Probabilistic Models in Geographic Information Systems - Bayesian Networks for Management of Uncertainty

摘要

The challenges of managing uncertainty in GIS require general but very powerful statistical models. This paper reports on an implementation of Bayesian Network (BN) techniques to construct probabilistic models that meet the requirements of managing uncertainty in GIS applications. BNs can be considered as a generalization of Generalized Weighted Least Squares (GLS) techniques used in geodesy and photogrammetry. BNs extend GLS "like" techniques to the general GIS case of categorical data, any mathematical model, and any probability distribution. BNs can alternatively be considered as a generalization of rule-based expert systems to a more powerful reasoning capability that correctly incorporates uncertainty estimates. BNs have been successfully applied to Artificial Intelligence problems that require modeling of, and reasoning with, uncertain or incomplete knowledge. BNs define a complete probabilistic model from local conditional probability distributions. This characteristic can be exploited to automatically build situation specific BNs. This paper demonstrates that the challenges of management of uncertainty in GIS can be met with software that can automatically construct and solve BNs for GIS models.