Context and objectives
The aim of this study is to evaluate the practical utility of existing error models taking into account the type of error information that is available, to identify relevant problems encountered in the modelling of uncertainty, to suggest strategies to explore the spatial structure of error in different GIS coverages and to incorporate this knowledge in the error modelling.
Expected scientific results
- The usefulness of simple error models which are based on global descriptions of attribute uncertainty strongly depends on the type of analysis one wishes to perform. While not suited for estimation at one location, the application of simple error models may yield satisfactory results in inventory studies, where output values obtained for different locations am summed (e.g. the calculation of area) and local differences are averaged out. Next to the absence of detailed information on data uncertainty at a specific location, a lack of knowledge of the spatial structure of uncertainty proves to be a serious handicap for uncertainty modelling. If analysis encompasses global operations (e.g. visibility analysis), or if regional estimates are to be obtained, the level of autocorrelation in the error field has a major impact on the estimation of output uncertainty. It is therefore important that in the future more attention is paid to the reporting of the spatial structure of error in source data.
- The only method for the modelling of uncertainty in categorical data which takes spatial correlation into account only works for two classes, which substantially reduces its practical utility. In this study the method was generalised for use with more than two classes. Within each class inclusions of other classes can be simulated through specification of the proportion and the mean size of the inclusions of these classes within the original mapping unit.
- A much more detailed modelling of image classification uncertainty becomes possible if next to the classified image also the probabilities on which the classification is based are available within the GIS. Class membership probabilities not only describe classification uncertainty at the level of each pixel, but also provide information on the spatial structure of uncertainty which can be incorporated in the modelling.