Post-classification change detection with data from different sensors: some accuracy considerations.

Change detection from remote sensing data is often done by simple overlay of classified maps. However, such analyses can contain a significant proportion of boundary errors, especially when combining data from different sensors. This paper presents a protocol that allows reliable post-classification comparisons by taking into account classification accuracies, landscape fragmentation, planimetric accuracies, pixel sizes and grid origins. The proposed protocol has been applied, with little extra effort, in a fragmented agricultural Mediterranean zone using MSS (1970s) and TM (1990s) images. Applying the protocol, change detection had an accuracy of 85.1%, while for a direct overlay it was only 43.9% accurate. The drawback of this method is that it reduces the useful area of comparison. As the accuracy of individual classifications is critical, the paper also describes and tests a hybrid classifier that combines an unsupervised classification approach with training areas. This approach has proved more successful than maximum likelihood classifiers. [ABSTRACT FROM AUTHOR]