A fully automated design of binary decision tree for land cover classification

A fully automated design method of binary decision tree is proposed for land cover classification of remote sensing data. The proposed method has the following features: (1) all possible binary combinations of identification classes are tested as splitting pattern of classes; (2) additional variables by linear combination of paired bands are applied in data segmentation; (3) automatic interaction detector technique is utilized to divide the data set into binary tree nodes. Completely enumerated Landsat MSS data with land-cover data are used for the test of the method and classification accuracy. The authors' new method is compared with Bayesian discriminant classifier in terms of accuracy and training sample set sizes.