Preliminary Results

Presented at the North American Water and Environmental Congress '96

American Society of Engineers
Anaheim, California
June 22-23, 1996

Neural Networks Predict Pesticide Leaching From Turfgrass Covered Areas


By Steven K. Starrett1, Associate Member, ASCE, Shelli K. Starrett2, Yacoub M. Najjar3, Associate Member, ASCE, and Judy C. Hill4, Student Member, ASCE
1Res. Asst. Prof., Dept. of Civ. Engrg. Dept. ; 2Asst. Prof., Dept. of Elec. Engrg. and Comp. Engrg.; 3Asst. Prof., Dept. of Civ. Engrg.; 4Undergrad. Honors Student, Dept. of Civ. Engrg., Kansas State Univ., Manhattan, KS 66506.

Abstract

The goal of this work was to determine whether artificial neural networks (ANN) can be used to predict the percentage of applied pesticide that leaches through 50 cm of turfgrass covered soil. One- hundred seventy-five sets of data used to train the ANN. After training, a simulation of 25 test situations were run. We were pleased with the ANN predictions for this initial study. The ANN appears to be reasonably able to predict the percentage of pesticide that leached through 50 cm of turfgrass covered soil. Future work consists of incorporating additional data sets into the training and testing process of this ANN.

Introduction

Turfgrass managers apply pesticides on golf courses, home lawns, sports complexes, industrial parks, and other areas to improve turf quality. The public's increased concern for the environment has increased attention on the environmental effects of chemical applications to turfgrass areas. Possible adverse environmental affects of pesticides are numerous: pesticides are potentially harmful to humans, may reduce certain bird populations, can destroy nontarget organisms, and may elevate non-pest species to pest status (Balogh and Walker, 1992). Limited research has been conducted concerning the fate of pesticides applied to turfgrasses. The United State Golf Association, in 1990, started a research program that focused on environmental issues related to the golf industry. Published papers describing research results from this program are beginning to be printed (for example, Starrett et al., 1995a, 1995b, 1996a, and 1996b, Horst et al., 1996). The understanding of the transport of pesticides on golf courses will be greatly advanced due to this recent surge of funding from the USGA.

Concentrated efforts have been made in recent years to describe water and pesticide movement in agricultural fields. Consequently, a variety of new models have been proposed that vary widely in their conceptual approach and degree of complexity. These models are influenced by the environment, training, and biases of their developers. PRZM/RUSTIC, EXAMS, SWRRB, CREAMS/GLEAMS, SURFACE, HSPF/STREAM, and GUS (Gustafson, 1993) are some of the available models that attempt to predict the movement of pesticides in the environment or address the likelihood of pesticides to contaminant drinking water supplies. Unfortunately, as the degree of modeling sophistication has increased so has the number of the associated parameters required to be evaluated from the experimental data. A high level of knowledge is required in order to understand and implement these models in any predictive technique. Also, these models were developed for the agricultural industry making it difficult for the models to simulate golf course conditions.

The network topology and the form of the rules and functions can all be variable in an ANN. As a result, this variation leads to a wide variety of network types such as competitive learning, the Hopfield network, and the back-propagation network (Fausett, 1994; Simpson, 1991; Werbos, 1994). Despite this wide variety, back-propagation networks are currently being used by most neural-network application developers (Abdul Aziz and Wong, 1992; Basheer and Najjar, 1994a; Basheer et. al., 1994a; Chao and Skibniewski, 1994; Flood and Kartam, 1994a and 1994b; Gagarin et. al. , 1994; Ghaboussi et. al., 1994; Ghaboussi et. al., 1991; Ghaboussi et. al., 1991; Goh, 1994; Garret et. al., 1992a; Garret et al., 1992b, Karunanithi et. al., 1994; Lee and Sterling, 1992; Najjar et. al., 1994; Penumadu, 1993; Rizzo and Dougherty, 1994; Rogers, 1994; Rogers and Dowula, 1994; Wang and Feng, 1993).

The goal of this work was to determine whether Artificial Neural Networks (ANN) can be used to predict the percentage of pesticide that leaches through 50 cm of turfgrass covered soil. The benefits of developing an accurate computer simulation model are: its ability to evaluate alternate management practices, its assistance in the selection of alternate pesticides, and its assistance in optimizing irrigation practices to meet turfgrass and environmental quality goals (Balogh and Walker, 1992).

Research Methods

The Neural Network Toolbox of the computer software package MATLAB was used for this work. A step-by-step description of the back- propagating learning algorithm is given in numerous publications (for example, Basheer and Najjar, 1994a and 1994b; Basheer et. al., 1994a and 1994b; Najjar et. al., 1994). A brief description of ANNs follow.

The processing units in a back-propagating ANN are arranged in layers. Each ANN has an input layer, an output layer and a minimum of one hidden layer (Fig. 1). The presence of the hidden units permits the ANN to form a mapping scheme from the input units to the output units. In this ANN, activation propagation direction takes place in a feed- forward manner, from the input layer through the hidden layer to the output layer. On the other hand, the process of learning is achieved by back-propagating the errors at the output layer through the hidden layer(s) to the input layer. No communication is permitted between the processing units within a layer, however, the processing units in each layer may send their output to the processing units in higher layers. It is the presence of the hidden units (three for this model) that allows this ANN to represent and compute the complicated association between the input and output patterns.


Figure 1. Example Artificial Neural Network (ANN) with backpropagation.


The connections between the various nodes in the three-layer paradigm represent the most important part of the computational ANN. Associated with each connection is a numerical value representing the strength or the weight of that connection. The connection strengths are developed during the training process of the ANN. At the beginning of the training process, the connection strengths are assigned random values. As inputs-outputs are presented during the training, the adopted rule of learning modifies the connection strength in an iterative process. When the iterative process has converged, the weights thus learned are stored for later use by the ANN. As a result, the ANN will be able to utilize the stored weights to quickly evaluate new sets of inputs when presented. A step-by-step summary of the back-propagating learning algorithm is given in Basheer and Najjar, 1994 and 1996.

Data collected from research conducted by Dr. Nick Christians, Dr. Al Austin, (Iowa State University) and Steven K. Starrett that investigated the transport of pesticides applied to turfgrass was used to train our ANN. Some general characteristics of the seven pesticides used are presented in Table 1.



Table 1. Properties of pesticides.

Pesticide Water Solubility (mg L-1) Kow Koc Field Half-Life (days)
metalaxyl 8400.0 50.0 50 70
isazofos 69.0 1000.0 100 34
chlorpyrifos 2.0 1000000.0 6070 30
pendimethalin 0.3 1500000.0 5000 90
2,4-D 796000.0 0.4 20 10
MCPP 0.3 0.3 20 21
dicamba 400000.0 0.2 2 14
(Wauchope, et al., 1992)
Kow = octanol:water partitioning coefficient.
Koc = soil:water partitioning coefficient normalized by the organic carbon fraction of soil.


Initially the following inputs were used: the pesticide's solubility, Koc, Kow, half-life and rate of application values; the soil's organic matter, bulk density, and pH; the cumulative irrigation values, the irrigation application practice (heavy or light applications), and the time after the pesticide application that the leachate samples were collected. The importance of the various inputs were determined. Many inputs had no affect on the predicted outputs. After much investigation, the number of inputs used to train the ANN was reduced to the following five: pesticide solubility, Kow and rate of application values, time after application, and the irrigation application practice. One-hundred seventy-five sets of data were used to train the ANN. After training, a simulation of 25 test cases were run.

Results

The 25 data sets used for testing of the ANN were not included in the training set. The 25 sets of data were given to the trained ANN and a prediction of the percentage of pesticide that leached through 50 cm of turfgrass covered soil was made. Test cases 1-20 are from the earlier mentioned research conducted at Iowa State University. To determine how well the ANN could make predictions on data from other test situations, we also used data from a study conducted by Drs. Cisar and Snider from Florida (cases 21-25). A comparison was made between the measured values and the predicted values (Fig. 2 and Fig. 3).


Fig. 2. A comparison between measured values and ANN predicted
values for the percentage of pesticide that leached through 50 cm of
turfgrass covered soil for test situations 1-13.



Fig. 3. A comparison between measured values and ANN predicted
values for the percentage of pesticide that leached through 50 cm of
turfgrass covered soil for test situations 14-25.



Discussion and Conclusions

Several of the initial inputs proved to be not significant contributors in the training of the ANN. Possible explanations are that these inputs varied little between training cases and, therefore, it was difficult for the ANN to develop relationships based on such small input deviations. Also, several of the input characteristics are difficult to determine accurately, therefore, discrepancies between changes in inputs and their effect on the outputs were possible. Half-life of particular pesticides, for example, is practically impossible to generalize since it depends on the specific environmental conditions. Redundancies in input data could also cause some inputs to be unneeded. For example. time after application and irrigation values are closely linked due to the use of a fixed irrigation schedule. It is anticipated that with the addition of new data sets some previously unimportant inputs may need to be re- included.

We were pleased with the ANN predictions for this initial study (Fig. 2 and Fig. 3). The ANN appears to be able to reasonably predict the percentage of pesticide that leached through 50 cm of turfgrass covered soil. Future work will consist of incorporating additional data sets into the training and testing process of this ANN.


Acknowledgments

The authors would like to thank the Green Section of the United States Golf Association for their financial support.

References

Balogh, J.C., and W.J. Walker. 1992. Golf course management & construction environmental issues. Chelsea, Michigan: Lewis Pub.

Basheer I.A. and Najjar Y.M. (1994), "Designing and Analyzing Fixed- Bed Adsorption Systems With Artificial Neural Networks", J. of Env. Systems, 24:3, 291-312.

Basheer, I.A., L.N. Reddi, and Y.M. Najjar, 1996, " Site Characterization Using Neuronets: An application to the Landfill Siting Problem," Ground Water Journal, Sept./Oct. 1996.

Horst, G.L, P.J. Shea, N.E. Christians, D.R. Miller, C.L. Stuefer-Powell, and S.K. Starrett. 1996. Pesticide Dissipation Under Golf Course Fairway Conditions. In press. Crop Sci.

Starrett, S.K., N.E. Christians, and T.Al Austin. 1995. Fate of 15N Amended Urea in a Turfgrass Biosystems. Comm. in Soil Sci. and Plant Anal. 26:1595-1606.

Starrett, S.K., N.E. Christians, and T.Al Austin. 1995. Fate Of Nitrogen Applied to Turfgrass Covered Soil Columns. 121(6):390-395. J. of Irrigation and Drainage Eng.

Starrett, S.K., N.E. Christians, and T.Al Austin. 1996. Comparing Dispersivities And Soil Chloride Concentrations Of Turfgrass Covered Undisturbed And Disturbed Soil Columns. In press. J. of Hydrology.

Starrett, S.K., N.E. Christians, and T.Al Austin. 1996. Fate Of Isazofos, Chlorpyrifos, Metalaxyl, And Pendimethalin Applied To Turfgrass Covered Undisturbed Soil Columns. In Press. J. of Environmental Quality.

Wauchope, R.D., T.M. Buttler, A.G. Hornsby, P.W.M. Augustijn-Beckers, and J.P. Burt. 1992. The SCS/ARS/CES pesticide properties database for environmental decision-making. In G. W. Ware, (ed.) Reviews of environmental contamination and toxicology. Springer- Verlag, New York. N.Y. 123:1-35.




Last Modified December 16, 1996