A probabilistic decision analytical approach for watershed planning: A mercury total maximum daily load case study

This work develops a decision analytical approach to water quality management at the watershed scale through a mercury Total Maximum Daily Load (TMDL) development case study. This approach treats the key environmental variables as causally-related random variables that may be influenced through mitigation actions (interventions) to an uncertain degree. Starting from the perspective that water quality management falls under the rubric of "decision-making under uncertainty", I explore the application of state of the art probabilistic tools for decision support. This work goes beyond the current deterministic paradigm in which conservative modeling choices are used to deal with predictive uncertainty. The proposed decision model frames the TMDL setting process as a set of regulatory decisions that may involve large uncertainties (limited data bases and incomplete knowledge) and that is subject to tight regulatory deadlines and small decision process budgets.; Probabilistic source analysis and linkage analysis models based on the available data, standard environmental science and engineering theory, and mercury biogeochemistry expertise were created for the case study mercury TMDL decision situation. Discrete conditional probability distributions based on these models and expertise were incorporated in a Bayesian network model, a tool for solving prediction and inference queries. In conjunction with a parametric value model, this mercury Bayesian network serves as the basis of a mercury TMDL decision model for the case study. This decision model demonstrates a formal context for considering the importance of uncertainty in TMDL decisions, for prioritizing information collecting activities, for considering trade-offs between compliance uncertainty and mitigation costs, and for considering and representing hypotheses within a TMDL decision-modeling framework. Sensitivity analysis using the Bayesian network is used to demonstrate approaches for prioritizing information collection activities and for estimating the value of perfect information on variables of interest. As demonstrated, future information activities should be based on preliminary models of the uncertain relationships between possible interventions and environmental targets. Very importantly, the Bayesian perspective of decision analysis allows decision participants to interpret new information (monitoring and knowledge) in light of previous information and knowledge, which is a good basis for an adaptive management framework.