JEQ Grow Your Career With ASA
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

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 (6)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Miao, Z.
Right arrow Articles by Del Re, A. A. M.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Miao, Z.
Right arrow Articles by Del Re, A. A. M.
Agricola
Right arrow Articles by Miao, Z.
Right arrow Articles by Del Re, A. A. M.
Related Collections
Right arrow Multicomponent Transport Models
Right arrow Water Flow Models
Right arrow Runoff
Right arrow Agricultural Pesticides
Right arrow Water Pollution
Published in J. Environ. Qual. 33:2217-2228 (2004).
© ASA, CSSA, SSSA
677 S. Segoe Rd., Madison, WI 53711 USA

TECHNICAL REPORTS

Surface Water Quality

Uncertainty Assessment of the Model RICEWQ in Northern Italy

Zewei Miao, Marco Trevisan*, Ettore Capri, Laura Padovani and Attilio A. M. Del Re

Istituto di Chimica Agraria ed Ambientale, Università Cattolica del Sacro Cuore, 29100 Piacenza, Italy

* Corresponding author (marco.trevisan{at}unicatt.it)

Received for publication January 14, 2004.

   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Model predictions are often seriously affected by uncertainties arising from many sources. Ignoring the uncertainty associated with model predictions may result in misleading interpretations when the model is used by a decision-maker for risk assessment. In this paper, an analysis of uncertainty was performed to estimate the uncertainty of model predictions and to screen out crucial variables using a Monte Carlo stochastic approach and a number of statistical methods, including ANOVA and stepwise multiple regression. The model studied was RICEWQ (Version 1.6.1), which was used to forecast pesticide fate in paddy fields. The results demonstrated that the paddy runoff concentration predicted by RICEWQ was in agreement with field measurements and the model can be applied to simulate pesticide fate at field scale. Model uncertainty was acceptable, runoff predictions conformed to a log-normal distribution with a short right tail, and predictions were reliable at field scale due to the narrow spread of uncertainty distribution. The main contribution of input variables to model uncertainty resulted from spatial (sediment–water partition coefficient and mixing depth to allow direct partitioning to bed) and management (time and rate of application) parameters, and weather conditions. Therefore, these crucial parameters should be carefully parameterized or precisely determined in each site-specific paddy field before the application of the model, since small errors of these parameters may induce large uncertainty of model outputs.

Abbreviations: 1REAT, chemical concentration in first paddy water runoff • CUM, cumulative chemical concentration in paddy runoff after a 21-day treatment period • DAT, days after treatment • IPEU, input parameter error uncertainty • Kd, sediment–water partition coefficient • LOD, limit of detection • MCS, Monte Carlo simulation • MEU, model error uncertainty • NVU, natural variability uncertainty • TU, total uncertainty • TW21, time-weighted chemical concentration in water runoff over a 21-day treatment period • VBIND, mixing depth to allow direct partitioning to bed


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
PESTICIDE USE IN RICE (Oryza sativa L.) cultivation affects the quality of surface waters. The potential for pesticide contamination of water bodies is high in areas where rice is cultivated under flooded conditions. Paddy water management (e.g., irrigation and drainage) increases the likelihood of transport of pesticides to surface water via runoff or drainage (Capri and Miao, 2002) and raises the complexity of pesticide behavior in paddy fields, so measuring representative paddy water quality is difficult and expensive. Previous monitoring studies have shown that water quality in paddy rice should be evaluated at the watershed level. However, variability of rainfall, topography, soil properties, vegetative cover, and farming practices make it difficult to accurately predict water runoff, soil erosion, and pesticide off-field transport (Fontaine et al., 1992; Bobba et al., 2000; Wolt et al., 2002).

Recently, mathematical models have been used to investigate pesticide concentrations in the paddy environment and to assess the fate and transport of pesticides at various scales; RICEWQ (Williams et al., 1999), PADDY (Inao and Kitamura, 1999), and PCPF (Watanabe and Takagi, 2000) can be cited. However, predictions of a deterministic model are often uncertain: (i) the model structure can be invalid and the equations describing the chemical processes can be incorrect; (ii) the parameters can be chosen incorrectly; and/or (iii) some input data can be wrong. Sources of uncertainty can be grouped into three categories: (i) errors resulting from the conceptual scheme of the world (i.e., model error); (ii) stochasticity of the real world due to temporal and spatial variability (i.e., natural variability); and (iii) uncertainty of the model parameters (i.e., input parameter error) (Dean et al., 1989; Jian and Schilling, 1996; Trevisan et al., 2001; Hession and Storm, 2000; Keller et al., 2002).

The importance of incorporating uncertainty analysis into pesticide fate models has been emphasized by many authors (Dean et al., 1989; Tiktak, 1999; Hession and Storm, 2000). Ignoring the uncertainty associated with model predictions may result in misleading interpretations, incorrect evidence might be reached by comparing with field measurements and isolated model predictions, and incorrect conclusions might be drawn when the model is used for risk assessment by the decision-maker (Soutter and Musy, 1998; Tiktak, 1999; Keller et al., 2001, 2002). Bobba et al. (2000) suggested that the inclusion of uncertainty analysis in modeling activities can give a truthful depiction of model limits, and that uncertainties must be estimated and included in modeling activities.

The RICEWQ model was developed as a simple deterministic model for pesticide exposure assessment in the United States, and may be a suitable model for predicting small-scale and edge-of-field pesticide concentration in surface water in Europe; it has been calibrated in Italian paddy fields (Williams et al., 1999; Capri and Miao, 2002; Miao et al., 2003a). The RICEWQ model was selected because existing pesticide transport models are not yet configured to simulate the flooding conditions, overflow, and controlled releases of water that are typical under rice production. Furthermore, RICEWQ has been validated on eight rice paddies within Arkansas and Louisiana of the United States (Williams et al., 1999). The model is user-friendly and was intentionally designed to be algorithmically parsimonious to minimize input and validation data that are seldom available. Nevertheless, its uncertainty at paddy-field scale is not known.

The purpose of this study is to evaluate the uncertainty of RICEWQ model outputs and to screen uncertainty sources both of the model itself and of input parameters. To this aim, paddy field measurements are objectively compared with model forecasting by means of a statistical analysis of stochastic Monte Carlo virtual simulations.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
The RICEWQ Model and Its Improvement
The RICEWQ model is a water quality simulation model used to evaluate the dissipation of chemicals in aquatic systems and to predict the losses of agrochemicals to receiving waters. The model was developed to simulate water and chemical mass balance associated with unique governing processes in paddy fields. Water mass balance takes into account precipitation, evaporation, seepage, overflow, irrigation, and drainage. Pesticide mass balance takes into account dilution, advection, volatilization, partitioning between water and sediment, decay in water and sediment, burial in sediment, and resuspension from paddy sediment. The model can simulate the fate of up to five chemicals or metabolites in the unique flooding, overflowing, and controlled water-release conditions that are typical of rice production. Chemical mass balance is represented by Eq. [1]:

[1]
where is the rate of change of pesticide concentration (10–6 kg m–3 d–1); {sum}Minflux (10–6 kg) and {sum}Moutflux (10–6 kg) are cumulative influx and outflow of chemical mass, respectively, from a given daily volume, V (m3 d–1) (i.e., the rice paddy); and {sum}Mreact is mass transformation in all processes (10–6 kg). The pesticide conservation of mass equations for water, sediment, and foliage subsystems are depicted in Eq. [2], [3], and [4], respectively:

[2]

[3]

[4]
where , , and are the rate of change of chemical mass in water, sediment, and foliage, respectively (10–6 kg 10–3 m–3 d–1); MWapp is the daily mass of applied pesticide not lost by drift and arriving to the water surface (10–6 kg 10–3 m–3 d–1); MFapp is the daily mass of applied pesticide intercepted by foliage (10–6 kg 10–3 m–3 d–1); Mwash is the daily mass washed off from foliage (10–6 kg 10–3 m–3 d–1); MWdeg, MSdeg, and MFdeg are the daily masses of pesticide degraded in water, sediment, and foliage, respectively (10–6 kg 10–3 m–3 d–1); MWtran, MStran, and MFtran are the daily masses of metabolite formed by transformation of parent compound in water, sediment, and foliage, respectively (10–6 kg 10–3 m–3 d–1); Mvolat is the daily mass volatilized across the air–water interface (10–6 kg 10–3 m–3 d–1); Mout is the daily mass lost in overflow or drainage (10–6 kg 10–3 m–3 d–1); Mseep is the daily mass lost in seepage (10–6 kg 10–3 m–3 d–1); Mbed is the daily mass transfer to bed sediment by direct partitioning (10–6 kg 10–3 m–3 d–1) ; Msetl is the daily mass transfer to sediment by particulate settling (10–6 kg 10–3 m–3 d–1); Mresus is daily resuspended mass (10–6 kg 10–3 m–3 d–1); Mdifus is the daily mass diffused between water and sediment (10–6 kg 10–3 m–3 d–1); and Mharv is the daily mass of pesticide removed after harvest (10–6 kg 10–3 m–3 d–1) (Williams et al., 1999; Capri and Miao, 2002).

The RICEWQ model uses the following equation to estimate water balance in the paddy field:

[5]
where is the rate of change of water storage (m3 d–1), which is equal to the cumulative sum of inflow sources ({sum}Ii m3 d–1) minus the cumulative sum of outflow ones ({sum}Oi m3 d–1).

Version 1.6.1 of RICEWQ uses two parameters to control water balance within a rice paddy: the minimum water depth (DMIN0) and the maximum water depth (DMAX0). If the water level is below DMIN0, then irrigation starts and continues until the water level equals DMAX0. If the water level is higher than DMAX0, then runoff occurs until the water level returns to DMAX0. While DMIN0 values change on a monthly basis, only one fixed DMAX0 value can be used to control overflow-based runoff during the whole cropping season. In reality, the depth of paddy may vary during the season. To account for this problem, the model code (Version 1.6.1) was modified to allow DMAX0 changes during the cropping season. This improved version of RICEWQ is now capable to simulate realistic paddy runoff during the cropping season.

Field Data and Model Parameterization
The field data were obtained from a monitoring study performed in the Mantova Province (45°09'N, 10°12'E; altitude = 20 m), located in an area intensively cropped of northern Italy where rice covers 62 different soil types in an area of 30895 ha (Fig. 1) . The soil properties (top horizon) in the basin strongly vary. Clay ranges from 1 to 47% (average 22%), sand from 6 to 82% (average 38%), organic carbon from 0.1 to 12.2% (average 3%), and pH from 6.0 to 8.3 (average 7.9). Sediments in paddies and water bodies are slightly different: they varied from silty-clay and 1.5% organic carbon in paddies to sandy and less than 0.5% organic carbon in water bodies.



View larger version (64K):
[in this window]
[in a new window]
  		Fig. 1. Location of Mantova province (Lombardia region, northern Italy) with the main irrigation–drainage canals (lines) and sampling sites (A and B) of the study area.

 
In two paddy fields, runoff water from the paddy was sampled twice in 1999 and 2000. The first sampling was made two days after application of the pesticide and the second sampling after rice harvesting, three months later. In both fields, tricyclazole (5-methyl-1,2,4-triazolo[3,4-b]benzothiazole) was applied as Beam (Dow Agrosciences, Indianapolis, IN) by airplane on 25 July in both years. An application rate of 0.6 kg a.i. ha–1 (label dose) was estimated after a survey of the whole basin. Owing to the crop sensitivity to fungus diseases, the whole watershed was covered in one day by the aerial application of tricyclazole. Average values of paddy water level in the basin were estimated to be around 5 cm at treatment time, and rose to 10 cm in June, to 20 cm in July, and to 30 cm in August. The estimated drainage rate was 2.59 cm d–1 ha–1. Rice emerged on 20 May 1999, and was harvested on 3 Sept. 1999. In the paddy field, the depth of active sediment was 5.0 cm.

Water samples (2 L) were extracted three times with 100 mL of dichloromethane and the organic phases were collected and filtered on anhydrous sodium sulfate, reduced to a small volume with a gentle nitrogen stream, and then dried by rotary vacuum evaporator and reconstituted with 1.0 mL of acetone. Final samples were quantified by gas chromatography (GC)–nitrogen and phosphorus detection (NPD) (GC100 DLS; Dani, Milan, Italy) and the positive sample confirmed by GC–mass spectrometry (MS) (GC MS 5973N; Agilent, Palo Alto, CA). The limit of detection (LOD) of tricyclazole was 0.1 µg L–1. The recovery of tricyclazole from untreated spiked water was 87.2 and 90.3% with fortification from 0.10 to 1.70 µg L–1, respectively.

The RICEWQ model input variables and parameters are summarized in Table 1. Input parameters were derived from field and laboratory sample measurements (e.g., pesticide application rate, organic carbon content, soil bulk density), from empirical estimates, from the literature (Cheng, 1990; Tomlin, 1994; van der Werf, 1996; Klepper and den Hollander, 1999; Williams et al., 1999; Brawley et al., 2000; Capri et al., 2001), or extrapolated from models like SOILPAR (Donatelli et al., 1996) or RadEst (Donatelli et al., 2000).


View this table:
[in this window]
[in a new window]
  		Table 1. Summary of the main input parameters and variables used in the RICEWQ model.

 
Daily climatic data from 1991 to 1999 were averaged from nine neighboring stations, which are scattered over the whole study area (Fig. 1). Evapotranspiration was assumed to be equal to pan evaporation, which was a valid assumption for a paddy aquatic environment (Linsley and Franzini, 1979; Williams et al., 1999). The potential evapotranspiration was calculated with the Penman–Monteith equation using the RadEst model.

Uncertainty Analysis
Monte Carlo Simulations
Monte Carlo simulations (MCS) were used to estimate the uncertainty of RICEWQ model forecasting. The Monte Carlo approach requires probability density functions for each parameter (i.e., soil properties, pesticide properties, and water management and crop practices). The most insidious weakness of the Monte Carlo technique is a mis-specification of the parameter distribution (Dubus and Brown, 2002; Dubus and Janssen, 2003; Warren-Hicks et al., 2002). Small changes in the parameter distribution assumptions can dramatically change the shape of the resultant Monte Carlo distribution. For example, the use of a series of independent normal distributions rather than multivariate normal distributions (with an appropriate covariance matrix) produces large differences in the resultant Monte Carlo distribution (Warren-Hicks et al., 2002), or if a normal distribution is misunderstood as a uniform one, model uncertainty may be enlarged.

So far, it has not yet been clearly defined how many model simulations should be done for a successful MCS (Warren-Hicks et al., 2002). The number of simulations usually depends on a number of factors: (i) the nature of the parameter that is being estimated, (ii) the form of underlying distribution, (iii) the variability in the observations, (iv) the degree of precision and/or accuracy desired, (v) the level of confidence to be associated with the estimates, and (vi) the actual statistical estimator used to provide the estimate (Warren-Hicks et al., 2002). For example, probabilities of PELMO modeling results were still significantly influenced by the seed number, even for those cases in which 5000 models simulations were undertaken (Dubus and Janssen, 2003). According to Parysow et al. (2000) and Dubus and Janssen (2003), when the number of simulations is greater than 2500 or 3000, the variability of a model projection generally tends to be stable and very close to the results with 5000 or 6000 runs. The MCS sample size was extrapolated from Eq. [6]:

[6]
where n is the sample size (i.e., number of simulations), p and q are the event probabilities, and L is the error interval beyond the confidence interval of 95 or 99% (Snedecor and Cochran, 1989). On the assumption that the probability of chemical runoff after the application is 50% and the confidential interval is 95%, the sample size was set to 400 simulations with 9-yr weather data sets to reach a 95% significant confidence. Thus, a total of 3600 runs were performed.

In this study, 20 variables, which are affected by either input parameter error or natural variability, were chosen to perform uncertainty analysis and to assess their effects on RICEWQ output variability. These selected variables can be broadly classified into three groups: (i) crop and water management descriptors (seven parameters), (ii) chemical properties (seven parameters), and (iii) soil properties (six parameters) (Table 2). The parameters associated with water depth in the paddy field contribute not only to model error but also to input parameter error.


View this table:
[in this window]
[in a new window]
  		Table 2. Input parameter distribution used in Monte Carlo simulation (MCS).

 
In Table 2 the input parameter distributions used in the uncertainty assessment are shown. The probability distribution of the organic carbon content (%) of the 62 soil types is log-normal, with a sample mean of 0.44 and a standard deviation of 0.77. The probability distribution of the soil bulk density is normal, with a sample mean of 1.43 g cm–3 and a standard deviation of 0.12 g cm–3. The probability distribution of the water half-life is assumed to be normal, with a sample mean of 30 and a standard deviation of 2 d. The probability distribution of the square root of the sample wilting point of the 62 soil types is uniform, with a minimum of 0.10 cm3 cm–3 and a maximum of 0.49 cm3 cm–3.

When a parameter distribution was unknown, a uniform distribution was chosen because it provides a reasonable alternative to guessing at the shapes of distributions with limited data or knowledge (Warren-Hicks et al., 2002). Ranges, arithmetic mean, and standard deviation reported in Table 2 were used to generate random variables of uniform or normal distributions. For each variable set used for a given simulation, pseudo-random values were generated by means of a spreadsheet-based Monte Carlo approach (Janssen et al., 1994; Capri et al., 2001; Hood, 2001; Dubus et al., 2003).

For the sediment–water partition coefficient (Kd) an uncertainty evaluation was not made, since Kd (L kg–1) is related to organic carbonic content. Instead, the organic carbon partitioning coefficient (Koc) of tricyclazole was set to 3424 L kg–1 (Knowles, personal communication, 2001) and the Kd value was calculated as a function of the pseudo-random organic carbon values.

Assessment of Model Performance
Model performance was assessed by comparing field measurements with model predictions of 62 soil-type field measurements during a 9-yr period (de Vries and Bakker, 1998; Heuvelink, 1998; Tiktak, 1999; Dubus et al., 2002; Dubus and Brown, 2002).

The following formula was applied to define the general model performance:

[7]
where PE is percentage exceedance of predicted data versus observed data (%), n is the number of Monte Carlo iterations, Mj {where j = 1 to the number of field observations} is an individual field measurement, Pi {i = 1 ordinal number of Monte Carlo iterations} is an individual model prediction, and Xi is an indicator variable (Warren-Hicks et al., 2002).

The expected value of PE is 50% when half of the model predictions are above and half below the measured values. When PE value is around 50%, the model can be considered to be reasonably predictive; when it is less than 25% or more than 75% the model is less accurate, but acceptable, given the variability of the model parameters; if it is near 0 or 100%, the model is likely to be too inaccurate (Warren-Hicks et al., 2002).

Determination of Uncertainty Contributions
Analysis of variance was used to estimate the model error, the natural variability, and the input parameter error (Miao et al., 2003c). The contributions of input parameters and weather conditions to total uncertainty were represented with a multiple regression. In the process of fitting the regression, the analysis was made by using the climatic conditions (9 yr) and the input parameter as class variables, and as dependent variables we used: (i) pesticide concentration (µg L–1) in the first runoff event after treatment (1REAT); (ii) cumulative pesticide concentration (µg L–1) in water runoff over a 21-day treatment period, calculated from cumulative fluxes of water and pesticide (CUM); and (iii) time-weighted average pesticide concentration (µg L–1) in water runoff over a 21-day treatment period (TW21). It is worthy to note that 1REAT heavily depends on time and strength of the first runoff and drainage event, that CUM points to the magnitude of pesticide runoff, and that TW21 represents the intensity of pesticide runoff.

The total uncertainty (TU) was assumed to be estimated by the coefficient of variation of each output and was thus calculated as:

[8]
where MSct is the sum of the square of the corrected total residuals and is the mean of model output dependent variable.

The TU could be split into (i) input parameter error uncertainty (IPEU), (ii) natural variability uncertainty (NVU), and (iii) model error uncertainty (MEU). Among them, IPEU, NVU, and MEU were respectively calculated as:

[9]
where MSr is the mean square of the input parameter variables residuals, {tau}0.05 is the confident level of 95% of Student's t test, and DFr is the number of degrees of freedom of the input parameter variables:

[10]
where MSy is the mean squares of the climatic conditions (i.e., years), {tau}0.05 is the confident level of 95% of Student's t test, and DFy is the number of degree of freedom of the climatic condition (i.e., years):

[11]
where MSe is the mean square of the estimate error, {tau}0.05 is the confident level of 95% of Student's t test, and DFe is the number of degree of freedom of the estimate error.

Forward stepwise multiple regression was used to show how input error affects output total uncertainty by building a multiple regression model with a variable-screening criterion of p = 0.05 (SAS Institute, 1985; Mazumdar et al., 1999; Wisniak and Polishuk, 1999; Miao et al., 2003c; Dubus et al., 2003). The dependent variables were 1REAT, CUM, and TW21.

The 20 class variables of ANOVA were used as nonforced variables for the stepwise multiple regression. The validity of regression methods was checked using the adjusted coefficient of determination value. After log(Y + 1) algorithm transformation of the three dependent variables (1REAT, CUM, TW21), their R2adj of linear regressions were 0.406, 0.420, and 0.346, respectively, which all rank significant at a confidence level of 99.99% (P < 0.0001) by the Fisher test. With the exception of a few outliers, residuals fluctuated in a random pattern around the base line 0, and a normal probability Q–Q plot of the residuals follows a roughly straight line, which reveals a good linear regression function between independent variables and log-transformed dependent variables (Fig. 2) . Moreover, a correlation test confirms that there is no significant mutual correlation between the 20 input parameters produced by Monte Carlo at a confidence level of 95%.



View larger version (36K):
[in this window]
[in a new window]
  		Fig. 2. The residual scatter plot of log(Y + 1) transformation of the three Monte Carlo simulation (MCS) dependent variables against the fitted values (3600 runs): (a) tricyclazole concentration (µg L–1) in first paddy water runoff event (1REAT), (b) cumulative tricyclazole concentration (µg L–1) in paddy runoff over a 21-day treatment period (CUM), and (c) time-weighted concentration (µg L–1) in water runoff over a 21-day treatment period (TW21).

 
The standardized regression coefficients of multiple regression model (SRCi) (Janssen et al., 1994) were used to account for the contributions of each input parameter to model output uncertainty:

[12]
where {sigma}Xi and {sigma}Y are the standard deviation of the independent variables Xi and the dependent variable Yj, p is the number of random variables, and bi is the regression coefficient of Eq. [13]:

[13]
in which {epsilon}j is the residual error and N the number of data points.


   RESULTS AND DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
CONCLUSIONS
 REFERENCES
 
Observed versus Predicted Data
Results of the field measurements are reported in Table 3. Tricyclazole was detected in paddy surface water and also in the irrigation water inlet after treatment. However, it quickly disappeared during the following days, and its residues were always below the detection limit during the winter.


View this table:
[in this window]
[in a new window]
  		Table 3. Measured tricyclazole concentrations (means of three replicates ± SD) in surface water at the paddy field.

 
Observed and predicted values (Table 4) can be compared at paddy level. The simulated mean concentration of tricyclazole in paddy surface water was 3.56 µg L–1 at day after treatment (DAT) 2 and <LOD at DAT 30, which is close to the measured value of 3.55 µg L–1 at DAT 2 and <LOD at DAT 93.


View this table:
[in this window]
[in a new window]
  		Table 4. Predicted tricyclazole concentrations in paddy runoff for several days, concentration in the first paddy water runoff event (1REAT), cumulative concentration in paddy runoff over a 21-day treatment period (CUM), and time-weighted concentration in water runoff over a 21-day treatment period (TW21).

 
Figure 3 shows the distribution of the simulated concentration of pesticide in the water runoff from the paddy field, obtained from 558 samples (62 soil types times 9 yr). The distribution is nearly normal at DAT 0, left-skewed normal at DAT 2, and log-normal both at DAT 7 and 28. Ranges are very wide and varying, for instance, from 0 to 2.39 µg L–1 at DAT 28.



View larger version (25K):
[in this window]
[in a new window]
  		Fig. 3. The probability distribution of daily runoff concentration (µg L–1) predicted at field level (558 runs): (a) DAT 0, (b) DAT 2, (c) DAT 7, and (d) DAT 28, where DAT is days after treatment.

 
A percentage exceedence value of 47.0% at DAT 2 indicates that measured values are in the middle of the predicted distribution and are in agreement with simulated data. At DAT 90, 99.9% of either the measured concentration or the 558 estimates is lower than LOD (<0.1 µg L–1). Thus, the model results can be accepted as reasonable forecasting of field data, notwithstanding field data uncertainty due to the site-to-site variability of soil properties at paddy-field scale. However, according to previous research (Miao et al., 2003b), upscaling of the model from field to watershed level was not applicable.

Uncertainties Due to the Model
The prediction uncertainties due to the model MCS were evaluated using different approaches: output distribution, ANOVA, and stepwise multiple regression.

Monte Carlo Simulation Model Output Frequency Distributions
The frequency distributions of the three indicators (1REAT, CUM, and TW21) appear to comply with a log-normal distribution (Fig. 4) . Their 9-yr averaged coefficients of variation are 151, 156, and 117% (Table 4). The distributions show a high degree of skewness and kurtosis, particularly in the case of CUM. The 95th percentile concentrations are roughly three times the concentration means.



View larger version (17K):
[in this window]
[in a new window]
  		Fig. 4. The probability distribution of the 9-yr runoff concentrations (µg L–1) predicted with Monte Carlo simulation (MCS) virtual parameters (3600 runs). 1REAT, chemical concentration in first paddy water runoff; CUM, cumulative chemical concentration in paddy runoff over a 21-day treatment period; TW21, time-weighted chemical concentration in water runoff over a 21-day treatment period.

 
Simulation uncertainty is directly related to the spread of the resultant distribution. For log-normal distribution, the uncertainty is related to the size of the right tail. As shown in Fig. 4, when values are classified into 10 equal classes, the distribution right tail is narrow, and as many as 89, 94, and 92% of the samples of the three indicators fall within the first two classes. This means that, although the variance of tricyclazole predicted concentrations in runoff is high, the overall uncertainty of the simulations is low. It has been stated that a model error of a factor of 2 to 3, for field measurement, can be considered acceptable if the error is not systematic (Klepper and den Hollander, 1999). By this criterion, the model RICEWQ can be useful to predict occurrence and size of paddy runoff contamination events.

Analysis of Variance of the Model Uncertainty
The uncertainty indicators (IPEU, MEU, NVU, and TU) have been estimated as explained above by means of ANOVA. These indicators were calculated using the mean of three dependent (output) variables 1REAT, CUM, and TW21: 21.27, 47.32, and 10.31 µg L–1, respectively. The IPEU indicator had values of 44, 46, and 35% for 1REAT, CUM, and TW21, respectively; MEU ranged from 3 to 4%; NVU had values of 49, 206, and 294% for 1REAT, CUM, and TW21, respectively. Both IPEUs and MEUs are independent from the output variable. However, NVUs are dependent on the output variable chosen. The TU indicator had values of 93, 74, and 121% for 1REAT, CUM, and TW21 respectively, confirming that the model uncertainty is reasonable.

Relationship of Input Variables to the Uncertainty
Results of stepwise regressions ({alpha} = 0.05) are reported in Tables 5, 6, and 7. The variables always positively correlated with the three output variables were application rate and minimum water paddy depth in June. The variables always negatively correlated were Kd, mixing depth of sediment depth to allow pesticide direct partitioning to bed (VBIND) (except for 1REAT), pesticide application time, pesticide application efficiency, and maximum crop coverage.


View this table:
[in this window]
[in a new window]
  		Table 5. Signs of the coefficient of the stepwise regressions of chemical concentration in first paddy water runoff (1REAT) (3600 runs).{dagger}

 

View this table:
[in this window]
[in a new window]
  		Table 6. Signs of the coefficient of the stepwise regressions of cumulative chemical concentration in paddy runoff after a 21-day treatment period (CUM) (3600 runs).{dagger}

 

View this table:
[in this window]
[in a new window]
  		Table 7. Signs of the coefficient of the stepwise regressions of time-weighted chemical concentration in water runoff over a 21-day treatment period (TW21) (3600 runs).{dagger}

 
Variables that increase pesticide concentration in paddy water will increase concentrations in runoff, and vice versa; VBIND is closely related to pesticide percolation and dilution. With higher values of VBIND, more pesticide is mixed with sediment, so the pesticide concentration in runoff will be lower. The negative relationship between Kd and chemical runoff concentration is supported by previous research (Tiktak, 1999; Capri et al., 2001).

Application rate, application efficiency, and the pesticide concentration in water outlet are positively correlated because the former increases the pesticide concentration in paddy water. Similarly, crop coverage affects the quantity of pesticide arriving in the paddy water, that is, when crop cover increases, the concentration in the water is lower. As a consequence, these variables have a strong influence not only on the runoff prediction, but also on the uncertainty of simulation.

For all three output variables, the minimum water paddy depth in June is positively correlated with the concentration in runoff, because the probability of runoff is higher if irrigation is applied to the field at higher initial paddy depth. For CUM, this relationship maintains in July and August.

The standardized regression coefficient values are used to illustrate the contribution of each parameter to the prediction uncertainty. As shown before (Tables 5, 6, and 7) the number of parameters significantly related with the three output variables are in the order: CUM > 1REAT > TW21. Perhaps the cumulative variable CUM also accumulates uncertainty, while 1REAT does so occasionally, and TW21 lowers the uncertainty by means of time-weighted averaging.

Figure 5 shows the contribution of each parameter to pesticide runoff uncertainty. Application time and Kd have the highest contribution, each more than 10% of the total uncertainty. For TW21, except for Kd and application time, the parameters including wash off rate and VBIND contribute more than 10% to model uncertainty as well.



View larger version (29K):
[in this window]
[in a new window]
  		Fig. 5. The contribution of process input parameters (%) to the prediction uncertainty of each modeling output variables (3600 runs): (a) chemical concentration in first paddy water runoff (IREAT), (b) cumulative chemical concentration in paddy runoff over a 21-day treatment period, and (c) time-weighted chemical concentration in water runoff over a 21-day treatment period (TW21). See Table 2 for input variable definitions.

 
For each output variable, the magnitude of the uncertainty due to input parameters is different, but it is possible to group some parameters totaling almost 55% of the uncertainty. They are spatial (including Kd and VBIND) and management parameters (such as time and rate of pesticide application and minimum water paddy depth). Therefore, these crucial parameters, particularly for Kd and VBIND, should be carefully parameterized or experimentally determined before the use of the model, since error in these parameters could produce large uncertainty of model outputs (Miao et al., 2003c; Capri and Miao, 2002). Moreover, the runoff output variables (e.g., 1REAT, CUM, and TW21) should be carefully selected. For example, when the goal of the simulation is to quantify event-based uncertainty, 1REAT should be selected. In summary, model theory and parameter errors strongly contribute to overall output uncertainty. This agrees with a previous report, which suggested that parameter errors were the strongest factors of the whole model uncertainty (Hession and Storm, 2000).


   CONCLUSIONS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
REFERENCES
 
The RICEWQ model has been shown to be useful for forecasting pesticide runoff in paddy fields. A higher-tier pesticide exposure assessment of surface water is being proposed for assessing the RICEWQ model at field scale in northern Italy. The assessment represents a realistic worst-case scenario of pesticide impact, the predicted runoff concentration is in agreement with the field measurement, and the model can be applied to simulate pesticide fate in site-specific paddy runoff.

Uncertainty analysis is an indispensable procedure to evaluate model accuracy. The uncertainty of the improved RICEWQ model was investigated using Monte Carlo techniques and a series of statistical methods including ANOVA and stepwise multiple regression. These methods were implemented for the identification of the reliability of model predictions and screening of crucial process variables. The results demonstrated that the paddy runoff concentration predicted by RICEWQ was in agreement with the field measurement, and the model can be applied to simulate pesticide fate at paddy-field level.

The results illustrated that the RICEWQ runoff predictions conform to a log-normal distribution with a short right tail, the model uncertainty is acceptable, and the model estimation is reliable at field scale due to its narrow spread of the uncertainty distribution and the center position of the measured runoff values in the prediction uncertainty distribution. For the different runoff output variables (i.e., 1REAT, CUM, and TW21), the contribution of input parameters to the prediction uncertainty is different, so the runoff output variables must to be carefully selected before using the model. Meanwhile, the contribution of parameters to prediction uncertainty is varied. Spatial parameters, including Kd and VBIND, together with management parameters like time and rate of application and climatic conditions, are the main contributors to the uncertainty. Therefore, these crucial parameters should be carefully parameterized or experimentally determined before application of the model, since small errors in these parameters may produce large uncertainty of model outputs.


   ACKNOWLEDGMENTS
 
The work was supported by the Italian national project MIUR (COFIN 2000), "Herbicide Fate in Paddy Field." This publication was financed by Università Cattolica del Sacro Cuore for the quality of the results obtained (esercizio 2004).


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


Related articles in JEQ:

This Issue in Journal of Environmental Quality

JEQ 2004 33: 1947-1953. [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
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 (6)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Miao, Z.
Right arrow Articles by Del Re, A. A. M.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Miao, Z.
Right arrow Articles by Del Re, A. A. M.
Agricola
Right arrow Articles by Miao, Z.
Right arrow Articles by Del Re, A. A. M.
Related Collections
Right arrow Multicomponent Transport Models
Right arrow Water Flow Models
Right arrow Runoff
Right arrow Agricultural Pesticides
Right arrow Water Pollution

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
The SCI Journals Agronomy Journal Crop Science
Vadose Zone Journal Journal of Plant Registrations
Journal of Natural Resources
and Life Sciences Education		 Soil Science Society of America Journal