ng L-1 (boscalid), 0.3 and 0.9 ng L-1 (penconazole), 0.3 and 1.2 ng L-1 (metazachlor), and 0.4 and 1.3 ng L-1 (flufenacet), as well as 0.6 and 2.2 ng L-1 (met-ESA) and 0.5 and 1.6 ng L-1 (met-OA) considering an enrichment factor of 5000. A detailed analysis of measurement precision can be found in the Supplement (Text S1).

Contaminant concentrations in stream water (G1) differed clearly depending on the flow conditions (Fig. 2). During stationary flow, concentrations of boscalid and the TPs of metazachlor ranged in the order of tens of nanograms, while penconazole, metazachlor, and flufenacet only occasionally exceeded the LQ. During discharge events in contrast, peak concentrations varied from a few nanograms (flufenacet) to several milligrams per liter (boscalid), spanning a range of 6 orders of magnitude. Concentration increase during events compared to stationary flow was different among the compounds. Median concentration of boscalid increased by a factor of 48, while concentrations of met-ESA and met-OA only increased by a factor of 3 and 5, respectively. Similar patterns were found for contaminant mass. Contaminant mass mobilized in the catchment during discharge events ranged from several hundreds of micrograms (flufenacet) to several hundreds of milligrams (boscalid) and even several grams in exceptional cases (boscalid, metazachlor). Based on compound medians, about 76 times more boscalid but only about 4 times more met-OA were transported during discharge events than during an equally long period under stationary flow conditions.

The 10 events were characterized by different discharge magnitudes and dynamics (Fig. 3). Mean discharge during the events ranged between 0.7 (E10) and 32.0 L s-1 (E2), with respective peak values between 4.4 (E10) and 199.7 L s-1 (E2). The recorded event hydrographs included events with one single discharge peak (E4, E5, E6, E10), with one major peak followed by one or more secondary peaks (E2, E3, E7, E9), and events in which a major peak followed an earlier smaller peak (E1, E8). In most cases discharge had recessed to pre-event levels by the end of the 12 h sampling procedure; only E1 and E2 showed ongoing flow recession. In many cases, concentrations in the final event samples were still elevated compared to pre-event conditions. However, due to flow recession, mass flux was usually very low by the time the last sample was collected (Fig. S1).

Cluster A (Fig. 4) was characterized by absence of a clear peak during the first 2 h of sampling but elevated concentrations during later times, resulting in low iDS. Cluster B showed a quick response; i.e., concentrations increased sharply within the first 30 min. Concentrations were the highest of all clusters and still elevated in the last sample compared to pre-event levels. Cluster C was characterized by a clear peak within the first 2 h and a low tailing and was the cluster with highest median iDS. Cluster D showed the most inconsistent pattern, and maximum concentrations appeared later compared to clusters B and C. A relatively clear pattern was evident in the attribution of compounds to the clusters. Chemographs of the fungicides boscalid and penconalzole were mainly assigned to cluster B, while the herbicides and the TPs were assigned to the remaining three clusters. Cluster A was composed of herbicide and TP chemographs, particularly from events with multiple discharge peaks. Cluster D represented chemographs of herbicides and TPs mainly during the events E5 to E8, which were all characterized by sharp discharge peaks during periods of generally low flow (Fig. 3). Almost all chemographs of the events E2 and E4 were attributed to cluster C.

Contaminants were mitigated in the wetland in terms of both RC (Eq. 1) and RM (Eq. 2). RC (Fig. 5a) was close to zero for boscalid and poorly constrained for the remaining compounds during stationary conditions, partly due to an insufficient number of detections. During discharge events, in contrast, peak concentrations of all compounds were clearly reduced. While RC was narrowly constrained for the fungicides and herbicides, TPs exhibited higher variability. Mean RC and corresponding standard deviations were 29.8 ± 18.4 % (boscalid), 42.1 ± 11.5 % (penconazole), 47.9 ± 16.4 % (metazachlor), 53.8 ± 22.6 % (flufenacet), 29.5 ± 84.7 % (met-ESA), and 47.9 ± 29.5 % (met-OA), respectively. RC was clearly different among chemograph clusters, with the lowest values in cluster A and highest values in cluster D (Fig. 5b). Moreover, RC was higher for the lower half of peak concentrations than for the upper half (Fig. 5c) and systematically increased with dispersion sensitivity (Fig. 5d). RC was also related to discharge conditions. Highest RC values were reached when mean discharge was low (Fig. 5e) but the ratio of maximum to mean discharge was elevated (Fig. 5f), i.e., in events characterized by low pre-event discharge and sharp discharge peaks (in particular events attributed to cluster D in Fig. 4). Although there was evidence for major water surpluses and deficits in the event water balance between G1 and G2, particularly in events with low discharge such as events E6 and E10, an imbalanced water balance had only minor effects on RC (Fig. 5g). We did not find clear relationships between RC and compound properties such as KfOC, water solubility, or soil half-life (Fig. 5h–j).

Relative mass removal during discharge events (Fig. 6a) resulted in smaller rates and higher variability compared to RC. Mean RM and corresponding standard deviations were 7.7 ± 29.6 % (boscalid), 17.3 ± 26.0 % (penconazole), 18.1 ± 27.8 % (metazachlor), 27.0 ± 28.1 % (flufenacet), 35.2 ± 68.4 % (met-ESA), and 44.0 ± 28.7 % (met-OA). These values show that mass removal was limited for most compounds. Although the general pattern in RM for the different compounds was similar to RC, behavior of RM among the chemograph clusters was different. While RC increased from cluster A to cluster B and C, RM decreased (Fig. 6b). Cluster D exhibited high values of both RC and RM. No clear response was found to different levels of input mass (Fig. 6c); however, median RM was lowest when peaks in chemographs were sharpest (Fig. 6d). This means the relationship of RM to increasing sharpness of chemograph peaks was inverted compared to RC. RM was not obviously related to discharge dynamics, neither to mean discharge (Fig. 6e), nor to the ratio of maximum to mean discharge (Fig. 6f). Disregarding events with very low discharge (events E6 and E10), it seemed possible that much of RM was the result of water imbalances during the events (Fig. 6g). However, RM of most chemographs plotted above the 1:1 line of RM and relative water balance, indicating that RM was higher than water imbalance would explain. RM showed a tendency to decrease with increasing KfOC (Fig. 6h), but no clear pattern was found for solubility (Fig. 6i) and soil half-life (Fig. 6j).

Regarding chemographs and calculation of RC, uncertainties arose from timing and frequency of sampling and analytical error and additionally from discharge measurement when calculating masses and RM. Analytical methods used in this study usually produced very consistent results so that variability in concentrations of parent compounds in duplicate samples was low (SD < 10 %). However, in individual samples collected at G1 analytical variability was elevated for met-ESA and met-OA (Fig. 3), reducing confidence in concentrations and the derived measures RC and RM of TPs in the affected chemographs (E2, E5, E7, E8, E9). Uncertainty related to timing and frequency of sampling can hardly be quantified but certainly depends on how well the sampling intervals captured variability in concentrations during flood events and how well the time lag between upstream and downstream sampling matched the residence time of solutes in the wetland. Lefrancq et al. (2017) assessed the effect of sampling frequency in pesticide monitoring data collected during runoff from a single vineyard and found that acute toxicity of pesticide flushes was underestimated by up to 4 times when calculated from event means and by up to 30 times when calculated from random samples. Although these data were collected on the plot scale, and we assume that variability in our catchment is lower due to longer flow paths and mixing processes on the catchment scale, uncertainty of the chemographs in our study could have been reduced by increasing sampling frequency. Regarding the timing of upstream and downstream sampling, there is evidence that water residence time in the wetland was in fact shorter than 1 h. The observation that for quickly responding compounds, such as boscalid, concentration in the first sample at G2 was often elevated compared to the first sample at G1 indicates that the contaminant flush had already reached G2 when sampling started. This did not influence determination of Cout,max and RC in the outlet of the wetland, as concentrations were still rising from the first to the second sample (Fig. 3). However, effects on Mout were higher, since a relevant fraction of contaminant mass leaving the wetland was not recorded and thereby caused overestimation of RM (Fig. S1). Another source of uncertainty exclusively affecting contaminant mass and not concentrations was the use of different gauging systems at G1 and G2. Different shapes of the measurement cross-section (triangular at G1 and rectangular at G2) caused G2 to be less precise and water imbalances on the event scale, particularly when flow was low. Summarizing the setup constraints above, we have high confidence that the experimental setup produced realistic chemograph shapes and captured peak concentration reasonably well but are less confident regarding contaminant loads.

Peak concentrations of mobilized contaminant flushes were different depending on the compound. This may be due to the application of different amounts and due to temporal patterns of application. The fact that maximum concentrations of metazachlor and flufenacet in specific events exceeded concentrations during most other events by a factor of more than 100 suggested application of these compounds shortly before the onset of runoff. Despite highly variable application patterns, our cluster analysis resulted in four groups with similar chemograph shape. Many factors have been shown in literature to influence the mobilization of pesticides in catchments, including catchment properties, event properties, and physiochemical compound properties. As catchment properties, here we consider factors associated with runoff generation such as catchment geometry, terrain slopes, and in particular the delineation of areas where different compounds were applied. The interplay of these factors defines hydrological activity and connectivity (i.e., by shortcuts like roads and drainage pipes) of critical source areas for different compounds (Doppler et al., 2012; Gomides Freitas et al., 2008). Event properties include intensity and dynamics of rainfall (Imfeld et al., 2020) and subsequent runoff (Doppler et al., 2014). Relevant physiochemical compound properties are, e.g., mobility and degradability (Gassmann et al., 2015).

These properties are reflected to varying degrees in the results of the cluster analysis. Cluster A was characterized by a quick response and a concentration plateau towards the end of sampling and was mainly composed of TPs. The fact that concentration maxima in cluster A were delayed compared to fungicides (cluster B), although their parent compounds were applied closer to the stream in the flat valley bottoms, suggests that they were transported with a slower flow component. Due to flatter terrain, surface runoff played a less important role, and the main transport pathway was subsurface flow. Where fields were undrained, however, transit time of water from the infiltration point to the stream would likely exceed the temporal scale of event sampling. Most of the water reaching the stream from the fields in the valley during discharge events would therefore be pre-event water, enriched in TPs formed in the soil, corresponding to the formation site of TPs of the chloracetamide herbicides to which metazachlor belongs (Mersie et al., 2004). Seepage of pre-event TP-rich water thus explains the immediate response of chemographs in cluster A. The quick response was often followed by a local concentration minimum between samples 2 and 5, i.e., between 30 min and 6 h after sampling was initialized. Coincidence of this minimum with concentration peaks of fungicides might suggest dilution of TP concentration by mixing with event water carrying high loads of fungicides but fewer TPs of metazachlor.

Cluster B represented differences between fungicides and the remaining compounds. Considering land use distribution in the studied catchment, it is unclear whether this partition reflects different compound properties or catchment properties or both. The fact that concentration in cluster B quickly increased with discharge (within 30 min) is in line with fast transport from the vineyard terraces to the stream via roads and drainage pipes as described by Gassmann et al. (2012) for suspended solids in the studied catchment. Along such preferential pathways, compound properties, such as sorption affinity, may be less important (Gomides Freitas et al., 2008) compared to, e.g., percolation through the soil with intense contact to sorption sites in the soil matrix. Moreover, fungicides are applied by sprayers into the foliage and can drift to, e.g., paved surfaces from which they can be quickly mobilized by subsequent rainfall (Lefrancq et al., 2013). We therefore hypothesize that cluster B was mainly produced by surface flushing and fast transport pathways of fungicides. This explained the quick rise and subsequent decline in concentrations (concurrent with plateaus produced by slower flow components in cluster A).

Cluster C was composed of chemographs of all compounds but mainly from events E2 and E4, indicating event dependence. Two aspects were found to support this idea. First, there was a secondary discharge peak in event 2 that did not contribute much in terms of contaminant concentration but rather caused dilution and produced particularly flat chemograph tails. Second, peaks of herbicides and TPs were less delayed compared to fungicides. This may be the result of recent herbicide application and active surface runoff in the flat valleys. Timing of pesticide application was identified as the main export driver of currently used pesticides by Imfeld et al. (2020), who performed a cluster analysis on rainfall data from a headwater vineyard catchment. Based on the magnitude of discharge and amount of mobilized contaminants (concentration of metazachlor ≈ 10 µg L-1), both explanations seem plausible in event E2. Event E4, however, did not show particularly high herbicide concentration, nor a secondary discharge peak. Although it is obvious that chemograph shapes in cluster C differed from the other clusters, unfortunately, the responsible factors remain unclear.

Cluster D included chemographs of both herbicides and TPs and presented a clear peak that was often defined by a single sample 2 h after the beginning of the event. In contrast to cluster A, cluster D was characterized by a single sharp discharge peak (except in event E7, where a second peak occurred shortly after the first) and mainly included chemographs during periods of low flow. Our interpretation is that cluster D represented flow events in which no dilution of herbicide and TP fluxes by fungicide fluxes or secondary discharge peaks occurred. Low pre-event discharge in cluster D compared to cluster A may indicate low water levels, which may have caused a slower response as no enriched pre-event water was released from the soils in the valleys.

The unclear interpretation of cluster C suggests that we missed important factors for the formation of chemographs. In fact, variables like spatial distribution of rainfall or pesticide application rates and timing (Imfeld et al., 2020) and possibly other factors likely influenced chemograph shapes. Knowing all these variables would not change the results produced by the clustering algorithm but rather increase our ability to interpret them. Nevertheless, the cluster analysis helped us to explore how the catchment and processes within it influenced concentration signals of mobilized contaminants. Particularly, the analysis helped us to understand under what conditions and for which pollutant sharp-peaked chemographs, associated with high acute toxicity, can be expected. We therefore see a high potential of this type of analysis for the identification of influential factors for contaminant mobilization in other catchments, although these factors may not be universal but catchment-dependent.