Categorical models for spatial data uncertainty

Considerable disparity exists between the current state of the art for categorical spatial data error modeling and the current state of the practice for reporting categorical data quality. On one hand, the general Monte Carlo simulation-based error propagation framework is a fixture in spatial data error handling; researchers have identified potentially powerful approaches to characterizing categorical data error so that its effects on application uncertainty may be assessed. On the other hand, standard data quality assessments for categorical data are 'spatially unaware,' fail to provide critical information for error propagation models, and neglect the fitness for use paradigm underlying the longstanding rationale for accuracy metadata. Many error assessments rely on area-averaged indicators of map error that do not reflect spatial variability, such as the confusion matrix. How might this gulf between state of the art and state of the practice be bridged? In the present work we lay the foundation for such an edifice: we contrast several categorical error models proposed in the literature in terms of input parameters and performance for a heterogeneous land cover dataset. Familiar methods such as the confusion matrix are considered for their utility in developing error propagation models, as well as theoretically-based, spatially explicit methods like indicator simulation that are not commonly employed in applied research. We develop a comparative matrix to summarize different model requirements, characteristics, and performance, and utilize available secondary data sources where possible to develop improved inputs for the analysis of uncertainty propagation.