CHANGE DETECTION UNDER UNCERTAINTY: MODELING THE SPATIAL VARIATION OF ERRORS

This research focuses on modeling the spatial variation of errors in a land cover map and in the detection of changes when local uncertainties are identified. The motivation for this research stems from the fact that although a plethora of classification accuracy indices exists, most of them are global measures, such as the percentage correct and kappa index which are being the dominant measures so far. There is certainly a need to know the spatial distribution of errors in order to take informed decisions for further research and better interpretation of the results. The methodology that was followed, initially focused to land cover mapping and change detection mapping, for land use changes that have taken place since 1975. The latter was achieved by devising a simple and operational rule-based approach to map land cover changes, based on the classification of Landsat imagery (MSS for 1975 and TM5 for 1990, 1999, 2007) and the conceptual analysis of the information regarding change detection. The use of ancillary GIS data such as a Digital Elevation Model, existing thematic maps and the knowledge of the island's vegetation dynamics, formed the basis for setting the rules (e. g. impossible changes, valid transitions) for the post-processing of the classified images that led to a more accurate assessment and mapping of land cover changes. Consequently, a model-based approach to estimate local uncertainties is proposed by using independent ground truth samples and GAM models to estimate the probability of error occurrence in the study area. Change detection maps were subsequently cross-tabulated with error maps in order to provide information on the reliability of the change estimates. The results can be used in various ways. Firstly, as an iterative classification process corrected by independent ground truth samples, taken at the areas with low accuracy. Secondly, the analyst can proceed to use the enhanced change detection maps, making informed decisions and incorporating uncertainties in future modeling.