Modeling unobserved true position using multiple sources and information semantics

We present a modeling framework that supports semantically informed statistical inference about unobserved true location and positional uncertainty for geographic information spanning multiple sources. We demonstrate the use of a semantic representation of information sources to support construction of a Bayesian belief network that operationalizes the data integration process. Our approach allows for positional data, metadata, and other ancillary and derived information to inform inference regarding unobserved true position. In our application, we use two line feature datasets and a set of GPS data points describing a portion of the street network in Laramie, Wyoming. Using source metadata we inform prior distributions. Additionally, we use feature straightness to illustrate how form and process - gridded streets and the process of the Public Land Survey - can be used to improve inference for true position. The presented modeling framework is suitable for multiple data sources when the best data are not necessarily known and when the information semantics associated with the input data can be described in a systematic way.