ABSTRACT
As all modeling is a mere abstraction of much more complex processes that in many cases might not be fully understood, uncertainties are also an intrinsic part of the approach. Uncertainties are thus not necessarily a problem in modeling, but rather an inherent component of the process, as long as the sources and bounds of the uncertainties associated with individual models are known and understood. Where this is the case, sensitivity tests can be conducted to assess the susceptibility of model results to uncertainties in certain data, parameters, or algorithms and compare these uncertainties with the sensitivity of model runs to variations in individual parameters. Decision makers have become increasingly familiar with such methodologies, through, for example, the scenarios presented in IPCC reports (IPCC, 2001). While uncertainties inherent in spatial data have been the focus of a number of research projects in the GIScience community, many users of spatial data either completely neglect this source of uncertainty or consider it less important than, for example, parameter uncertainties. However, even if a modeler is aware of the uncertainties introduced through, for instance, a Digital Elevation Model (DEM), it is not always straightforward or even possible to assess them, e.g., when metadata from the data producers are incomplete, incorrect, or missing. If this information cannot be reconstructed, assumptions have to be made that might or might not be realistic and sensible for testing the impact of uncertainties in spatial data on a model. In this chapter, we use the term “error” when referring to the deviation of a measurement from its true value. This implies that the elevation error of a DEM can only be assessed where higher accuracy reference data are available (Fisher and Tate, 2006). Error is inherent in any DEM, but is usually not known in terms of both magnitude and spatial distribution, thus creating uncertainty. “Uncertainty” is used in this context, where a value is expected to deviate from its true measure, but the extent to which it deviates is unknown, and it can only be approximated using uncertainty models (Holmes et al., 2000).
