JEQ Journal of Natural Resources and Life Sciences Education
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online 7 June 2005
Published in J Environ Qual 34:1174-1185 (2005)
DOI: 10.2134/jeq2004.0044
© 2005 American Society of Agronomy, Crop Science Society of America, and Soil Science Society of America
677 S. Segoe Rd., Madison, WI 53711 USA
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Related articles in JEQ
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via ISI Web of Science (10)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Lindahl, A. M. L.
Right arrow Articles by Jarvis, N. J.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Lindahl, A. M. L.
Right arrow Articles by Jarvis, N. J.
Agricola
Right arrow Articles by Lindahl, A. M. L.
Right arrow Articles by Jarvis, N. J.
Related Collections
Right arrow Vadose Zone Processes and Chemical Transport
Right arrow Surface Water Quality
Right arrow Watershed-Scale Studies
Right arrow Pesticides
Right arrow Dual Porosity/Permeability Models

TECHNICAL REPORTS

Landscape and Watershed Processes

Stochastic Modeling of Diffuse Pesticide Losses from a Small Agricultural Catchment

Anna M. L. Lindahla,*, Jenny Kreugera, John Stenströmb, Annemieke I. Gärdenäsa, Ghasem Alavic, Stéphanie Roulierd and Nicholas J. Jarvisa

a Department of Soil Sciences, SLU, Box 7014, 750 07 Uppsala, Sweden
b Department of Microbiology, SLU, Box 7025, 750 07 Uppsala, Sweden
c Sida, Dusjanbe (DHL), 105 25 Stockholm, Sweden
d Institute of Terrestrial Ecology, ETH Zürich, CH-8952 Schlieren/Zürich, Switzerland

* Corresponding author (anna.lindahl{at}mv.slu.se)

Received for publication February 4, 2004.

   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
The objective of this study was to identify the main sources of variation in pesticide losses at field and catchment scales using the dual permeability model MACRO. Stochastic simulations of the leaching of the herbicide MCPA (4-chloro-2-methylphenoxyacetic acid) were compared with seven years of measured concentrations in a stream draining a small agricultural catchment and one year of measured concentrations at the outlet of a field located within the catchment. MACRO was parameterized from measured probability distributions accounting for spatial variability of soil properties and local pedotransfer functions derived from information gathered in field- and catchment-scale soil surveys. At the field scale, a single deterministic simulation using the means of the input distributions was also performed. The deterministic run failed to reproduce the summer outflows when most leaching occurred, and greatly underestimated pesticide leaching. In contrast, the stochastic simulations successfully predicted the hydrologic response of the field and catchment and there was a good resemblance between the simulations and measured MCPA concentrations at the field outlet. At the catchment scale, the stochastic approach underestimated the concentrations of MCPA in the stream, probably mostly due to point sources, but perhaps also because the distributions used for the input variables did not accurately reflect conditions in the catchment. Sensitivity analyses showed that the most important factors affecting MACRO modeled diffuse MCPA losses from this catchment were soil properties controlling macropore flow, precipitation following application, and organic carbon content.

Abbreviations: LAI, leaf area index • LOD, limit of determination • PRCC, partial rank correlation coefficients


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
IN RECENT DECADES, many carefully designed and performed experimental studies, coupled to theoretical approaches, have led to an improved understanding of the complex interaction of processes that govern pesticide fate and mobility at the scale of small soil columns or field plots (Boesten, 2000; Jarvis, 2001). However, the smallest scale of interest for managing and minimizing undesirable impacts on the environment is the scale at which farmers apply pesticides (i.e., the field scale). Moreover, an understanding of the factors controlling losses of pesticides at even larger scales (i.e., catchments) is becoming increasingly paramount, for example with the advent of the EU "Water Framework Directive" (European Union, 2000). It is only when we have a sound understanding of the factors responsible for pesticide contamination of surface waters at larger scales that effective mitigation measures can be designed.

Despite the need for progress in this field, there are still many unresolved difficulties in extrapolating results and process understanding gained from small-scale experiments to the field and catchment scales that are most relevant for management. Monte Carlo approaches have been applied to the "upscaling" problem and have clearly demonstrated the potentially large effects of soil heterogeneity on leaching at field (Wu and Workman, 1999) and catchment scales (e.g., Soutter and Pannatier, 1996). However, the results of such analyses are conditional on the assumptions underlying the process descriptions in the selected model and on how the model is parameterized. In particular, a lack of information on parameter distributions, and especially parameter correlations, has severely limited progress. For example, many studies have treated the sorption coefficient and degradation rate coefficient as independent parameter distributions (Di and Aylmore, 1997). This may lead to overestimation of variability in leaching, since sorption and degradation are sometimes found to be inversely correlated (Allen and Walker, 1987; Cantwell et al., 1989; Guo et al., 1999), as sorbed compounds are less available to soil microorganisms. Other studies have focused on variability in transport characteristics, but have ignored variability in sorption and degradation, presumably due to lack of data (Wu and Workman, 1999). Even when sufficient input data exist to properly parameterize stochastic approaches, no one has (to our knowledge) actually compared predictions with measured pesticide export from catchments.

There is still some uncertainty regarding the relative significance of variability in soil organic carbon content (which largely controls sorption and to some degree degradation processes), and variation in solute transport properties and parameters, especially in the presence of macropore flow. Sensitivity analyses for "chromatographic" flow and transport models show that leaching is highly sensitive to parameters describing sorption and degradation (Boesten, 1991), while corresponding analyses for models that incorporate a treatment of non-equilibrium macropore flow suggest that parameters related to the macropore region show a similar degree of sensitivity (Dubus and Brown, 2002). Variation in pore water velocity rather than organic carbon content was found to explain most of the variability in leaching risk of a weakly adsorbed pesticide at a study site with a shallow water table (Mulla et al., 1996). Similarly, in a stochastic modeling approach, atrazine transport predictions were found to be relatively insensitive to variations in organic carbon content compared with the variability of hydraulic conductivity and soil physical properties (Lafrance and Banton, 1995).

The objectives of this study were to use the process-based macropore flow model MACRO (Jarvis, 1994) to identify the important controls on pesticide leaching at both field and catchment scales. Monte Carlo simulations were run for the Vemmenhög catchment in southern Sweden based on cultivation and pesticide usage data derived from farmer interviews and stochastic soil parameter sets representing the variability of soil characteristics. Simulated pesticide losses were compared with measured concentrations in drainflow from a single field for one year (1999) and in the stream at the outlet of the catchment for seven years (1993–1999). Sensitivity analyses were performed to identify the main sources of variation in pesticide losses.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Catchment Description
The Vemmenhög catchment, located in the very south of Sweden (55°26' N, 13°27' E), is at an altitude of about 45 m above sea level, and has undulating topography with glacial till–derived calcareous sandy loam and loam soils. The total catchment area of 9 km2 (900 ha) consists of 95% arable land, with four major crop types (winter cereals, spring cereals, winter rape, and sugar beets) comprising 95% of the cropped area (Table 1). The climate in the region is maritime with a mean annual temperature of 7.2°C, and mean summer and winter temperatures of 16 and –1°C, respectively. The length of the growing season (mean daily temperature of >5°C) is around 220 d. The long-term (1961–1990) average annual precipitation is 662 mm and falls mainly as rain.


View this table:
[in this window]
[in a new window]
  		Table 1. Cropping and MCPA applications in Vemmenhög, 1993–1999.

 
Extensive drainage systems have been installed within the area and open ditches were covered and replaced by a culvert system in a larger part of the catchment during the late 1950s. About 40% of the field area is systematically drained with tiles spaced at 16 m and at a depth of approximately 1 m. On the remaining area, tile drains are installed in an irregular manner, following the natural drainage routes and connecting isolated depressions to the culvert system. The culvert collects tile drainage and also runoff water from surface runoff inlets located in topographic depressions in the landscape. Surface runoff inlets can also be found along roads and in some farmyards. The culvert water is discharged into a small open stream that flows another 1.1 km to the outlet of the catchment.

Data Collection for Management Practices
Information on the crops grown, fertilization, and pesticide handling and usage on the field scale (i.e., type of pesticide, dose, time of spraying, and field location and size) have been collected annually through personal interviews with the farmers (Kreuger, 1998a). Before 1997, all but one farmer (2.5% of the arable land) cooperated with this investigation. Due to joint cultivation with a neighboring farm, the investigation now also incorporates information from this farm (except for one field of 4.5 ha) since 1997 (Kreuger, 1998b). In this study, we focus on MCPA (4-chloro-2-methylphenoxyacetic acid), one of the more widely used herbicides in the catchment. It is applied postemergence mainly to spring-sown cereals, but 12 spring applications in the period 1993 to 1999 were also made to autumn-sown cereal crops and one on peas. Table 1 shows the amounts of MCPA applied in the catchment, the number of applications, and the total sprayed area each year between 1993 and 1999. In total, 85 fields were sprayed at least once during this period, covering a total area of 598 ha (equivalent to 72% of the catchment area). Treated fields were sprayed with MCPA one to four times during the 7-yr period.

Validation Data
Monitoring of water discharges and pesticide concentrations at Vemmenhög started in 1990 and is still ongoing (Kreuger, 1998a). In this study, we focused on the measurements made in the period 1993 to 1999. Culvert flow rates are measured using a 90-degree V-notch weir and an ultrasonic sensor (Model 3210 flow meter; ISCO, Lincoln, NE). Time-weighted water samples (each sample being a composite of subsamples taken at 10-min or hourly intervals) were collected once or twice a week from the stream using programmable automatic samplers (ISCO) during January–June 1993, May–October 1994, and May–November 1995–1999. More details of the sampling procedures can be found in Kreuger (1998a).

In addition to monitoring outflows and pesticide export from the catchment, a representative field, 30 ha in size, located within the catchment at Näsbygård, was used for model calibration and testing at the field scale. At the Näsbygård field, measurements of piezometric pressure (1992–1999, with 8–13 measurements each year) were available at two depths for each of three locations, one in a hollow, another at a local watershed or hilltop position, and one on a midslope position. The piezometers located at the hilltop and midslope positions mostly indicated downward-directed hydraulic gradients (i.e., local recharge). On average, long-term (1993–1999) calculations with the MACRO model (Jarvis, 1994) suggest a yearly ground water recharge of approximately 100 mm and a discharge of approximately 260 mm to the field drainage systems. Piezometers located in the hollow indicated a discharge area with negligible vertical hydraulic gradients. Thus, for topographic depressions, piezometric pressure can be considered equivalent to the water table depth. Drain flows are continuously measured in a V-notch weir at the main drain outlet from the field. MCPA was applied on the Näsbygård field on 21 May 1999 at 0.40 kg ha–1 and 13 instantaneous grab samples of the outflow were taken during individual peak flow events between 27 May and 24 June for determination of MCPA concentrations. A sample collected two days before application indicated that residues of MCPA in the drainage water were close to the level of determination, possibly due to intrusion of shallow ground water contaminated with MCPA from previous applications. In a recent study, small concentrations of MCPA (up to 0.1 µg L–1) have been detected in ground water at the Näsbygård field at a 7-m depth (J. Kreuger, unpublished data, 2000–2004).

Pesticide Analyses
Samples were first hydrolyzed with alkali for 15 h at room temperature. After acidification to a pH less than 2, the acid was extracted. These extractions were performed as liquid–liquid extraction (using dichloromethane) until 1997 when it was replaced by solid-phase extraction (ENV+; International Sorbent Technology, UK). To preclude any systematic differences between the two extraction methods, the first 65% of the samples in 1997 were run using both extraction methods. Subsequent extractive alkylation with pentafluorobensylbromide and gas chromatography was conducted as described by Åkerblom et al. (1990). The analyses were performed using gas chromatography and mass spectrometry (GC–MS), with a recovery efficiency of 93% for the liquid–liquid extraction and 99% for the solid-phase extraction. The results were not corrected for recovery efficiency. The limit of determination (LOD) was 0.1 µg L–1 in samples analyzed in 1993–1996, which then improved to 0.02 µg L–1 from 1997. A surrogate analyte (2,4,5-TP) was used to monitor the accuracy and precision of the analytical procedures and concentrations were corrected according to extraction efficiency of the analyte.

Soil Properties
Three separate soil sampling campaigns were performed with the objectives to (i) obtain basic soil data that could be used to develop local-scale pedotransfer functions for model parameters, (ii) compare the spatial variation of fundamental soil properties at the field and catchment scales, and (iii) investigate the relationships between soil properties and landscape position. Thus, a grid soil survey covering the entire Vemmenhög catchment was performed in 1998 at a measuring distance of 333 m, yielding the particle size distribution (in nine size classes) and organic carbon content at 51 locations in the topsoil, 24 locations in the subsoil at both 50- and 75-cm depths, and 18 locations at the 100-cm depth (Svensson, 1999). Also, at four locations across the catchment, soil pits were dug for soil profile descriptions, and cylinder samples (7.2-cm diameter, 5-cm height) taken for measurements of soil water retention and saturated hydraulic conductivity.

At Näsbygård, the relative elevation was surveyed at 77 measuring points and the organic carbon content and particle size distribution of the topsoil (0- to 25-cm depth) were analyzed. At 16 locations, the soil was also sampled at the 25- to 50-, 50- to 75-, and 75- to 100-cm depth. Measurements were also made of near-saturated hydraulic conductivity of the topsoil using a tension infiltrometer, soil organic carbon content, soil particle size distribution, and soil water retention at three landscape positions representative of hilltops, slopes, and hollows. Replicate soil columns (20-cm diameter, 20-cm height) were sampled at each landscape position for subsequent determination of solute transport parameters (e.g., dispersivity) in breakthrough experiments conducted with a nonreactive tracer (Cl–) and MCPA (Roulier and Jarvis, 2003). These studies showed that the soil properties were dependent on topography. On hilltops, the soil has a higher clay content, lower organic carbon content, and higher bulk density than soils located at other landscape positions. Soil located on midslopes is relatively depleted of clay, presumably due to erosion. In the hollows, the soil has a higher organic carbon content and therefore a lower bulk density and higher saturated water content. At tensions of –10 cm H2O, the hydraulic conductivity is smaller for the soil at the hilltop (Roulier and Jarvis, 2003) due to the higher clay content.

The results of the soil surveys showed that the correlation scale of the variation in soil properties was small in relation to typical field sizes, and that the clay content at field and catchment scales had similar means, medians, and ranges. Organic carbon content also showed similar means and medians at both scales, but the range was somewhat larger at the catchment scale (Jarvis et al., 2001). This implies that soil property distributions and pedotransfer functions derived from data collected at Näsbygård may be reasonably representative of the Vemmenhög catchment as a whole. The reason for this is that the catchment is relatively small and homogenous with respect to agro-environmental conditions (soil type, cropping pattern, soil management, and topography).

Incubation Experiments
Soil samples were collected at different depths and landscape positions at Näsbygård in January 2002 for estimation of MCPA degradation rates in laboratory incubation experiments. Two replicate subsamples were taken from each of eight sampled locations, giving sixteen incubation experiments in total. Particle size distribution and organic carbon content were also determined for each sample location.

MCPA degradation did not strictly follow first-order kinetics. Instead, the samples demonstrated the fast metabolic degradation that phenoxy-acid herbicides often undergo (Stenström, 1992). Nevertheless, fitted pseudo-first-order rate coefficients (0.69 < R2 < 0.93) were used in the model analysis since MACRO only allows for first-order degradation kinetics. No significant differences between degradation rates were observed between topsoil and subsoil samples (p < 0.05) and likewise no significant statistical relationships (p < 0.05) could be established between the derived rate coefficients and particle size distribution or organic carbon content.

The Model
MCPA leaching at field and catchment scales is analyzed and interpreted with MACRO Version 4.3b (Jarvis, 1994) using a Monte Carlo technique to account for the spatial variability of model parameters. MACRO is a one-dimensional profile-scale model, and therefore may be an inappropriate choice for many catchment-scale studies. However, due to some special features of the Vemmenhög catchment, MACRO is a suitable model for studying diffuse pesticide losses in this case. The catchment is small, the variability in soil characteristics is limited, and the input and transport routes are all well-defined. There is very little surface runoff and no spray drift input (due to the culverting), and the drainage systems at Vemmenhög create rapid connections between field leaching losses and the catchment outlet. Although base flow contributes significantly to stream flow 1.1 km further downstream, especially during low flow conditions (Kreuger, 1998a), our monitoring data at the outlet of the culverted drainage system showed very little influence of ground water inflow. Thus, the processes that influence pesticide loadings and concentrations in the stream correspond closely to those included in the model.

MACRO is a dual-permeability model that accounts for macropore flow by dividing the soil into two separate but interacting pore systems (micropores and macropores) each characterized by a degree of saturation, conductivity, and water flow rate. In the micropores, vertical water flow is calculated with Richard's equation, the soil water release characteristic is described by the Brooks and Corey (1964) equation, while hydraulic conductivity is calculated by the Mualem (1976) model. In the macropore domain, water flow is treated as a noncapillary, gravity-driven process. If the precipitation intensity exceeds the infiltration capacity of the micropores at the soil surface the excess water is routed to the macropores in the uppermost soil layer. Lateral water exchange from macropores to micropores is calculated using a first-order approximation to the water diffusion equation (i.e., ignoring gravity). Pesticide transport in the micropores is calculated using the convection–dispersion equation with source–sink terms for mass exchange between macro- and micropore domains, lateral leaching losses to drains, and degradation. In the macropores, dispersion is neglected since solute transport is assumed to be dominated by convection. Mass exchange between macro- and micropores is calculated as a combination of diffusion and convection terms (Jarvis, 1994). The solute concentration in water routed to the macropores at the soil surface is calculated assuming an instantaneous local equilibrium and complete mixing of net rainfall with the water stored in a shallow surface soil layer or "mixing depth" (Steenhuis and Walter, 1980). Sorption is described as an instantaneous process calculated according to a Freundlich isotherm. Degradation is predicted in the model assuming first-order kinetics. Field degradation rates are predicted from reference rate coefficients measured in the laboratory, accounting for soil moisture and temperature effects using simple functions (Boesten and van der Linden, 1991).

Potential evapotranspiration is calculated using the Penman–Monteith equation (Monteith, 1965). Root water uptake is calculated as a function of the evaporative demand, soil water content, and root distribution (Jarvis, 1989). Empirical functions are used to define the leaf area and root development of annual crops. In this study, a water table is assumed to be present in the profile, and the vertical recharge to ground water is calculated as a linear function of the time-varying water table height. Water flow to field drains is calculated from all saturated soil layers in the profile using seepage potential theory (Leeds-Harrison et al., 1986).

Driving Data and Simulation Setup
Daily precipitation and meteorological variables were recorded at the synoptic meteorological station Skurup located 6 km to the northeast of the catchment. Measurements made within the catchment during the summer months showed that precipitation amounts and distribution were similar to those at Skurup. Total precipitation amounted to 93 to 95% of the precipitation at Skurup. Hence, 95% of the precipitation at Skurup was used as model driving data.

The simulations were run from 1 Jan. 1992 to 31 Dec. 1999, but we only report results for MCPA losses for the last seven years. The first year of each simulation was considered as a "warm-up" period for the model hydrology as the predictions are not independent of the (unknown) initial conditions.

Model Application
The information on soil properties at Vemmenhög, gathered in the soil surveys, was used to estimate probability distributions for uncorrelated "master" soil properties and to derive pedotransfer functions from which "slave" soil properties, dependent on the "master" soil properties, were calculated. The idea behind "master" variables is that they are fundamental characteristics that cannot be estimated from other properties. The "master" variables were sampled using the Latin hypercube method (McKay et al., 1979). Sensitivity analyses were performed based on the results of Monte Carlo simulations.

Uncorrelated Distributions of "Master" Variables
Based on measured data for the catchment, five fundamental properties (particle size distribution index, characteristic particle size, organic carbon content, pH, and the first-order degradation rate coefficient) were found to be uncorrelated (p < 0.05). Three additional soil properties, reflecting soil structure (mixing depth, topsoil macroporosity, and subsoil macroporosity), were assumed to be uncorrelated, since the amount of data was not sufficient to test for possible correlations. These eight uncorrelated soil properties were treated as stochastic Monte Carlo "master" variables (Table 2).


View this table:
[in this window]
[in a new window]
  		Table 2. Input distributions to Latin hypercube sampling for stochastic variables (error terms excluded).

 
The particle size distribution index and the characteristic particle size were derived by least square regression analysis, fitting a curve of van Genuchten (1980) type to the measured cumulative particle size distributions (Haverkamp and Parlange, 1986). The fits were good, with an average R2 value of 0.99. There were no significant differences between topsoil and subsoil distributions of particle size distribution index (p < 0.05). For the characteristic particle size, the means were not significantly different but the variance of the subsoil was significantly larger (p < 0.05). These variables were treated as constant with depth (Table 2).

Organic carbon content was found to decrease and pH increase with depth. Each of these trends was described by a reference function and associated distributions of scale factors (Table 2), by applying functional normalization based on least squares regression analysis (Tillotson and Nielsen, 1984).

Pedotransfer Functions and "Slave" Parameters
"Slave" parameters were linked to one or more of the eight "master" variables through local pedotransfer functions (Table 3) derived by applying least square regression analyses to the data acquired in the soil survey and measurement campaigns. Six of these pedotransfer functions are accompanied by normally distributed regression error terms ({epsilon}), reflecting the strength of the functional relationships.


View this table:
[in this window]
[in a new window]
  		Table 3. Local pedotransfer functions for calculating "slave" parameter distributions. The error of the estimate, {epsilon}, is normally distributed, with zero mean and standard deviation denoted by {sigma}.{dagger}

 
Deterministic Parameters
A number of model parameters were treated deterministically (Table 4), either because they were considered a priori as less sensitive (Dubus and Brown, 2002), or because insufficient data was available to predict their variation. The organic carbon partition coefficient and the Freundlich exponent were set to average values of 70 cm3 g–1 and 0.75 respectively, based on data contained in unpublished confidential company reports, and results in the published literature (e.g., Riise et al., 1994; Socías-Viciana et al., 1999). Interception of pesticide by the growing crop was calculated as a linear function of the simulated leaf area at the time of application. "Washoff" and dissipation of the intercepted pesticide was calculated using default values in the model.


View this table:
[in this window]
[in a new window]
  		Table 4. Parameters treated deterministically.

 
As a special case, the parameter controlling ground water recharge was determined in preliminary deterministic simulations for the Näsbygård field based on the "average" soil (see definition below) by calibration against measured drainflows and piezometric pressures.

Monte Carlo Analysis at the Field Scale
The "master" variable distributions shown in Table 2 and the pedotransfer functions in Table 3 were used to generate distributions of model parameter values. These are shown in Table 5 for the field-scale simulations at Näsbygård. The distributions of effective diffusion pathlength, saturated hydraulic conductivity, and sorption coefficient show, as expected, significant skewness. The generated distributions used for the catchment-scale simulations are nearly identical since the same functions and distributions were used. However, the generated parameter ranges are slightly different in some cases, since the number of simulations differed. "Cut-off values," based on expert judgement, were used in some cases to prevent unrealistic extreme values (Table 5). Two hundred simulations were run for the Näsbygård field for the period 1992 to 1999, each nominally representing an area of approximately 0.15 ha. The means and coefficients of variation for the two model outputs that we focus on here (the total amount of MCPA leached and maximum concentration of MCPA in the stream) were constant for sample sizes larger than approximately 105.


View this table:
[in this window]
[in a new window]
  		Table 5. Parameter distributions for the field-scale simulations.{dagger}

 
The aggregated output was compared with water discharges and concentrations of MCPA in the outflow from the field in spring 1999. For comparative purposes, a single "deterministic" run using the "average" soil was also performed in addition to the stochastic simulations. The "average" soil is defined by the means of the uncorrelated input distributions, and for the "slave" parameter distributions, estimates calculated with the pedotransfer functions assuming {epsilon} = 0.

Monte Carlo Analysis at the Catchment Scale
We adopted a deterministic treatment of the management of the individual fields within the catchment sprayed with MCPA, utilizing our almost complete knowledge from farmer questionnaires of the field size, dose, timing of application, and dates of crop sowing and harvest (Table 1). Each treated field was randomly assigned stochastic "soil parcels" or "sub-areas," with a probability of assignment in proportion to the size of the field. A total of 620 Monte Carlo simulations were assigned to an area of 598 ha that was treated at least once during the simulation period. Each "sub-area" therefore nominally represents 0.965 ha. For each simulation, three different loss routes are possible: surface runoff to surface drainage inlets, subsurface flow to field drainage systems, and deep percolation to ground water. As noted above, ground water inflow at the monitoring station is thought to be very small, so both percolation and MCPA leaching to ground water simulated by the model were excluded from consideration.

Part of the catchment area (28%, i.e., 230 ha) was not treated with MCPA during the simulation period. There seemed little point in treating this area stochastically. For each of the crops in the rotation (Table 1), hydrological simulations were run for the "average" soil. The drainage and runoff predicted from these five simulations were mixed with drainage and runoff from the stochastic "soil parcels," in proportion to the yearly areal coverage of the different crops.

Sensitivity Analysis
Sensitivity analyses were performed to identify which input variables contributed most to variability in the output. The information obtained from such an analysis is useful when trying to reduce predictive uncertainty, since it determines which variables should be given priority in additional data collection or research to strengthen the knowledge base. It should be noted that although the results of the sensitivity analysis can be used to assess the relative importance of those model input parameters treated stochastically, the actual variability in MCPA leaching will also be significantly influenced by variations in variables that are not included (e.g., spatial variation in rainfall across the catchment, Freundlich exponent, organic carbon partition coefficient).

Two outputs of ecotoxicological interest were investigated, namely the maximum concentration of MCPA and the total loss of MCPA to the stream. Large but transient concentrations of pesticides may lead to acute toxicity in plants and animals, while a prolonged "chronic" exposure to lower concentrations may also be harmful. The first year of application was chosen for the sensitivity analyses for the "sub-areas" in Vemmenhög that were sprayed more than once during the simulation period. The total amount of leached MCPA resulting from one application was calculated from 1 April (i.e., month of earliest application) to 31 December (i.e., winter leaching of MCPA is assumed to be insignificant). Measurements of MCPA concentrations in the stream during the period January to April in 1993 (Kreuger, 1996) and 2002 (Kreuger, 2003) support this assumption. No MCPA was detected in January to April 1993 and only traces of MCPA (concentrations less than 0.02 µg L–1) were found during the same period in 2002.

In addition to the eight "master" variables (Table 2) and six error terms of the local pedotransfer functions (Table 3), four other spatially variable parameters related to the application pattern were included in the analysis: dose of applied MCPA, leaf area index (LAI) at application, and two different measures of precipitation. The precipitation pattern affects pesticide leaching throughout the application year, but this is impossible to describe by only one variable. Two important features of rainfall are the rain amount and distribution. Two sensitivity analyses were performed in which the variation of precipitation was expressed either as the total precipitation or as the maximum daily rainfall amount during a certain number of days following application. The number of days following application was varied to find the optimal representation of precipitation.

Partial rank correlation coefficients (PRCC; Kendall and Stuart, 1979) were calculated between the model inputs and the two "target" model outputs. Rank transformed data gives equal weight to all values in the distributions used in the analyses, irrespective of the type of probability distribution and size of standard deviation. Calculation of a PRCC between two variables eliminates any indirect influence on their relationships due to the other variables (Conover, 1980). Hence, even when the input variables are correlated, PRCCs will determine their independent effects on the output. It should be mentioned that, in the present study, all "master" soil properties and error terms are uncorrelated (except for correlations arising stochastically from the sampling), whereas the variables dose of applied MCPA, LAI at the time of application, and precipitation are to some extent correlated.

Analyses based on PRCCs are qualitative, that is, the input variables are ranked in order of importance and do not give any information on how much a given input variable is more important than another. Only first-order terms (the "main" effects) are accounted for. As a result, the importance of those parameters that influence output mostly through interactions may be overlooked in the analysis. Rank transformations are only appropriate when relationships between inputs and output are monotonic. The R2 value obtained from a multiple linear regression between ranked output and ranked uncorrelated input variables was calculated as a measure of the degree of monotonicity (Saltelli and Sobol, 1995).


   RESULTS AND DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
CONCLUSIONS
 REFERENCES
 
Field-Scale Simulations
Although not shown here, the aggregated output of the stochastic simulations closely resembled measured discharges from Näsbygård: total discharge was overestimated by 12% in the period 1993 to 1999, while the RMSE in daily discharge was 1.04 mm. Since maximum concentrations of MCPA occur shortly after application (Kreuger, 1998a), it is important to capture the small outflows measured during the dry summer months. For the summer months (May to August), the discharge was only overestimated by 6%, and the RMSE in daily discharge was 0.38 mm. Figure 1 shows that for the critical period following MCPA application in 1999, the stochastic approach closely matched the measurements at the Näsbygård field, both in pattern and magnitude, although simulated peaks occurred 5 to 20 h earlier than the measured peaks. These slight time lags are due to the use of daily rainfall data in the MACRO model. In MACRO, the total daily precipitation always starts at midnight, with a duration calculated from a user specified intensity (in this study 5 mm h–1), which is kept constant during the whole simulation period. As shown in Fig. 1, the simulation of a single deterministic parameter set using the means of the input distributions ("average" soil) failed to capture the measured peak outflows giving instead a smooth attenuated drainage response. This is presumably because peak flows are generated by combinations of parameter values (especially those controlling macropore flow) that cannot be captured in a deterministic simulation.



View larger version (28K):
[in this window]
[in a new window]
  		Fig. 1. Measured and predicted (stochastic and deterministic) outflows from the Näsbygård field site for a 40-d period following pesticide application on 21 May 1999.

 
Figure 2 shows that simulated MCPA concentrations were in good agreement with the measurements at Näsbygård, at least with respect to magnitude. MCPA was detected in the outflow from Näsbygård at concentrations up to 14 µg L–1 13 d after application in connection to a heavy rainfall of 22.1 mm that presumably generated macropore flow. In the simulation, there are three subsequent concentration peaks, occurring 15, 16, and 18 d after application corresponding to rainfalls of 8.9, 9.7, and 10.1 mm, respectively. Only the second of these three peaks is matched by the measurements, perhaps due to the relatively low sampling frequency during this period. According to the simulations, 0.05% of the applied amount leached from the field in the 40 d following application. Simulated loss of MCPA was dominated by drainage with 97% of the simulated loss transported by drain flow. Simulated MCPA leaching via drain flow occurred from 127 of the 200 "sub-areas" and one of these "sub-areas" also lost MCPA via surface runoff. Simulated MCPA loss from the "average" soil was extremely small (maximum concentration is 3 x 10–14 µg L–1), which is probably explained by the failure of the "average" soil to predict macropore flow during the time period following application.



View larger version (30K):
[in this window]
[in a new window]
  		Fig. 2. Measured and predicted concentrations of MCPA in outflow from the Näsbygård field for a 40-d period following pesticide application on 21 May 1999. Filled circles represent samples with concentrations below the limit of detection.

 
Catchment-Scale Simulations
Simulated aggregated outflow generally closely resembled measured outflows from the Vemmenhög catchment (RMSE = 0.64 mm d–1; Fig. 3) . In 1993 to 1996, the measuring instrument at the monitoring station often failed to register large outflows, so measured high outflow peaks cannot be compared with simulated flows. With the exception of the very dry month of May in 1993, it is evident that the important outflows occurring during summer months are successfully simulated (discharge was overestimated by 6%, RMSE = 0.20 mm d–1 for May to August, 1993 to 1999). Although not shown here, running the model for the "average" soil gives a total outflow that is very close to that measured, but the simulation fails to capture the small but critically important summer outflows. The stochastic approach manages to capture these small outflows due to drainage resulting from certain input parameter combinations that are missing in the simulation with the "average" soil.



View larger version (27K):
[in this window]
[in a new window]
  		Fig. 3. Measured and predicted outflows from the Vemmenhög catchment during January 1993 to December 1999. Summer outflows in 1993, 1995, 1997, and 1999 are enlarged.

 
Measured MCPA concentrations in the stream are compared with those simulated with the Monte Carlo approach in Fig. 4 , while Table 6 compares total loads on an annual basis. Since both simulated and measured concentrations are often either below or close to the LOD, the estimated loads in Table 6 underestimate the true totals. To protect the aquatic environment, Crommentuijn et al. (1997) defined a no-effect concentration for MCPA of 1.7 µg L–1. Thirteen measured concentrations exceeded the no-effect concentration, with seven of these occurring in 1993, a year when concentrations as large as 10, 30, and 60 µg L–1 were detected in the stream (Fig. 4). These large concentrations were clearly caused by point-source pollution from filling or cleaning spraying equipment, since they occurred during recession flows with no connection to flow peaks in the stream. As an example, Fig. 5 shows this pattern for 1997. Such evidence of significant point sources was noted in all other years except 1998 and 1999. Kreuger and Nilsson (2001) reported that total pesticide loadings to the stream decreased by 90% during the 1990s probably as a result of advisory campaigns targeted at farmers' handling of spraying equipment and the introduction of biobeds (Torstensson, 2000) on many farms in the catchment. Most simulated concentrations are smaller than the LOD, and in 1993, 1995, and 1997 the LOD is never exceeded. All but 4 of the 122 simulated concentrations were smaller than the equivalent measured concentrations. Apart from the fact that the model does not include point sources, this underestimation may also partly arise from model errors or modeller subjectivity, or because one or more input distributions did not accurately reflect conditions in the catchment. In agreement with the limited available measurements, simulated MCPA concentrations during winter (January to April) were very small (<<LOD).



View larger version (20K):
[in this window]
[in a new window]
  		Fig. 4. Measured and predicted concentrations of MCPA above the limit of determination (LOD) at the catchment monitoring station. The dotted line marks the LOD. A total of 46 measured concentrations and 113 simulated concentrations were less than the LOD. Simulated concentrations were averaged to match the time resolution in the measurements (composite samples taken once or twice weekly).

 

View this table:
[in this window]
[in a new window]
  		Table 6. Simulated and measured losses as percentage of applied amount of MCPA in the Vemmenhög catchment.

 


View larger version (22K):
[in this window]
[in a new window]
  		Fig. 5. Measured outflow and concentrations of MCPA in the outflow from the Vemmenhög catchment in 1997.

 
Although it is recognized that the model is only validated against the integrated observed response at field and catchment scales, it is interesting to note that between 1 and 17% of the sprayed "sub-areas" were predicted to contribute 90% of the annual amount of MCPA leached in the years 1993 to 1999. This implies that a "best management strategy" to minimize impacts on surface waters would be avoidance of pesticide applications to fields or parts of fields within a catchment identified as "hot spots" for diffuse losses.

Sensitivity Analysis
Field-Scale Sensitivity Analysis
The R2 values for the multiple linear regressions at Näsbygård were 0.74 and 0.75 for the total amount of MCPA leached and the maximum concentration, respectively. Hence, the model is at least 74 to 75% monotonic. The results of the field-scale sensitivity analysis for the two different output variables were almost identical since they were very strongly correlated. Therefore, only PRCC values for the total amount of MCPA leached are shown in Table 7. These results should be interpreted with care, since the mutual order of ranking among variables with similar PRCC values is uncertain, due to possible non-monotonicity of the model (A. Saltelli, personal communication, 2003). Moreover, the replicability of the mutual order of ranking among variables with little influence on model predictions is very low (Dubus and Janssen, 2003).


View this table:
[in this window]
[in a new window]
  		Table 7. Partial rank correlation coefficients (PRCC) for total amount of leached MCPA for Näsbygård.{dagger}

 
Of the 14 input variables, 9 were significantly (p < 0.05) correlated to the simulated amount of MCPA leached (Table 7). The pedotransfer error for the effective diffusion pathlength, regulating the exchange of water between the macropore and micropore domains, was the most important variable. This calls for further research to elucidate what these unexplained variations depend on. Since it is impossible to measure, a substantial part of the pedotransfer error may originate from the uncertainty in estimating effective diffusion pathlength by inverse procedures (Roulier and Jarvis, 2003). The significant influence of bulk density and the Brooks–Corey pore size distribution index is probably due to the influence of these variables on the saturated matrix conductivity through the pedotransfer functions (Table 3). As bulk density increases, the saturated matrix water content decreases, and so does the Brooks–Corey pore size distribution index. Both lead to a decrease in the saturated matrix hydraulic conductivity, which will in turn generate macropore flow more frequently. This is supported by the high ranking (third) of saturated matrix conductivity in the sensitivity analysis. Thus, four of the five most influential input variables are strong determinants of the strength of macropore flow as simulated in the model. The Brooks–Corey pore size distribution index may also be important in itself: a decrease in the Brooks–Corey pore size distribution index decreases the drainable porosity, which increases susceptibility to leaching to drains.

Since the influence of mixing depth and matrix dispersivity on predicted leaching are small compared with other variables, time-consuming calibration activities to find good representations of these variables may not be justified. Indeed, in the inverse modeling effort reported by Roulier and Jarvis (2003), these two parameters were so insensitive that their initial uncertainty domains could not be reduced in some cases.

Of the "master" soil properties, organic carbon content was the most important variable, ranking fourth overall. The influence of organic carbon content is due to its combined effect on the effective diffusion pathlength (Table 3) and the sorption distribution coefficient of the pesticide. Variation in sorption is mainly explained by the variations in organic carbon content (R2 = 0.99 for unranked data), since pH in the calcareous soil at Näsbygård is much larger than the pKa value of MCPA. The influence of pH on predicted MCPA leaching was too small in relation to the other variables to be ranked at this level of significance (p < 0.05).

It is interesting to note that the influence of variations in degradation rate was small in relation to the other variables. This is quite contrary to the results of previous studies. For example, Dubus and Brown (2002) investigated the sensitivity of pesticide leaching simulated by the MACRO model to 43 primary input parameters in four contrasting scenarios, and found the degradation rate constant to be one of the most influential input parameters (ranked first to fourth). This discrepancy is probably explained by the fact that the nominal values for the degradation rate constants assumed for the two pesticides studied by Dubus and Brown (2002) were smaller and their range much broader than for MCPA in our study. Indeed, the variability in MCPA degradation found in the incubation experiments for Näsbygård soil was unusually small (CV = approximately 8%, Table 2). This may partially be explained by the small number of samples (N = 16), but it is more probably due to the fact that rapid metabolic degradation was observed. Other studies of field-scale variability of pesticide degradation usually point to larger variation. Walker et al. (2002) found that within a single field the half-life of isoproturon varied from 6 to 30 d and for chlorotoluron from 34 to 203 d. This comparison emphasizes that parameter rankings for sensitivity to pesticide leaching are highly site and compound specific.

Catchment-Scale Sensitivity Analysis
Like the field-scale sensitivity analysis, the results for the catchment-scale sensitivity analysis were very similar for both target outputs (MCPA load and MCPA maximum concentration) (R2 = 0.99 for ranked data). This is because of the very short half-life of MCPA, which means that almost all the leaching occurs soon after application in a limited number of flow "events." Thus, results for only one of the target outputs are presented in Table 8.


View this table:
[in this window]
[in a new window]
  		Table 8. Partial rank correlation coefficients (PRCC) for total amount of leached MCPA for Vemmenhög.{dagger}

 
Of the 17 input variables, 13 significantly influenced the predicted loss of MCPA (Table 8). The variation in precipitation following MCPA application proved to be extremely significant, only outranked by the pedotransfer error of the effective diffusion pathlength. However, the ranking of the precipitation variable was very sensitive to how it was calculated. A maximum rank (second) was reached when it was calculated for a 17-d period following application. Precipitation was not among the top 10 influential variables when it was calculated for a period of 1 to 12 d following application. This suggests that the heavy rainstorms that generate MCPA leaching at Vemmenhög are infrequent events.

Ranked sixth, LAI at the time of application was almost as influential as organic carbon content. Consequently, application timing apparently had a larger effect on pesticide leaching than the variation of many of the soil properties over the catchment. Later MCPA applications generally result in smaller pesticide leaching since a smaller fraction reaches the soil surface due to crop interception and because the soil is generally drier. Excluding autumn sown crops, the median application of MCPA took place 43 d after sowing, but with a wide range (11–73 d). A large range of leaf area indices (0.03–3.72) was therefore simulated at the time of application. Ranked ninth, the variation in the actual dose of MCPA was apparently much less important than variations in crop interception in controlling predicted leaching via the effective dose at the soil surface.

For the soil properties and pedotransfer errors, the mutual ranking was very similar to the Näsbygård field-scale sensitivity analysis. This is not surprising since the Monte Carlo samplings were based on identical parameter distributions in the two cases.


   CONCLUSIONS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
REFERENCES
 
Cultivated soils are often very heterogeneous with respect to hydraulic and chemical properties, allowing a large range of transport velocities, and conditions for degradation and sorption. The stochastic parameter sets created to account for these variations successfully predicted the hydrologic response of the field and catchment. In contrast, a single deterministic parameter set using the means of the input distributions ("average" soil) failed to reproduce the small but important summer outflows captured by the stochastic simulation. For the field-scale simulation in spring 1999, the "average" soil also greatly underestimated MCPA leaching, due to the low drainage simulated during the critical period following application. With the stochastic approach, however, there was a good resemblance between the aggregated output of the simulations and measured MCPA concentrations in drainflow at the field scale. At the catchment scale, the stochastic approach underestimated the concentrations of MCPA in the stream. This is mostly attributed to the occurrence of point-source contamination, but other factors probably also contributed, including model errors and the fact that the distributions used for the input variables may not have accurately reflected conditions in the catchment.

The error in calculating the effective diffusion pathlength was the most influential variable in the sensitivity analysis. This reflects the critical importance of soil structure and macropore flow and this is also supported by the fact that three other parameters regulating macropore flow were among the five most important soil properties affecting MCPA leaching losses. Ranked second, the variation in precipitation during the 17 d following application appeared to be very important for predicting MCPA leaching. Of the "master" soil properties, organic carbon content was the most significant variable, since it influenced both the effective diffusion pathlength and the sorption distribution coefficient. With regard to management practices, LAI at the time of application was ranked sixth, and was more important than the variation in many soil properties over the catchment. In contrast to many other studies, the measured variability in degradation rate was small and had little effect on variations in MCPA leaching losses.


   ACKNOWLEDGMENTS
 
This study was funded partially by the EU project CAMSCALE ("Upscaling, Predictive Models and Catchment Water Quality") under the contract ENV4-CT97-0439, and also by the FORMAS (the Swedish Research Council for Environment, Agricultural Sciences and Spatial Planning) project "Understanding the Variability of Pesticide Losses to Surface Waters and Groundwater at Field and Catchment Scales." Thanks are due to Gunborg Alex, Thomas Berglöf, Göran Jonsäll, Eva Lundgren, Märit Peterson, and Åsa Ramberg at the Organic Environmental Chemistry Unit, Department of Environmental Assessment at the Swedish University of Agricultural Sciences for carrying out pesticide residue analyses. The MCPA incubation experiments were carried out by Elisabet Börjesson at the Department of Microbiology, Swedish University of Agricultural Sciences. The authors are grateful to the farmers in the catchment for their cooperation, Sten Hansson for his assistance with field work, and Göran Tuesson at the Rural Economy and Agricultural Society Kristianstad who maintained the sampling equipment. Thanks are also due to Professor Lars Bergström for helpful comments in reviewing this manuscript.


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 


Related articles in JEQ:

This Issue in Journal of Environmental Quality

JEQ 2005 34: v. [Full Text]  




This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Related articles in JEQ
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed