Fuzzy decision tree classification for high-resolution satellite imagery

There are many algorithms for remote sensing image classification. This study discusses the problems involved in classification of high spatial resolution satellite images and investigates the use of different traditional classification algorithms and decision tree classifiers. Maximum likelihood classifier, which is the most widely used traditional distance-based classifier, is compared with CART (Classification and Regression Tree) classification, which is a decision tree classifier. Maximum likelihood performs better for coarse resolution images as the data is in normal distribution. But as the spatial resolution increases the data has more within-class variations and it is no longer in normal distribution. CART algorithm was observed to be better than maximum likelihood for classifying data which do not have a Gaussian distribution and are multi-modal with overlapping classes.; Maximum likelihood is a parametric distance-based classifier and therefore it can be used to determine quantitatively how close the mis-classified pixels are to other classes in the spectral space. It can be used to derive fuzzy partitions of the classification. But CART does not provide any information about the mis-classified pixels. Therefore we build a fuzzy decision tree classifier which can determine the fuzzy partitions of classification from the crisp CART algorithm. A membership function was defined to assign class memberships to the arcs of the tree. Using these values we produce an ambiguity index image from which we can determine areas with high ambiguity.