Target and suspect screening for biocides in drinking water resources of Honduras Open Access

In 2010, the United Nations (UN) achieved Millennium Development Goal (MDG) 7.c and declared that the world had halved the proportion of people living without sustainable access to safe drinking water. According to the official UN report, Honduras had achieved its target, with 87% of the population able to access safe water (WHO and UNICEF 2012). However, the MDGs defined ‘safe water’ as water from ‘an improved water source,’ with no stipulations about the quality, or the degree of treatment of said water. In fact, a study conducted in Honduras found that 88% of ‘improved water systems’ left the water untreated, and 70% delivered highly contaminated drinking water (Los desafios de los sistemas de agua potable rural (The challenges of rural drinking water systems) 2005). In 2015, with the release of the Sustainable Development Goals (SDGs), the UN redefined safe drinking water as ‘water free from fecal and priority chemical contamination’, with arsenic and fluoride listed as the priority chemicals (UNICEF and World Health Organization 2017). While an improvement, water classified as safe under this definition is not necessarily safe, as many risks remain unaccounted for and poorly understood. One argument for the looseness of the UN definition of safe water is that basic parameters can be indicators of contamination. However, work conducted in Honduras has demonstrated that standard water quality parameters, including fecal coliforms, are not reliable predictors of water safety (Cordona 2003).

The apparent overarching success of the MDGs and SDGs in improving access to safe water masks a large disparity between rural and urban populations. Rural communities are still underserved, with 20% of the rural population in Latin America lacking access to improved water sources, without accounting for water quality (Pearce-Oroz et al. 2011). This is especially concerning because rural communities face a unique set of threats to water quality, one of which being pollution derived from agricultural activity. Even as water systems develop in rural Central America, they may not be sufficient to provide safe drinking water in the context of these unique water quality threats because the UN definition of safe water is so general.

We hypothesize that biocide use, which includes pesticides, herbicides, and fungicides, may be one such threat to water quality in Honduras. As of 2018, nearly 30% of Honduran land was devoted to agriculture (US CIA 2018), with the most important crops by export value being coffee, banana, palm oil, beans, sugar, and melons (Secretaría de Agricultura y Ganadería et al. 2017). Although no recently published data are available, biocide application at the beginning of the 2000s was reported to be on average 13.3 kg/ha of cultivated land per year, a number that was projected to increase throughout Honduras, especially in the southern agricultural regions of the country (Kammerbauer & Moncada 1998; Jansen 2002). Over the last 20 years, the same crops continue to be produced at a high intensity, leading us to expect that biocide application remains high. During this time, the relevant classes of bioicides, ordered by total import weight, were dithiocarbamates, organophosphates, bipyridyliums, carbamates, triazines, and organochlorides. Of all imported bioicides, 16% by weight are classified by the World Health Organization (WHO) as toxicological class 1A (extremely hazardous) or 1B (highly hazardous) (Jansen 2002). The Honduran National Service of Agri-food Health and Quality (SENASA) maintains a list of 26 prohibited biocides (Productos prohibidos y restringidos en Honduras (Products prohibited and restricted in Honduras) 2021), issues permits for biocide imports, and maintains records and approves new biocides for use within the country. However, these national records are not publically available in the present and have been unreliable due to poor record keeping practices in the past (Balluz et al. 2001; Jansen 2002). The last available report was published in 2008, and indicated that biocides from all of the aforementioned chemical classes continued to be used, although there were no data published on quantity (Gabrie 2008). Furthermore, biocides on the market in Latin America are frequently diluted (Porfido et al. 2014; Fernández 2015), which often leads farmers to apply more than the recommended dose (Jansen 2002). All of this makes biocide use in Honduras very difficult to study. Overall, legislation has been ineffective in reducing biocide use in Honduras (Balluz et al. 2001; Jansen 2002).

Biocides can be transported from fields into surface water by runoff and can thus become a threat to water quality. Biocides have been found ubiquitously in surface waters adjacent to agricultural activities all around the world (Malaj et al. 2014; Bradley et al. 2017; Carpenter & Helbling 2018). In Honduras, 90% of the water consumed across all sectors comes from surface waters, with the remaining 10% coming from groundwater (Guillén & Tabora 2015). When consumed by humans, even in incredibly small quantities, biocides can have negative health effects such as endocrine system disruption, reproductive and immune system suppression, and birth defects in offspring (Colburn et al. 1993; McKinlay et al. 2008). The government of Honduras has imposed maximum contaminant levels (MCLs) on 31 biocides in drinking water (Norma Técnica Nacional para la Calidad del Agua Potable (National Technical Standard for Drinking Water Quality) 1995). Honduran law states that biocide testing must be conducted with liquid or gas chromatography-based techniques; however the law acknowledges that this is outside the capacity of the vast majority of Honduran drinking water treatment plants (DWTPs), and thus, there are no monitoring requirements (Reglamento de calidad del servicio (Service quality regulation) 2005).

Due to the prevalence of agricultural activity and biocide use in Honduras, combined with the effective lack of monitoring, we hypothesize that bioicide contamination could pose a serious, yet unaccounted for, threat to drinking water safety. In the few studies that have screened for biocides in water systems in Honduras, biocide concentrations exceeded MCLs (Estrada 1995; Kammerbauer & Moncada 1998; Cordona 2003). Biocides have also been detected in the water resources of other South and Central American countries (Britto et al. 2015; Hunt et al. 2016). However, these studies were limited by outdated technology and were only able to measure a select few biocides. The objectives of this study were to: (i) identify watersheds-of-interest in Honduras by examining national agricultural practices in combination with hydrology and drinking water infrastructure using a Geographic Information System (GIS); (ii) collect water samples from watersheds-of-interest during the dry season and the rainy season, and along the treatment train of associated DWTPs; and (iii) quantify the concentration of 55 target biocides and estimate the abundance of four priority suspect biocides and other high-abundance biocides in each of the water samples by means of high-resolution mass spectrometry (HRMS). By employing HRMS, we were able to study a wide range of biocides and examine geospatial and temporal trends, as well as evaluate the efficiency of existing DWTPs in removing biocides. As Honduras strives to meet SDG 6.1 and achieve universal safe drinking water access, this work highlights the importance of accounting for biocide contamination as a threat to water safety and elucidates patterns in biocide contamination, which may be applicable to surface water resources throughout the majority world.

Our study focuses on water resources in southern Honduras, where agricultural activity is concentrated. We collected water samples from seven watersheds and/or associated DWTPs in this region. A map of Honduras including the location of the sampling sites is provided in Figure 1. All DWTPs were designed as part of the AguaClara program at Cornell (AguaClara Cornell; González Rivas et al. 2014) and are operated and funded by the communities in which they operate. All DWTPs are gravity powered and include rapid mix coagulation with polyaluminum chloride, flocculation, sedimentation, and granular media filtration followed by a chlorination phase. During our sampling campaigns, we measured the average influent pH of the DWTPs to be 7.76 (standard deviation (SD) 0.32); the average conductivity 111.12 (SD 43.64); the average total dissolved solids 90.42 (SD 52.41); and salt concentration consistently 0. During one sampling event, when samples were notably turbid, turbidity was recorded at 720 NTU. For all other sampling events, turbidity was not measured but was visually assessed to be low. Further information about the DWTPs is provided in Tables S1a and S1b of the Supplementary Material. The San Matias (SM) and Moroceli (MO) sites were sampled repeatedly. Five DWTPs at Marcala (Mar), Cuatro Comunidades (CC), San Nicolas (SN), La Brea (LB), and Tamara (Tam) were sampled only once.

We performed geospatial analyses to quantify the extent of agricultural activity in the watersheds of each treatment plant using QGIS version 3.16.1-Hannover. Digital elevation models (DEMs) were obtained from USGS EarthExplorer (U.S.G.S. 2021). Areas of low elevation (i.e., lakes) were filled using the Fill Sinks tool in the SAGA Terrain Analysis – Hydrology toolbox (Wang & Liu 2006). Surface water flow was modeled with the Catchment Area tool (also from SAGA Terrain Analysis – Hydrology), and locations of DWTPs were snapped to the point of highest flow accumulation within a 1 km radius from the DWTP location using the Snap Points to Grid tool in the SAGA Vector point tools toolbox. This new point was assumed to be the source of water for the DWTP. Based on correspondence with local treatment plant operators and field surveying, 1 km was assumed to be a reasonable distance. Finally, the Upslope Area tool within the Saga Terrain Analysis – Hydrology toolbox was used to delineate watersheds based on the snapped points and filled DEM. Shapefiles for each of the delineated watersheds were imported into RStudio version 1.3.1073 (running R version 3.3.3) using the st_read function from the sf library (Pebesma et al. 2021). Land-use data were downloaded from EarthStat's Harvested Area and Yield for 175 Crops year 2000 dataset (Monfreda et al. 2008) and imported into R using the raster tool from the raster library (Hijmans et al. 2021). The raster corresponding to each unique crop, as well as a total agricultural area raster, was clipped to the area of each individual watershed, and the fraction of land use within each watershed pertaining to each crop was calculated.

Two sampling campaigns were conducted with the goals of capturing: (i) the temporal variability of bioicide occurrence; (ii) the geographic variability of bioicide occurrence; and (iii) the potential of the local DWTPs to remove bioicides from raw source water. The first sampling campaign was conducted from October 2018 to January 2019 during the dry season in Honduras. We collected 40 mL grab samples from two DWTPs (MO and SM) during five sampling events over the period of the first sampling campaign. Duplicate samples were collected simultaneously from the influent to each DWTP (referred to as influent samples in the following), from the effluent of the sedimentation basin (prior to any chlorination, referred to as pre-chlorination samples in the following), and from the post-chlorination effluent (referred to as effluent samples in the following). Additional samples were collected from the source water (referred to as source samples in the following) for both DWTPs. We also collected influent samples from five additional DWTPs (Mar, CC, SN, LB, and Tam) to further capture geographic variability. The second sampling campaign was conducted in summer 2019, which represents the rainy season in Honduras. We collected duplicate 40 mL grab samples from the influent, pre-chlorination, and effluent of the MO and SM DWTPs during one sampling event (see Figure 1 for a map of all sampling locations and Table S2 for full sample data sheets). We note that samples collected in October 2018 were collected in 40 mL amber glass vials. However, due to vial breakage during storage, all subsequent samples were collected in 40 mL LDPE vials. The effluent samples were spiked with 5 mL of 23 mg L⁻¹ sodium metabisulfite to stabilize residual chlorine. All samples were stored at Zamorano University at −20 °C until transport to Cornell University (holding time was less than 3 months). During transport, samples were wrapped in aluminum foil and packed into a Styrofoam box filled with freezer packs. Samples arrived at Cornell University in a partially to completely frozen state. We note that samples were neither filtered nor sterilized prior to storage and shipment. Therefore, we cannot rule out biocide adsorption onto suspended particles or microbiological activity affecting biocide concentrations during this period, and all reported biocide concentrations consequently represent conservative estimates of actual biocide concentrations in the native samples.

We selected 55 biocides from our in-house library of reference standards to include in this study. Biocides were selected based on their global high-volume usage and whether they have MCLs in Honduras, are classified as toxicological class 1A or 1B, or are prohibited biocides in Honduras. We also included other water-relevant compounds as positive controls (e.g., caffeine, abscisic acid) as they have been ubiquitously found in surface water around the world (Moschet et al. 2013; Wu et al. 2014; Carpenter & Helbling 2018; Carpenter et al. 2019b; Peteffi et al. 2019). Stock solutions of all biocide reference standards (Sigma-Aldrich) were prepared at 1 g L⁻¹ in LC-MS-grade methanol (OmniSolv, VWR), nanopure water (EMD Millipore), LC-MS-grade acetonitrile (Fisher Chemical), ethanol (Decon Labs), or dimethyl sulfoxide (Macron Fine Chemicals) and stored at −20 °C. A mixture of all reference standards was created in nanopure water at 5 mg L⁻¹ and stored at −20 °C. The mixture was diluted with nanopure water to create a calibration curve at 0, 1, 5, 10, 25, 50, 100, 250, 500, 1,000, 1,500, and 2,000 ng/L. Similarly, a mixture of isotope-labeled internal standards (ILISs) was prepared in nanopure water at 5 mg L⁻¹ and stored at −20 °C. We spiked 20 μL of this ILIS mixture, diluted to 0.2 mg L⁻¹, into each calibration standard and sample prior to sample analysis; ILISs were added to account for differences in analyte ionization efficiency resulting from differences in sample matrix or the addition of sodium metabisulfite. Reference standard and ILIS information is provided in Supplementary Tables S3 and S4.

Upon arrival at Cornell University, all water samples were completely thawed at room temperature and transferred from the 40 mL LDPE vials into 40 mL amber glass vials. To extract any bioicides that may have adsorbed to the walls of the LDPE vials, 2 mL of methanol was added to the emptied LDPE vial and vortexed. The methanol was then transferred to an empty 40 mL amber glass vial and completely evaporated under a gentle stream of high-purity nitrogen gas. The sample water was then transferred into the 40 mL amber glass vial containing the residual extract. The vials were capped and vortexed. Samples were then transferred to 50 mL conical tubes (VWR) and centrifuged at 4,700 rpm at 20 °C for 15 minutes (Sorvall Legend XTR, Thermo Scientific). The supernatant was transferred into 20 mL glass beakers and spiked with 28 μL of 1 M ammonium acetate buffer. The sample pH was adjusted to between 6.3 and 6.7 with formic acid (5%) and ammonia (1.4 N). Then, 8 mL of the sample was transferred into each of three 10 mL glass LC-MS sample vials and spiked with 20 μL of the ILIS mixture. Prepared samples were stored at 4 °C until analysis within 20 days.

All prepared samples were measured once, and the resulting data were used for both target screening and suspect screening as described in the following. We adapted a previously described analytical method that implements large volume injection (LVI) and high-performance liquid chromatography (HPLC) coupled with HRMS (QExactive hybrid quadrupole orbitrap, Thermo Fisher Scientific) (Carpenter et al. 2019b). Briefly, samples were injected at 5 mL volumes onto a Hypersil GOLD aQ trap column (2.1×20 mm, particle size 12 μm, Thermo Fisher Scientific) at room temperature (21–22 °C) and eluted with a mobile phase gradient onto an XBridge C-18 analytical column (2.1×50 mm, particle size 3.5 μm, Waters) at 25 °C for analyte separation. Each sample was analyzed in a positive polarity mode with heated electrospray ionization with the following parameters: voltage: +4.0 kV; sheath gas flow: 40 arbitrary units (AU); auxiliary gas flow: 20 AU; capillary temperature: 320 °C; S-lens RF level: 50 AU, auxiliary gas heater temperature: 50 °C. The instrument method acquired full-scan MS data with the following parameters: a mass-to-charge ratio (m/z) range of 100–1,000; resolution: 140,000 at 200 m/z; automatic gain control (AGC) target: 500,000; maximum injection time: 200 ms. Data-dependent MS2 scans were acquired with an inclusion list consisting of the target biocides and the following parameters: resolution: 17,500 at 200 m/z; AGC target 200,000; maximum injection time: 100 ms; isolation window: 1 m/z; underfill ratio: 0.1%; dynamic exclusion: 6 s. Additional details on the HPLC gradient program can be found in Supplementary Table S5.

Data collection and processing were conducted using XCalibur v4.1 (Thermo Fisher Scientific). The target bioicides were quantified using ILISs based on the ratio of the area responses of the target biocide to the internal standard and by 1/x weighted linear least-squares regression. Limits of quantification (LOQs) were determined by the lowest linear calibration point with five sticks and the presence of a diagnostic fragment. Method blanks, instrument blanks, and continuing calibration checks were included in the analyses to account for laboratory sources of contamination, sample carryover, and to verify the precision and accuracy of the calibration.

Our laboratory maintains a suspect list containing over 2,000 contaminants of interest. These substances have all been measured in our laboratory using the analytical method described in the preceding, but we do not have reference standards for the majority of these suspect chemicals. Therefore, we can confidently screen for the presence or absence of these chemicals in any water sample and estimate their abundance, but we cannot perform a sensitive quantification. We first cross-referenced our complete suspect list with 61 biocides that have MCLs in Honduras, are classified as toxicological class 1A or 1B, or are prohibited biocides in Honduras. We identified 15 suspect biocides that overlapped with one of these site-relevant lists. Eleven of those were already included in this study as target biocides; the remaining four were identified as priority suspect biocides. We then performed a suspect screening for the priority suspect biocides as previously described (Pochodylo & Helbling 2017; Wang et al. 2020). Further details on suspect screening can be found in text in the Supplementary Material and in Table S6.

All statistical analyses were performed in RStudio version 1.3.1073 (running R version 3.3.3). Based on all >LOQ detections from the target screening, for each individual sample, the average concentration and complexity were calculated using the mean and length functions, respectively, from the base package of R. For each plant, the mean concentration and complexity were calculated based on influent and source samples only for both the dry and rainy seasons. For land-use correlation analyses, we included only the December sampling event, in which all plants were sampled, to eliminate the effects of temporal variation. We tested for non-parametric correlations between complexity or concentration and land use (calculated as described above) by calculating the Spearman's rank correlation coefficient and p-value.

We used QGIS to delineate watersheds around the source water for each of the seven DWTPs. The proximity and relative sizes of the seven watersheds are provided in Supplementary Figures S1–S3. The size of the watersheds for the seven DWTPs ranges from 196 ha (SN) to 3,115 ha (CC). The mean watershed size is 1,102 ha. On average, cropland represents 22% of land use within these watersheds, although this value ranges from 5.7% in the CC watershed to 50.2% in the LB watershed. Across all watersheds, the dominant crops with regard to land use were maize, coffee, beans, sorghum, sugar cane, oil palm, and bananas, although the production of no single crop dominates more than 10% of land use in any given watershed. With this analysis, we confirmed that there is agricultural activity in all of the watersheds, and we identified that the most significant crops by percent land use are maize, coffee, and beans. The total watershed area, the area devoted to agriculture, and the area devoted to the production of all crops representing >0.5% of land use in any watershed are provided in Table 1.

We report biocide occurrence data for 46 samples collected from the seven watersheds and/or associated DWTPs. For MO, we report biocide occurrence data for three source samples and six influent, pre-chlorination, and effluent samples. For SM, we report biocide occurrence data for three source samples and six influent and pre-chlorination samples, and five effluent samples. For Mar, CC, SN, Tam and LB, we report biocide occurrence data for one influent sample. All biocide concentrations are reported in Supplementary Tables S7–S11.

We detected 30 of the 55 target biocides in at least one sample and 24 of the 55 target biocides in at least one of the source or influent samples; a summary of these latter data is provided in Table 2. Across all DWTPs, the most consistently detected compounds in influent samples were caffeine (100% of samples), metamitron (70%), abscisic acid (50%), diuron (45%), and aminobenzimidazole (30%). The mean concentration for all >LOQ detections was 96.0 ng/L in the dry season and 10.0 ng/L in the rainy season, with a maximum concentration of 9.0 μg/L for caffeine (non-biocide) in the influent of the SM DWTP. The influent and source samples from MO contained the highest concentrations of biocides. The maximum concentrations of biocides measured in the whole study were 709 ng/L for 2-methylisothiazolin-3-one (MI) in a MO influent sample, 649 ng/L for alachlor in a MO source sample, 195 ng/L for benzisothiazolin in a MO source sample, and 126 ng/L for metamitron in a MO influent sample. These values are toward the lower end of the range of concentrations typically observed in surface water studies in the United States (Bradley et al. 2017; Pochodylo & Helbling 2017; Carpenter & Helbling 2018; Carpenter et al. 2019b). To our knowledge, the only study measuring biocide concentrations in southern Honduras was carried out by Kammerbauer and Moncada between 1995 and 1997. This earlier study reported concentrations of DDT ranging from 0 to 100 μg/L, chlordane ranging from 20 to 250 μg/L, dieldrin ranging from 5 to 35 μg/L, and aldrin ranging from 0 to 10 μg/L (Kammerbauer & Moncada 1998), which is nearly three orders of magnitude higher than the concentrations we measured, although it is important to note that both the analytical methods and the compounds of interest were different between the two studies, which limits direct comparison. More recently, triazine fungicides have been measured in Brazilian river water at concentrations between 80 and 480 ng/L (Britto et al. 2015), which are comparable values with what we measured. Additionally, insecticide concentrations in stream sediments have been reported in streams across Argentina, Paraguay, and Brazil (Hunt et al. 2016). Based on the EPISuite Level III fugacity model from the U.S. EPA (EPA 2015), their results correspond to aqueous concentrations ranging from 20 ng/L to 9.1 μg/L, which also agrees well with our findings. Of all the compounds on our target list, eight have MCLs under Honduran law, but we did not measure any contamination in excess of these MCLs. A summary of all of the quantification data for the MO and SM DWTPs are provided in Supplementary Tables S12 and S13.

Of the four priority suspect biocides (bentazone, endrin, pyridat, and propanil), only propanil was measured in the source sample and a pre-chlorination and effluent sample from MO in November 2018. We do not have an authentic standard for propanil to allow for sensitive quantification, but we can compare the peak areas of the suspect biocides with those of the target biocides and make rough estimates of abundances (Backe et al. 2013; Jacob et al. 2021). The base 10 logarithm of the propanil peak area was in the 6.3–6.5 range, and we estimate that propanil is present in the MO samples in the range of 10 ng/L. We also identified five high-abundance biocides in at least one source or influent sample including 4-5-dichloro-2-n-octyl-isothiazol-3 (DCOIT, fungicide), N,N-diethyl-3-methylbenzamid (DEET, insect repellant), diazoxon (metabolite of diazinon), icaridin (insect repellant), and iodocarb (herbicide). These substances were detected in a variety of samples as detailed in Supplementary Tables S14–S18 and with peak areas that suggest that they are the most abundant substances in the water samples. We conservatively estimate that DCOIT and diazoxon are present at concentrations of >10 ng/L, icaridin and iodocarb are present at concentrations of >100 ng/L, and DEET is present at concentrations of >1,000 ng/L. Although these biocides do not have MCLs in Honduras, are not classified as toxicological class 1A or 1B, and are not prohibited biocides in Honduras, their occurrence at relatively high levels in some water samples suggests that more targeted investigations are warranted.

We expected to observe a correlation between the extent of agriculture in a watershed and biocide occurrence, as defined by both complexity and concentration. Complexity refers to the number of biocides present at >LOQ in a given sample, and concentration refers to the mean concentration of all >LOQ detections in that sample. We calculated the concentration and complexity of each sample, as well as percent agriculture in the watershed from which each sample was taken. We then tested for non-parametric correlations by calculating the Spearman's rank correlation coefficient and p-value. No such correlation was observed between the extent of agriculture and either biocide complexity or concentration. However, bean production (% of land use) was significantly and strongly correlated with biocide complexity (Spearman's ρ=0.96, p=0.005). Bean production was also correlated with the concentration of aminobenzimidazole (Spearman's ρ=0.80 p=0.03) and abscisic acid (Spearman's ρ=0.80 p=0.03), which are, respectively, a transformation product of the widely used fungicide carbendazim and a naturally occurring plant hormone (plots of correlation data are provided in Supplementary Figure S4). In most of the watersheds, beans are the third most important crop (behind maize and coffee, see Table 1). Nationally, beans are the second most important crop in Honduras (Tshering 2002). The fact that beans are a predictor of contamination, but maize and coffee are not, could suggest that biocides are less important for maize and coffee production in Honduras, or that we happened to include biocides relevant to bean production in our target list, but did not include biocides relevant to maize and coffee production.

The lack of correlation between the overall extent of agricultural within a watershed and biocide occurrence could suggest that land use is not a good predictor of biocide occurrence, due to inconsistent biocide use. Indeed, previous work in southern Honduras has found that biocide concentration and complexity were associated not with the extent of agriculture, but with specific agricultural practices such as the relative intensity of export-driven agriculture (Kammerbauer & Moncada 1998) or the absence (or limited width) of riparian buffers (Hunt et al. 2016). Nevertheless, the correlations observed with bean production suggest that land use can be a relevant driver of contamination, which suggests that the lack of correlation could also be a function of the limitations of a target screening approach to capture the full picture of contamination in water samples, about which little is known. We also expect that a sampling campaign that relied on more continuous sampling as opposed to grab sampling might provide more insights on the overall relationship between land use and biocide occurrence.

We observed a significant but weak negative correlation between the octanol–water partition coefficient (logK_ow) of each biocide and its measured concentration (Spearman's ρ=−0.12, p=2.22×10⁻¹², Supplementary Figure S5). This illustrates that the affinity of a biocide for water is correlated with the overall concentration of that compound in surface water, although there are other important factors determining the contamination profile of our water samples, such as the quantity of bioicides applied (Wang et al. 2018). Additionally, peak biocide application in Latin American agricultural areas typically occurs during the rainy season (Hunt et al. 2016). Bearing these facts in mind, we expect that during the dry season, biocides with lower logK_ow values (higher affinity for water) will partition from the soil to the water phase, but that during the rainy season, due to increased biocide application and increased overall contact with water (i.e., higher runoff volumes), biocides with higher logK_ow values (lower affinity for water) may also enter the water phase. Therefore, we expect to observe higher contamination complexity in our rainy season samples. Indeed, in the MO and SM samples, we did detect more than double the number of compounds in the rainy season samples versus the dry season samples. Additionally, biocides unique to the rainy season samples were detected in the SM and MO samples. In the MO samples, isoproturon, propoxur, and atrazine-2-hydroxy were detected in the rainy season sample but never during the dry season. In the SM samples, propoxur, trinexapac-ethyl, and dehydroacetic acid were detected in the rainy season sample, but never during the dry season. Across all samples, propoxur and isoproturon were only detected in the rainy season.

Conversely, the mean concentration of biocides is almost an order of magnitude higher in our dry season samples than in our rainy samples. Although we observe a large range of concentrations in the dry season samples, the rainy season concentrations still fall below the lower end of this range. The lower concentrations of bioicides in the rainy season are likely the result of dilution due to increased overall flow (Bahlmann et al. 2012; Prescott et al. 2017; Carpenter et al. 2019a). Although we do not have streamflow data for the source waters of our DWTPs, weather reports from the nearby Toncontín International Airport indicate daily precipitation for all of June and July (Datos históricos meteorológicos en el verano de 2019 en el Toncontín International Airport (Historical meteorological data from summer 2019 in the Tonctontín International Airport) 2019), so we can safely assume that surface water flow was elevated during this time period. Intense precipitation since June suggests that the ‘first flush’ (Louchart et al. 2001; Olsson et al. 2013) of biocides likely occurred well before sample collection in July. We expect that the complexity of samples collected at the beginning of the rainy season would have been even higher.

Differences in biocide contamination profiles between sampling campaigns could also be related to varying agricultural practices at different times of the year. In southern Honduras, there are two major sewing periods – the ‘Spring Period’, which occurs between mid-May and mid-June, and the ‘Posterior Period’, which occurs at the end of the rainy season (September through October). Biocide application is typically recommended for several weeks after seeds are sewn, or as problems occur (Escoto 2015). This means that our rainy season samples line up with the Spring Period of biocide application, and our dry season samples correspond to the Posterior Period of biocide application. Therefore, the presence of specific biocides may be linked to particular crops that are produced only during a certain season. For example, 70–80% of bean production occurs during the Posterior Period (Escoto 2015), so it is not surprising that we found a correlation between bean production and biocide contamination in data from samples collected in December.

We conclude that biocide contamination profiles differ between the dry and rainy seasons. Contamination complexity is higher during the rainy season than during the dry season. Across all watersheds, we observe lower contaminant concentrations in the rainy season than in the dry season likely due to dilution.

Existing drinking water treatment processes in the DWTPs we studied show no consistent impact on biocide concentration (Supplementary Figure S6). We calculated percent removal based on the effluent and the pre-chlorination concentrations, for biocides in every complete sampling event (samples taken along the treatment train), in which the given biocide was measured at >LOQ in the pre-chlorination sample. We calculated percent removal based on pre-chlorination samples rather than influent samples, because influent samples are highly susceptible to temporal variation in contaminant concentration (Bahlmann et al. 2012; Carpenter et al. 2019b), and drinking water treatment processes prior to chlorination have been shown to have little effect on biocide concentrations (Benner et al. 2013).

We expect biocides with similar chemical structures to behave similarly when treated with chlorine, so we explored biocide removal by chemical class. Those biocides classified as organochlorides, and those classified as other, show no consistent trends in removal by chlorination. Carbamates display an average increase in concentration during chlorination, and triazines seem to experience some extent of removal. All biocide classes have very high variability in percent removal, which could have several causes. Because our sampling campaign was not designed to explicitly evaluate DWTP performance, samples from all three treatment phases were taken at once, rather than a hydraulic residence time apart, so samples do not actually represent the same parcel of water moving through treatment. Thus, results could be skewed by temporal variation in the biocide occurrence of influent water. Furthermore, 60% of all biocides detected were within 10% of their LOQ. At such low concentrations, false negatives, or quantification inaccuracies are possible, which could further skew results. However, previous studies that have investigated the impact of chlorination and conventional drinking water treatment processes for removing biocides from water have also generally reported variable and poor removal (Benner et al. 2013). In agreement with the existing literature, our findings demonstrate that existing DWTPs are not consistently effective at removing biocides, although they may partially reduce triazine concentrations.

The data reported here represent the first extensive study on biocide occurrence in drinking water resources of Honduras. We demonstrated that biocides are present in drinking water in watersheds with agricultural activity. Biocide occurrence varies throughout the year, with lower average concentrations but higher biocide mixture complexity during the rainy season versus the dry season. Bean production is significantly correlated with biocide complexity, but we did not measure a correlation between overall agriculture and biocide occurrence. Based on our findings, DWTPs do not have any consistent impact on biocide concentrations. Although we did not measure biocides at levels that exceed current MCLs, our results illustrate that biocide occurrence is relevant and should not be ignored in Honduras. We expect that our findings can be extrapolated to other countries in Central America, and perhaps more broadly to other regions in the majority world where climate, agricultural practices, and environmental policies are comparable. Our study also demonstrates the limitations of using target screening to study biocides when little is known about the target biocides that may be present in the water samples. Future studies utilizing non-target screening approaches would enable a more comprehensive, and less biased, understanding of biocide contamination in Honduras (Carpenter et al. 2019b). Additionally, the use of passive sampling or automatic sampling techniques is recommended to avoid the bias that may be introduced by depending on individual grab samples from a temporally variable system. Finally, our data can be used to identify biocides that could be used as predictors of biocide occurrence in water resources in southern Honduras. For example, aminobenzimidazole was present in watersheds with relatively high bean production and absent in those without; this substance may be a suitable indicator of biocide occurrence in water resources more broadly.

Supporting information is available that includes additional details on study area, methods, and results.

This work was supported by the United States National Science Foundation (CBET-1748982) and Cornell University Engineering Learning Initiatives.

All relevant data are included in the paper or its Supplementary Information.