Evaluation of streamflow as a covariate in models for predicting daily pesticide concentrations

Abstract

Several models have been developed with streamflow as a covariate for predicting daily pesticide concentrations in surface water systems. Among these models, the SEAWAVE-QEX model has been proposed by the United States Environmental Protection Agency for regulatory assessments. In this paper, the model was modified to include alternative transformations of streamflow data, and to include no streamflow covariates. The predictive performance of the modified models was evaluated and compared with the original SEAWAVE-QEX model using a high frequency sampling dataset that includes 9 sites with 10 years of data from the Atrazine Ecological Monitoring Program. Streamflow transformations evaluated included those in the SEAWAVE-QEX model (short-term and mid-term flow anomalies), reduced models with only short-term flow anomaly or without any flow covariates, normalized Box-Cox transformation of flow, and combinations of normalized Box-Cox and flow anomalies. Loglinear interpolation was also evaluated. The normalized Box-Cox transformation provided best predictive performance and significantly better predictive performance than that of the SEAWAVE-QEX model for a target quantity of regulatory interest, such as the maximum 1-day rolling average (similarly for the maximum 60-day rolling average, but not significantly so). The no-flow covariate model was only slightly worse than Box-Cox. Significant differences in predictive performance of the SEAWAVE-QEX model were detected across sites.