Improving Automation in Rule-based Interpretation of Remotely Sensed Data by Using Classification Trees

The definition of good classification rules for rule-based interpretation of remotely sensed data is a laborious and demanding task. One interesting method that could be used to automate the process is the classification tree method. It can be used to create a tree-structured classification hierarchy and rules automatically from training data. In this study, tests were carried out using the classification tree method in two applications: building detection using laser scanner and aerial image data and land-use classification using E-SAR data. The method was applied to segments with a large number of different attributes. The results were satisfactory and the classification accuracy was near to that obtained in previous studies using manually created classification rules. The most important benefit of the classification tree method was its high level of automation and speed compared with the process of defining the rules manually. A combination of the classification tree method and permanent, up-to-date reference data could be a useful tool in developing new classification applications and testing the feasibility of new remotely sensed datasets. Together with segments and attributes derived from remotely sensed data, it could be used for the rapid construction of classification trees, which could then be directly applied to classification or used as a starting point for further development of the rules.