Monitoring Network Design for Phytoremediation Systems Using Primary and Secondary Data Sources

Abstract

Phytoremediation, or contaminant removal using plants, has been deployed at many sites to remediate contaminated soil and groundwater. Research has shown that trees are low-cost, rapid, and relatively simple-to-use monitoring systems as well as inexpensive alternatives to traditional pump-and-treat systems. However, tree monitoring is also an indirect measure of subsurface contamination and inherently more uncertain than conventional techniques such as wells or soil borings that measure contaminant concentrations directly. This study explores the implications for monitoring network design at real-world sites where scarce primary data such as monitoring wells or soil borings are supplemented by extensive secondary data such as trees. In this study, we combined secondary and primary data into a composite data set using models to transform secondary data to primary, as primary data were too sparse to attempt cokriging. Optimal monitoring networks using both trees and conventional techniques were determined using genetic algorithms, and trade-off curves between cost and uncertainty are presented for a phytoremediation system at Argonne National Laboratory. Optimal solutions found at this site indicate that increasing the number of secondary data sampled resulted in a significant decrease in global uncertainty with a minimal increase in cost. The choice of the data transformation model had an impact on the optimal designs and uncertainty estimated at the site. Using a data transformation model with a higher error resulted in monitoring network designs where primary data were favored over colocated secondary data. The spatial configuration of the monitoring network design was similar with regard to the areas sampled, irrespective of the data transformation model used. Overall, this study shows that using a composite data set, with primary and secondary data, results in effective monitoring designs, even at sites where the only data transformation model available is one with significant error.