Impacts of Error on the Predicted Pattern of Change in a Post-Classification Change Analysis
Amy C. Burnicki
Department of Geography, University of Wisconsin-Madison,Madison, WI, USA
Abstract: One of the most common uses of time-series classified imagery is the monitoring of changes in land-cover composition and structure over time. A common approach to map changes in land-cover is post-classification change detection. The resulting accuracy of a post-classified map of change depends directly on the patterns of error associated with the time-series land-cover maps. This research examined the impact of classification error on the spatial pattern of change observed in a change map. A series of Iand-cover-change maps were produced using a simulation approach that controlled the: 1. amount and pattern of change occurring between the time-1 and time-2 classified maps; and 2. amount and pattern of classification error associated with the time-1 and time-2 classified maps. Both error-free and error-perturbed maps of change were produced and compared using landscape pattern indices. Results showed an increase in fragmentation within the land-cover-change classes (e.g., increase in number of land- cover-change patches, total edge, shape complexity) under all error conditions. Fragmentation was greatest when the spatial autocorrelation of the change class increased. Regardless of the pattern of change considered, errors associated with classified maps significantly altered the pattern of change simulated in the post-classification change analysis.
Keywords: land-cover change; simuattion; error propagation; landscape metrics