Integrating parametric uncertainty and modeling results into an advisory system for watershed management

This paper describes an approach to integrate complex modeling experience i nto a decision support framework for non-point source pollution modeling of a watershed. The approach employs probabilistic reasoning techniques and d erives probability distributions from previous model simulations. Thus, the sensitivity of a given model to its inputs is captured in such a way that the system can be used to find solutions to management problems through the application of probabilistic inference. A graphical probability model is a visual formalism encoding random variables and relationships between rando m variables as nodes and directed links in a graph. In our model, the nodes represent individual parameters for each of the fields in a watershed. Dir ected links connect correlated nodes, such as nodes representing the manage ment practice in a field to nodes representing soil loss rates. The directe d links between the nodes in the probability model follow the drainage netw ork of the watershed. The relationships (or links) between the nodes are qu antified via two methods. The first method integrates data (cases) derived from Monte Carlo simulation of a non-point source (NPS) pollution model. In the second method, deterministic functions defined in the NPS model are us ed to specify the relationships. The Monte Carlo simulations are performed to include the influence of parametric uncertainty on model results. The ne twork for an entire watershed is complex with a lar-e number of nodes, ther efore, a spatial analysis/visualization tool was developed for interacting with the large probability model. (C) 2001 Elsevier Science Ltd. All rights reserved.