4Meteorological Service of Canada, Environment Canada, Toronto, Ont., Canada

Abstract: A long term pesticide monitoring program was carried out in paired sub-watersheds near Kintore Watershed, Ontario, Canada between 1988 and 1991 to determine the impact of agricultural conservation practices on pesticide delivery to surface water. Three types of water sampling schemes including grab, storm event, and continuous water sampler were employed in this study. A total of 608 water samples were analyzed for concentrations of atrazine. Atrazine concentrations exceeded national guidelines in both watersheds during runoff events. Atrazine concentrations achieved higher peak values in the control watershed. Two different statistical approaches, multiple linear regression (MLR) and the artificial neural network (ANN) model were used to estimate the runoff losses of pesticide. Four important factors were selected as inputs for the MLR and ANN models. These factors include pesticide residues, daily rainfall, soil temperature, and soil moisture conditions. An attempt is made to explain differences in stream concentrations of atrazine under different weather scenarios by examining the four independent factors. The model results were evaluated with monitoring data show the good predictive ability of ANN model with R2=0.90 and RMS=0.83 in comparison with that of MLR model with R2=0.75 and RMS=1.3.

 

Keywords: agriculture, environment, modeling, monitoring, nonpoint source, pesticide, runoff, watershed

The Ecosystem Health Division of Environment Canada and the Upper Thames River

Conservation Authority were involved in a study of a paired watershed of Kintore Creek to examine the effects of farm conservation practices on pesticide transport to surface water between 1988 and 1991. Two adjacent sub-watershed in Kintore Creek area, near Thamesford, Ontario (Figure 1) were selected for monitoring runoff losses of atrazine, which is one of most heavily used pesticide in this area. The sub-watersheds are equal in size, adjacent to one another, and have similar highly erodible landscapes, and cropping patterns. Detailed Kintore Creek sub-watershed characteristics are listed in Table 1. Both sub-watersheds originate in swampy headlands that provide a year round source of water. Kintore Creek flows into the Middle Branch of the Thames River, which drains the corn belt of Ontario, before discharging into Lake St. Clair. Further details of the watershed project may be found in Merkley and Glasman (1984).

The primary goal of this study has been to determine the impact of agricultural conservation practices, including erosion control structures, tillage, and cover crops, on the delivery of pesticides to surface water. There are many other factors may also influence pesticide losses from agricultural fields based on the complex fate process of pesticide, such as weather conditions, soil characteristics, application rates and modes, crop cover etc. (Zhang et al 1997; DeLaune et al 1997; Pantone et al 1992).

The quantitative relationship between runoff concentration of pesticide and a knowledge of one or several of the above factors could be used to calculate and predict the runoff losses of pesticides from agricultural fields (Wauchope et al., 1992; Cohen et al., 1995). In this study, two different statistical approaches, multiple linear regression (MLR) and the artificial neural network (ANN) model were used to predict the runoff losses of pesticide from agricultural soils based on these important factors which include pesticide residues, daily rainfall, soil temperature, and soil moisture conditions. An attempt is made to explain differences in stream concentrations of pesticide under different weather scenarios by examining the four independent factors.

                Table 1    Kintore Creek sub-watersheds characteristics

Fig. 1    Locations of the Kintore Creek and Thames River Watersheds

Research approach employed in this study may be divided into two broad categories: data collection and statistical modeling. Data collection involved water sampling to establish pesticide concentrations and loads as well as a pesticide usage questionnaire to determine pesticide application rates and locations.

Statistical modeling included using two models, MLR and ANN to estimate the pesticide runoff losses. Four important factors were selected as inputs for the MLR and ANN models. These factors include pesticide residues, daily rainfall, soil temperature, and soil moisture conditions.

(1) Sampling Methods

Runoff events were monitored using ISCO model 2100 automatic water samplers. ISCO model 1640 liquid level sampler actuators were set to initiate water sampling during storm events. Samples were collected at one-hour intervals for the first five samples; the remaining three were collected at two-hour intervals. Each event therefore, was monitored for a total of eight hours. Repeated equipment failures prevented the monitoring of every runoff event, however, it is believed the samples that were obtained provide an accurate reflection of pesticide concentrations and loads.

Modified Quality Environment Automatic Liquid Samplers were utilized as the third method of collecting water quality information. These flow proportional samplers provided a composite weekly water sample by continuously collecting a small quantity of water over a seven day period. Average weekly atrazine concentrations were then combined with weekly discharge data to determine loadings.

All samples were collected in 350ml glass bottles, transported from the field in coolers, and refrigerated at Fanshawe Conservation Area until delivered to the lab at the Canada Centre for Inland Waters (CCIW), Burlington. Analysis of samples for atrazine concentrations was carried out at the CCIW lab and results were returned to both the project scientific authority and staff at the U.T.R.CA. Only results from runoff event and base flow monitoring were included in this analysis. Continuous sampler data has been reported elsewhere.

(2) Pesticide Usage Survey

Two different statistical approaches, multiple linear regression (MLR) and the artificial neural network (ANN) model were used in this study to predict the runoff losses of pesticide. The aim is to establish a quantitative relationship between the runoff concentration of pesticide and the factors determining the runoff losses of pesticide from agricultural fields.

Runoff losses of pesticide may be influenced by many factors that potentially interact in a complex manner (Leonard, 1988). Those factors include climatic conditions, soil characteristics, pesticide properties, application rate and methods, and agricultural managements, etc. In this study, four independent factors, including daily rainfall, mean temperature, soil moisture condition, and atrazine residues are selected as input parameters for models. The tillage practice is not considered here since the influence of the practice is not significant (Burgoa B. et al., 1995). The outputs of models were the mean pesticide concentration in runoff.

(1) Multiple Regression Model

                         (1)

is a multiple linear regression model with k independent variables, where b _{0, b 1} and b _k are called regression coefficients, while x₁, x₂ and x_k are called regression variables and e is the error (Bethea, et al., 1984). The estimation of the regression coefficients is done by minimizing the sum square error SSE.

(2) Artificial Neural Network Model

Many kinds of ANNs have been suggested, but the most commonly used artificial neural network is Back Propagation Neural Network (BPNN) (Rumelhart et al., 1986). In this study, the BPNN method is selected to calculate the runoff concentration of atrazine. The BPNN is a multiplayer feed forward neural network architecture that accepts continuous or discrete inputs and emit continuous valued outputs. Data are fed into the input neurons which transfer the data to one or more hidden layers (internal layers of neurons). Each of the signals from the input neurons is multiplied by a weighting factor and these weighted signals are accumulated by the hidden neurons. A transfer function is applied to these accumulated signals and the hidden neurons distribute their signal to output neurons, again a weighting factor is applied to each hidden neuron signal. The output neurons take the accumulated signals from the hidden neurons, apply a transfer function and emit the final outputs.

Input data and desired output values are provided to train the BPNN. Mean squared errors are calculated from the outputs that the neural network calculated from the input data and the desired values. These errors are back propagated through the network and used to adjust the weights. This training process is repeated over and over until the weights converge to a final value.

The input layer contains four neurons which receive four input signals of pesticide residues (x₁), daily rainfall (x₂), daily mean temperature (x₃), and soil moisture (x₄). The output layer contains one neuron which produces a corresponding output, daily mean runoff concentration of atrazine (y), for a given input vector (x₁, x₂, x₃, x₄).

The average of eight measured values from each event was used in MLR and BPNN models. Corresponding daily data of rainfall, mean temperature, were collected from the weather station in the City of London. Soil moisture was calculated as cumulated rainfall during five days before the rainfall event occurred (Ward, et al., 1995). Residues of atrazine were calculated by equation (Li, et al., 2000a)

                                  (2)

where R_t is the residue of pesticide in time t (days) after pesticide application; U is the total application of pesticide; F_t is the daily emission factor of pesticide; t_1/2 is the half-life of pesticide in the soil. For atrazine t_1/2 is approximately equal to 60 days (Mackay et al., 1997). The residues left in soils due to the use of atrazine in previous years are not considered here. The daily emission factors F_t are calculated from the annual emission factors from Scholtz et al. (1997).

The predictive results obtained from MLR and BPNN were compared with observed data and evaluated by following two statistical parameters, the Root-Mean-square error (RMS) and R square (R²). RMS quantifies the general difference between predictive results and measured data, and R² shows the agreement between predictive results and measured data. RMS and R² were calculated by

                                    (3)

                                         (4)

where SSE is the sum square error, n is number of data sets, p is number of terms in the model, and term n-p indicates the Error Degrees of Freedom. S_yy, the Total Sum of Squares, represents the total variability in the n observations as the sum of the squared differences between the responses y_i (k=1…n) and the average of all responses, .

Current provincial water quality guidelines for freshwater aquatic life have been set at five micrograms per litre (5 ug\l) for atrazine (Canadian Council of Resource and Environmental Ministers, 1987). Atrazine concentrations from bi-weekly grab samples and runoff events were reviewed within the context of these guidelines. Table 3 shows the number of samples that exceeded the Canadian Water Quality Guideline at two watershed from 1988-1991 by three sample methods. Storm event samples mostly exceeded the guideline with 46.9% of samples. Continuous samples and grab samples exceeded the guideline with 6.5% and 3.1% of samples respectively.

Both runoff event and grab samples were combined with hourly stream discharge data and calculated through the Beale Ratio Estimator to arrive at yearly atrazine loadings for each watershed. Table 4 summarizes loadings, estimate errors and significance tests for each of the three years.

Atrazine loads generally increased for both watersheds during the study period. Loadings were also significantly higher in the control than demonstration watershed for each year. The ratio of demonstration to control loadings (Table 4) suggests that, on average, loadings in the demonstration basin are less than the control, however, during 1990, the demonstration’s load exceeded the control’s. Atrazine loading results, then, support the notion that the control watershed has conditions more conducive for the delivery of atrazine to surface waters than the demonstration’s basin.

Figure 2a-d presents the first four sampled runoff events for 1988, which occurred between July 22 and Oct 2. It is obvious that decrease in atrazine concentration and calculated load with each successive rainfall runoff event, with the exception of the second runoff event on July 28 which had the greatest atrazine loadings because of the high concentration. The concentration peaks in four runoff events are all lags behind the discharge peaks by 50~15 minutes. It may be explained that the delay existed in adsorption-desorption processes when pesticide in the soil enter in runoff

Table 3    The number of Atrazine samples exceeded the Canadian Water Quality Guideline (Drinking Water: 60 m g/L, Protection of Aquatic life: 2 m g/L) at two sub-watersheds in 1988-1991

Input-output pair sets can be constructed for every storm event. The input values include daily rainfall, daily mean temperature, soil moisture, and atrazine residues when the storm started. Data of daily rainfall and mean temperature were obtained from the Canadian National Climate Data Archive Database (GRP, 2000). The soil moisture was calculated by cumulating five days rainfall before storm event, and the atrazine residues were calculated from Equation (2). The output values are the concentration of atrazine, which was calculated as a mean value of eight storm samples collected during every storm event. There are 40 input-output pair sets between 1988 and 1991 in total.

Since the performance of BPNN and MLR models is based on the quality and quantity of training data and fitting data, respectively, different data used for training and fitting processes will produce different results from the models. In this study, 30 input-output pairs were randomly selected from the total 40 input-output pairs, and used as the training datasets for BPNN model and the fitting datasets for the MLR model. The rest 10 input-output pairs have been used as the test set for both models.

               Table 4    Pesticide (Atrazine) Loading

Fig. 2a    Runoff discharges, water concentrations, and loads of atrazine from the east sub-watersheds on July 22 (Day 204), 1988

Fig. 2b    Runoff discharges, water concentrations, and loads of atrazine from the east sub-watersheds on July 27 (Day 209), 1988

Fig. 2c    Runoff discharges, water concentrations, and loads of atrazine from the east sub-watersheds on August 10 (Day 223), 1988

Fig. 2d    Runoff discharges, water concentrations, and loads of atrazine from the east sub-watersheds on October 2 (Day 276), 1988

(1) The performance of the MLR and BPNN Models

After a fitting process by using 30 input-output data sets, an equation relating the input data and the output data has been derived from the MLR model as

Y_p = –2.042 + 27.93X₁ + 0.00652X₂ + 0.00145X₃ + 0.00168X₄                                           (5)

where Y_p is the predicted daily mean runoff concentration of pesticide (ug/L), X₁ is the residue of pesticide (t), X₂ is the daily rainfall (mm), X₃ is the daily mean temperature (° C), X₄ is the Soil moisture (mm). The predicted daily mean runoff concentration from Equation (5) and the monitoring data are correlated with R² = 0.626 and RMS =1.616.

To improve the quality of the relation, quadratic terms were added into the model and Equation (5) becomes

Y_p = 2.001 + 114.2X₁² + 6.914´ X₂² – 0.02343X₂                                 (6)

In Equation (6), less important parameters, daily mean temperature (X₃) and soil moisture (X₄) are disappeared. The predicted daily mean runoff concentration from Equation (11) and the monitoring data are correlated with R² = 0. 746 and RMS =1. 305. The data are shown in Figure 3 along with the least square line

Fig. 3   Observed concentration of atrazine as a function of predicted values using MLR model

Y_obs = Y_p                                    (7)

where Y_p is the predicted mean concentration of atrazine and Y_obs is the observed mean concentration of atrazine. Ten input-out put pairs were used to test the performance of the MLR model. The results are shown in Table 8. The average, maximum and minimum absolute percent errors are found to be 27.93%, 79.18% and 5.1% for the MLR model.

(b) BPNN model

It has been reported that the number of hidden layers and the number of neurons in the hidden layer significantly affects the performance of the neural network for a given input-output space (Li et al., 1993). In this study, various numbers of hidden layers and neurons in the hidden layer were tested in the BPNN model and eight neurons in the hidden layer were finally chosen for the model.

To compare the BPNN with MLR models, the same 30 input-out pairs used as the fitting data sets for the MLR model were applied as training datasets for BPNN model. The predicted daily mean runoff concentration and the monitoring data are correlated with R² = 0.90 and RMS = 0.83. The least square equation is given

Y_obs = 0.90983 Y_p + 0.45822                               (8)

where Y_p is the predicted mean concentration of atrazine (ug/L) from the BPNN model, and Y_obs is the observed mean concentration of atrazine.

Figure 4 is a graph of Equation (8) superimposed over the data of the observed mean runoff concentration of atrazine and the predicted mean runoff concentration of atrazine from the BPNN model. The same 10 input-output pairs were used to test the performance of BPNN model, and the results are shown in Table 5. The average, maximum and minimum absolute percent errors were found to be 18.44%, 46.72% and 0.08%, respectively. The results show that the BPNN model in this study has a lower absolute percent error in predicting runoff concentration of pesticide than that from MLR.

Fig. 4    Observed concentration of atrazine as a function of predicted values using BPNN model

A long term pesticide monitoring program was carried out in paired sub-watersheds near Kintore Watershed, Ontario, Canada between 1988 and 1991 to determine the impact of agricultural conservation practices on pesticide delivery to surface water. A total of 608 water samples were analyzed for concentrations of atrazine. Atrazine concentrations exceeded national guidelines in both watersheds during runoff events. Atrazine concentrations achieved higher peak values in the control watershed.

Two statistical approaches, artificial neural network (BPNN model) and multiply linear regression (MLR) were compared in order to predict the runoff concentration of pesticide occurred in streams. The results show that BPNN model is more satisfactory than MLR based on the higher R² value and lower RMS. BPNN model can offer more advantages over the MLR and other conventional methods because it is not necessary to specify the form of the mathematical model before fitting the data. This is important when the appropriate equation is difficult to find due to the complexity exist in physicochemical process of runoff pesticide losses.