A Bayesian Method for Groundwater Quality Monitoring Network Analysis

Water resources management is best done when an efficient and reliable information system is available that provides data about the state of the water system being managed. Herein, a new methodology is developed and demonstrated that incorporates different aspects of monitoring, including vulnerability assessment, environmental health risk, the value of information, and redundancy reduction in an existing monitoring network. These various objectives reflect a desire for cost- efficiency, while taking into account uncertainty in predictions of groundwater contamination potential. This methodology is presented within an integrated approach to groundwater quality monitoring that includes several multi-objective aspects associated with decisions about monitoring. The result is an integrated, unified conceptual framework for groundwater quality monitoring. The importance of this framework appears in its iterative approach toward monitoring and the capture of information about the possible tradeoffs between different monitoring objectives. A Bayesian framework, based on the use of the relevance vector machine (RVM) modeling approach, plays a basic role in this conceptual framework. The RVM is employed to reduce redundancy and create a probability map of contaminant distribution. It is also used to estimate the expected value of information. The results obtained in this research show that the RVM is a useful tool for improving the efficiency of monitoring systems, both in terms of reducing redundancy (i.e., reducing costs) and increasing the information content of the data that are collected (i.e., decreasing the uncertainty in the understanding of the system being monitored). This is demonstrated through a case study application to nitrate contamination monitoring in the West Bank, Palestine. In this case study, the RVM analysis showed that only 30% of existing monitoring sites were needed to produce 98% statistical efficiency in the information obtained (as measured by the index of agreement). In addition, the results also show better vulnerability assessment performance compared to well known and recently applied vulnerability assessment methods in terms of sparsity of the model that is developed and accuracy of the predictions of the model. This is demonstrated from the misclassification or overlap error of less than 10% for the RVM modeling approach based on 190 monitoring sites. However, in this application, the results of health risk assessment and the evaluation of monitoring investments were less encouraging due to the minimal elasticity of the nitrate health effect with respect to monitoring information and uncertainty.