AWRA SYMPOSIUM ON GIS AND WATER RESOURCES
Sept 22-26, 1996
Ft. Lauderdale, FL

USING GIS AND REGRESSION TO ESTIMATE ANNUAL HERBICIDE CONCENTRATIONS IN OUTFLOW FROM RESERVOIRS IN THE MIDWESTERN USA, 1992-93

William A. Battaglin(1) and Donald A. Goolsby(2)

Table of Contents

INTRODUCTION

Background

Reservoirs are a prominent feature of the hydrology of river systems in the midwestern United States (Midwest), where they are used to impound and store water for public water supply, flood control, irrigation, recreation, hydropower, and wildlife habitat. There is no recent count of the number of reservoirs in the United States, but a 1982 inventory of artificial barriers that are more than 25 feet high or that impound more than 50 acre-feet of water, listed approximately 66,000 structures (U.S. Army Corps of Engineers, 1982). Reservoirs may also store undesirable substances such as agricultural chemicals (herbicides, insecticides, other pesticides, and nutrients) and sediment. Most sediment entering a reservoir is trapped behind the dam. However, soluble chemicals including some herbicides remain in the dissolved phase in water and are stored in the reservoir until they are discharged or are removed by biotic or abiotic processes.

Agricultural activities in the Midwest have resulted in the contamination of surface water with agricultural chemicals. Numerous recent investigations (Goolsby and Battaglin, 1995; Schottler et al., 1994; Baker and Richards, 1990) indicate that significant quantities of some herbicides are flushed from cropland to streams each spring and summer during rainfall events following application. Peak concentration of several herbicides can exceed 10 ug/L during these events (Coupe et al., 1995; Scribner et al., 1994).

Only a few studies have investigated the herbicide content of reservoirs or the way in which physical, hydrologic, and land use characteristics of reservoir drainage basins interact to affect herbicide concentrations in reservoir outflow. Preliminary results from this study (Scribner et al., 1996; Goolsby et al., 1993) and results from other investigations (Kalkhoff, 1993; Leung et al., 1982; Miller and Kennedy, 1995; Stamer et al., 1995) indicated that reservoirs and rivers regulated by reservoir outflows are likely to have lower peak herbicide concentrations but a longer period of elevated herbicide concentration than unregulated rivers. The implication is that water in or released from reservoirs is likely to have a higher time-weighted annual mean herbicide concentration than water in unregulated rivers. Time-weighted annual mean concentrations are used to determine compliance with U.S. EPA maximum contaminant levels (MCLs) and lifetime health advisory levels (HALs) for herbicides in drinking water supplies (U.S. Environmental Protection Agency, 1989; 1995).

Objective

The objective of this research is to develop models for the annual mean concentration of selected herbicides in midwestern reservoirs. To accomplish this objective, water quality data are analyzed both statistically and graphically. A geographic information system (GIS) is used to manage and quantify characteristics on land use, agricultural chemical use, climatic conditions, topographic character, and soil type within the reservoir drainage basins. Multiple linear and logistic regression equations are developed to estimate annual mean herbicide concentrations in reservoir outflow as a function of these land use, hydrologic, and physical characteristics.

Study Area

The study area consists of 76 reservoirs and reservoir drainage basins in parts of 11 midwestern States (figure 1). The two primary considerations in reservoir selection were (1) reservoir outflow was accessible for sampling, and (2) information on the volume and discharge of the reservoir was obtainable. Drainage basin boundaries for the 76 reservoirs were extracted primarily from hydrologic unit boundaries (Steeves and Nebert, 1994). Drainage basin areas range from 44 to 64,560 square kilometers (km2), and the mean basin area is 6,140 km2.

The study area encompasses a large region with considerable variability in land use, soil type, and climatic conditions. Harvested cropland represents more than 50 percent of the land area in about 25 percent of the basins and less than 5 percent of the land area in less than 5 percent of the basins. Soils in the eastern part of the study area tend to be finer textured and poorly drained, while soils in the western part tend to be coarser grained and well drained (U.S. Department of Agriculture, 1993).

Sample Collection and Analysis

Reservoir outflow was sampled eight times (approximately bimonthly) from April 1992 through September 1993 (table 1).



Samples were collected using non-contaminating and non-adsorbing pre-cleaned containers from near the centroid of flow or other outflow point that provided a representative sample of reservoir discharge. Prior to shipping to the lab for analysis, samples were filtered through 0.7 micron glass fiber filters and chilled (Scribner et al., 1996). The samples collected in this study are measures of water passing through or leaving the studied reservoirs and do not account for the variability of herbicide concentration within reservoirs.

All samples were analyzed for 11 herbicides (alachlor, ametryn, atrazine, cyanazine, metolachlor, metribuzin, prometon, prometryn, propazine, simazine, and terbutryn), 2 atrazine metabolites (desethylatrazine, desisopropylatrazine), and 3 cyanazine metabolites (deethylcyanazine, cyanazine amide, and deethylcyanazine amide) by gas chromatography/mass spectrometry (GC/MS). A metabolite of alachlor, alachlor ethane sulfonic acid (ESA), was analyzed by enzyme-linked immunosorbent assay after solid phase extraction (Scribner et al., 1996).

GIS Data

A GIS was used to manage and display the large quantity of spatial data in this investig ation. County-level estimates of agricultural chemical use, agricultural land use, crop acre age, and livestock were constructed by processing tabular data into GIS coverages (Battaglin and Goolsby, 1995). Other natural and anthropogenic variables defined within drainage basins from digital and non-digital data sources include basin area; population density (U.S. Department of Commerce, 1990); the soils variables hydrologic group, porosity, water-holding capacity, and permeability from STATSGO (U.S. Department of Agriculture, 1993); runoff (Rea and Cederstrand, 1994); and several hydrological parameters calculated for the TOPMODEL (Wolock, 1993) watershed model using digital elevation model and STATSGO data (Wolock, D.M., USGS, written commun., 1995). Area-weighted transfer and area-weighted sum algorithms, programmed in the GIS, were used to estimate the masses of agricultural chemical applied, acreages of crops, or numbers of people or livestock within drainage basins. In many cases, variable estimates were normalized by dividing by basin area. Explanatory variables that appear in models presented in this paper are listed in table 2.

Statistical Techniques

Multiple linear regression (MLR) and logistic regression (LGR) models were used to investigate relations between explanatory variables and the estimated annual mean concentrations of atrazine, cyanazine, and the total of 11 herbicides. Both MLR and LGR models provided satisfactory results and are presented in this paper.

MLR is a statistical technique that uses one or more explanatory variables to explain as much of the variation observed in the response variable as possible (Helsel and Hirsch, 1992). Once calibrated, MLR models can be used to estimate the response variable from knowledge of the explanatory variable values. MLR models with as many as six explanatory variables were tested. The best models generally had four or fewer explanatory variables. Models with more explanatory variables often had problems with autocorrelation among the explanatory variables, as indicated by variance inflation factor (VIF) values of greater than 10 (Helsel and Hirsch, 1992).

LGR is a statistical technique that uses one or more explanatory variables to predict the probability of a categorical response (Helsel and Hirsch, 1992). The response variable in LGR is the log of the odds ratio (logit). The logit transforms a variable (p) constrained between zero and one into a continuous variable that is linear with respect to the vector of the explanatory variables (equation 1):

Y = log [p/(1-p)] = b0 {b00} + b1X1 +... bkXk (1)

where: Y is the response variable; [p/(1-p)] is the odds ratio; b0 is the first intercept and b00 is the second intercept, X is the vector of k explanatory variables, b1 is the slope coefficient for the first explanatory variable X1; and bk is the slope coefficient for the kth explanatory variable Xk. LGR models presented here are comprised of two parallel linear equations. Models were compared based on fit statistics and on their ability to correctly estimate categories of annual mean herbicide concentration. The modeled category was selected as the one with the largest probability. Model accuracy is defined as the number of correct classifications (modeled concentration category matches observed category) divided by the number of attempted classifications.

RESULTS AND DISCUSSION

Herbicide Concentrations

Atrazine, its two metabolites, and alachlor ESA, were the most frequently detected herbicide compounds, all being detected in more than 60 percent of the samples. Cyanazine and metolachlor were also detected in about 50 percent of the samples. The highest herbicide concentration was for atrazine (12.4 ug/L) and the highest herbicide metabolite concentration was for ESA (19.7 ug/L).

A temporal pattern of herbicide concentration occurred in outflow from many of the 76 reservoirs. Like in streams and rivers (Goolsby and Battaglin, 1995) the highest herbicide concentrations occurred after the planting season in late spring and summer (sampling rounds 2 and 7). Herbicide metabolite concentrations occasionally peak later, in sampling rounds 3 and 8. The temporal patterns of alachlor, atrazine, cyanazine, metolachlor, desethylatrazine, and ESA concentrations in the outflow from all reservoirs are shown in
figure 2. Shown are concentrations in outflow from reservoirs with short residence times (Lake Vermillion, IL; 3 months), moderate residence times (Turtle Creek, KS; 1.5 months)and long residence times (Hillsdale Lake, KS; 8 months).

Annual Mean Concentrations

Estimates of the time-weighted annual mean herbicide concentration in reservoir outflow were calculated from the sample data. Annual mean herbicide concentration estimates are time-weighted not flow-weighted because time-weighted estimates are more representative of annual mean exposure via drinking water and are more relevant to water suppliers, who need to comply with federal regulations. For this report the eight sampling rounds were divided into two sets for the purposes of computing annual mean concentrations. The first set consists of samples collected in 1992 (rounds 1-4) and the second set consists of samples collected primarily in 1993 (rounds 5-8). Annual mean concentrations were estimated as the mean of the sample values. The resulting sampling frequency (quarterly) is probably inadequate to accurately estimate annual mean herbicide concentrations in reservoir outflow at some sites (Battaglin and Hay, 1996). Estimates of annual mean herbicide concentration are used to condition MLR and LGR models which can then be used to forecast expected annual mean herbicide concentrations from knowledge about the required explanatory variables.

Multiple Linear Regression Models

MLR was used to model the observed annual mean concentrations of atrazine, cyanazine, and the total of 11 herbicides in outflow from 76 midwestern reservoirs. Two of t he best models for atrazine, cyanazine, and total herbicide concentrations an d the coefficients of determination (R2) and variation (CV) for the m odels are listed in table 3.



A chemical use, chemical expense, or cropland variable is significant in every model, always having a direct relation to annual mean herbicide concentration. Subsurface contact time (CNCT) is significant in every model, always having a direct relation to annual mean herbicide concentration. Soil hydrologic group (HYGP) is significant in a several models, having a direct relation to atrazine and total herbicide concentration. Percentage of basin as pastureland (PSTP) is significant in several models, always having a direct relation to herbicide concentrations. TOPM is significant in all cyanazine models, having an inverse relation to annual mean concentration. Soil thickness (SLTK) is significant in two models and potential evapotranspiration rate (PET) and soil permeability (PERM) are each significant one model. All of the MLR models listed in table 3 have moderate predictive power and could only be expected to estimate annual mean herbicide concentrations values with a limited degree of confidence.

Logistic Regression

LGR was tested to determine if an improvement could be made over MLR models. LGR requires a discrete response variable; therefore, annual mean agricultural chemical concentration values were divided into three categories with roughly equal numbers of observations. The concentration categories and accuracy of two of the best models for atrazine, cyanazine, and total herbicide are listed in table 4.



The overall accuracy of the best LGR models, defined as the number of correct classifications (modeled concentration category matches observed category) divided by the number of attempted classifications, averaged about 76% (Table 4). LGR models are most effective at estimating when herbicide concentrations are in the lowest category (average accuracy of 86%), but had more difficulty predicting concentrations in the intermediate category (average accuracy of 67%), and concentrations in the highest category (average accuracy of 66%).

As with the MLR models, a chemical use, chemical expense, or cropland variable is significant in every model, having a direct relation to annual mean herbicide concentrations. CNCT is significant in several models, always having a direct relation to annual mean herbicide concentration. HYGP and PSTP area each significant in two models, both always having a direct relation herbicide concentration. PERM and TOPM are each significant in two models, both always having an inverse relation to herbicide concentrations. Annual mean temperature (TEMP) and SLTK are each significant in three models, both always having a direct relation to annual mean herbicide concentration.

Both MLR and LGR models helped to identify land use, chemical use, soil, and climatic variables in upstream drainage basins that affect annual mean herbicide concentrations in reservoir outflows. Results demonstrate a strong association between annual mean herbicide concentrations in reservoir outflow and the use of those herbicides within associated drainage basins. In many cases, models that use cropland acreage or agricultural chemical expenditure estimates work nearly as well as models using the harder to come by herbicide use estimates.

ACKNOWLEDGEMENTS

The authors thank James Fallon, Michael Meyer and John Flager for critical reviews of the manuscript; Gail Thelin and Naomo Nakagaki for providing raw Census of Agricultural data; and Laurie Boyer for html document preparation.

LITERATURE CITED

Baker, D. B. and R. P. Richards, 1990. Transport of soluble pesticides through drainage networks in large agricultural river basins. In: Long Range Transport of Pesticides: Lewis Publishers, Ann Arbor, MI. p 241-271.

Battaglin, W. A. and D. A. Goolsby, 1995. Spatial Data in Geographic Information System Format on Agricultural Chemical Use, Land Use, and Cropping Practices in the United States. U. S. Geol. Surv. Water-Resour. Invest. Rep. 94-4176.

Battaglin, W. A. and L. E. Hay, 1996, Effects of Sampling Strategies on Estimates of Annual Mean Herbicide Concentrations in Midwestern Rivers. Environ. Sci. Technol., 30(3), p 889-896.

Coupe, R. H., D. A. Goolsby, J. L. Iverson, S.D. Zaugg and D. J. Markovchick, 1995. Pesticide, Nutrient, Streamflow and Physical Property Data for the Mississippi River, and Major Tributaries, April 1991- September, 1992. U.S. Geol. Surv. Open-File Rep. 93-657.

Goolsby, D. A., W. A. Battaglin, J. D. Fallon, D. S. Aga, D. W. Kolpin and E. M. Thurman, 1993. Persistence of Herbicides in Selected Reservoirs in the Midwestern United States: Some Preliminary Results. U.S. Geol. Surv. Open-File Rep. 93-418, p 51-63.

Goolsby, D. A. and W. A. Battaglin, 1995. Occurrence and Distribution of Pesticides in Rivers of the Midwestern United States. In: Agrochemical Environmental Fate: State of the Art; CRC Press, Boca Raton, FL, p 159-173.

Helsel, D. R. and R. M. Hirsch, 1992. Statistical Methods in Water Resources. Elsevier, New York.

Kalkoff, S. J. 1993. Water Quality of Corydon Reservoir Before Implementation of Agricultural Best-Management Practices in the Basin, Wayne County, Iowa, September 1990 to September 1991. U.S. Geol. Surv. Water-Resour. Invest. Rep. 93-4099.

Leung, S. T., R. V. Bulkley and J. J. Richard, 1982. Pesticide Accumulation in a New Impoundment in Iowa. Wat. Resour. Bull. 18 (3), p 485-493.

Miller, J. G. and J. O. Kennedy, 1995. Study of Herbicides in Water and Sediment from 19 Iowa Water Supply Reservoirs January - February, 1995. Univ. of Iowa Hyg. Lab. Rep. No. 95-1.

Rea, A. and J. R. Cederstrand, 1994. GCIP reference data sets (GREDS). U.S. Geol. Surv. Open-File Rep. 94-388, CD-ROM.

Schottler, S. P., S. J. Eisenreich and P. D. Capel, 1994, Atrazine, Alachlor, and Cyanazine in a Large Agricultural River System. Environ. Sci. Technol., 28, p 1079-1089.

Scribner, E. A., D. A. Goolsby, E. M. Thurman, M. T. Meyer and M. L. Pomes, 1994. Concentrations of Selected Herbicides, Two Triazine Metabolites, and Nutrients in Storm Runoff From Nine Stream Basins in the Midwestern United States, 1990-92. U.S. Geol. Surv. Open-File Rep. 94-396.

Scribner, E. A., D. A. Goolsby, E. M. Thurman, M. T. Meyer and W. A. Battaglin, 1996 (in press) Concentrations of Herbicides, Herbicide Metabolites, and Nutrients in the Outflow from Selected Midwestern Reservoirs, 1992-93. U.S. Geol. Surv. Open-File Rep. 96-393.

Stamer, J. K., K. D. Gunderson and B. J. Ryan, 1995, Atrazine Concentrations in the Delaware River, Kansas. U.S. Geol. Surv. Fact Sheet FS-001-94.

Steeves, P. and D. Nebert, 1994. Hydrologic Units Maps of the Conterminous United States, 1:250,000-scale (nominal), ARC/INFO format, 1994.

U.S. Army Corps of Engineers, 1982. National Inventory of Dams Database. National Technical Information Service, Springfield, VA, digital file.

U.S. Department of Agriculture, 1993, State Soil Geographic Data Base (STATSGO) --Data Users Guide. U.S. Department of Agriculture, Soil Conservation Service, Misc. Pub. Num. 1492.

U.S. Department of Commerce, 1990. Census of Population and Housing, 1990. U.S. Department of Commerce, Bureau of Census, Data Users Division, Washington, D.C.

U.S. Environmental Protection Agency, 1989. Drinking Water Health Advisory: Pesticides. Office of Drinking Water Health Advisories, Lewis Publishers, Inc., Chelsea, Michigan.

U.S. Environmental Protection Agency, 1995. Drinking Water Regulations and Health Advisories. U.S. Environmental Protection Agency, Office of Water, Washington, D.C.

Wolock, D.M., 1993, Simulating the Variable-Source-Area Concept of Streamflow Generation with the Watershed Model TOPMODEL. U.S. Geol. Surv. Water-Resour. Invest. Rep. 93-4124.

Authors

2. Donald A. Goolsby
U.S. Geological Survey, WRD
Box 25046, MS 406,
Building 53, Wing F-1200, DFC
Denver, Colorado 80225
Email: dgoolsby@usgs.gov

GIS Symposium | home page
Maintainer: AWRA Webserver Team Last modified: 24Nov99 gaw
Copyright(c) 1996, American Water Resources Association