Kriging Models Predicting Atrazine Concentrations in Surface Water Draining Agricultural Watersheds. Issue 5 (1st September 2016)

• Record Type: Journal Article
• Authors: Mosquin, Paul L.; Aldworth, Jeremy; Chen, Wenlin
• Is Part Of: Journal of Environmental Quality. Volume 45:Issue 5(2016)
• Page Start: 1680
• Page End: 1687
• Publication Date: 2016-09-01
• Subjects: Agricultural ecology -- Periodicals; Environmental engineering -- Periodicals; Pollution -- Periodicals; 630
• DOI: 10.2134/jeq2015.10.0544
• Languages: English
• ISSNs: 0047-2425

Abstract

Abstract : The performance of kriging methods in predicting maximum m‐day (m = 1, 7, 14, or 30) rolling averages of atrazine concentrations in 42 site‐years of Midwest Corn Belt watersheds under two systematic sampling designs (sampling every 7 or 14 d) was examined. Daily atrazine monitoring data obtained from the Atrazine Ecological Monitoring Program in the Corn Belt region (2009–2014) were used in the evaluation. Both ordinary and universal kriging methods were considered, with the covariate for universal kriging derived from the deterministic Pesticide Root Zone Model (PRZM). For the maximum 1‐d rolling averages, prediction did not differ among methods. For rolling averages of longer duration (m > 1), predictions obtained by linear interpolation on a logarithmic scale were better (up to 15% lower for 7‐d sampling and 22% lower for 14‐d sampling in terms of the relative root mean squared prediction error) than those obtained by linear interpolation on the original linear scale and also less variable. For kriging methods, empirical semivariograms of daily atrazine time series suggest a negligible noise process, supported by replicate analysis of selected field samples; piecewise linear semivariogram models were found to perform best for predicting sampled data. We demonstrate that kriging prediction intervals offer close to nominal coverage for unsampled values. Core Ideas: For right‐skewed data, kriging methods on log scale are useful. For right‐skewed data, linear interpolation on log scale is useful. For right‐skewed data, linear interpolation on linear scale should be avoided. Estimate of nugget close to zero is extremely useful for semivariogram modeling. Linear semivariogram models worked at least as well as nonlinear models.