Enhancing post-classification change detection through morphological post-processing – a sensitivity analysis

Monitoring land-cover change is often done by simple overlay of two classified maps from different dates. However, such analysis tends to overestimate the rate of change. Main error sources are the mis-registration between classified maps and their thematic accuracies. This study proposes a change detection method with morphological post-processing to improve change detection accuracy in comparison with traditional post-classification by taking into account these error sources. The method is developed for binary maps and is based on standard morphological procedures that are generally integrated in common spatial processing or free software. A detailed sensitivity analysis of this method based on simulated data sets of different landscape characteristics and error levels demonstrated the potential improvement. The degree of improvement in change detection accuracy mainly depended on the error type and level and the fragmentation of the landscape. In particular, location error effects on change detection were strongly reduced independent of class proportion. Up to 60% improvement in user's accuracy of change could be achieved for maps with location error and characterized by fragmented landscapes. Coping with classification errors was shown to be more challenging. A user-friendly reference table summarizes the potential improvement through the proposed methods for various landscape characteristics and error sources.