Effect of alternative splitting rules on image processing using classification tree analysis

Rule-based classification using classification tree analysis (CTA) is increasingly applied to remotely sensed data. CTA employs splitting rules to construct decision trees using training data as input. Results are then used for image classification. Software implementations of CTA offer different splitting rules and provide practitioners little guidance for their selection. We evaluated classification accuracy from four commonly used splitting rules and three types of imagery. Overall accuracies within data types varied less than 6 percent. Pairwise comparisons of kappa statistics indicated no significant differences (p-value > 0.05). Individual class accuracies, measured by user's and producer's accuracy, however, varied among methods. The entropy and twoing splitting rules most often accounted for the poorest performing classes. Based on analysis of the structure of the rules and the results from our three data sets, when the software provides the option, we recommend the gini and class probability rules for classification of remotely sensed data.