Bayesian belief modelling

Progress update June 2012

The Bayesian Belief Network (BBN) model is intended to answer spatially oriented questions from the IFS workshops. BBN is a statistical tool. It has a spatial component and is more data intensive than the Mediated Model (MM) and oriented to natural science rather than social and economic concerns. The preliminary results of the BBN approach using Macroinvertebrates as an indicator is being described in Death et al (in preparation, 2012). The plan is to include fish and periphyton as additional indicators. As part of the IFS research project we explore if/how BBN and MM can be used in conjunction for adaptive management. The questions to be addressed by the BBN model are determined by what is identified as important to participants in the MM process and/or in the Action Plan.  

Figure 1: Water Quality in the Manawatū River Catchment

Water Quality in the Manawatu River Catchment

How the BBN links with mediated modelling

The Bayesian Belief Network (BBN) model links spatially explicit GIS maps of climatic, geological, hydrological and landuse/landcover to biological functions in the river so scientific data can be presented in an accessible way to stakeholders participating in the MM workshops. These models unite current scientific knowledge of the catchment, in terms of climatic, geological, ecological and other scientific data. This data will be collated and presented in a way that is both requested by and accessible to  MM workshop participants. It will be able to show the probable effects of changes in land uses on water quality and the ability of certain fish species to survive in rivers and streams.

Bayesian belief networks are statistical models, calculating probabilities that can be used in a spatially explicit GIS environment. A BBN is a graphical structure that allows for the representation of and reasoning about risk scenarios. The nodes in a network represent a set of variables in the domain being modelled. The nodes are connected by links representing the relationship between variables. These relationships can be learned from the data if these are available or can be elicited from experts in the field. There can be many independent or predictor variables and many dependant variables. 

BBNs can answer ‘how’ questions by selecting a desired outcome and looking at how the independent variables are changed. Or ‘what-if’ scenarios can be tested by changing predictor variables to states expected in the future, then observing changes in dependant variables. Any information supplied to the network will update the probabilities throughout the network immediately, and the strength of the prediction can be judged by the probability value for a given outcome. As with the connections between the variables, the probabilities can be supplied by experts or learned from the data if available. Utility nodes can be added to calculate costs of different decisions.

BBNs are a very flexible modelling tool with many unique features such as an ability to model multiple dependant variables, or handle missing data; they can update probabilities from only one piece of information, they can be predictive, diagnostic or classifiers, and they are intrinsically informative, given the ability to update immediately on screen when any information is provided. The ability of BBNs to model in real-time real scenarios makes them ideal as a tool for AM of environmental situations and they can be linked directly to GIS maps to visualise management changes immediately. BBN has been used elsewhere but not populated with real data as intended in this project.

Massey Contact Centre Mon - Fri 8:30am to 5:00pm 0800 MASSEY (+64 6 350 5701) TXT 5222 contact@massey.ac.nz Web chat Staff Alumni News Māori @ Massey