Abstract

In this paper, it is proposed an unsupervised methodology based on the Object Image Based Analysis (OBIA) paradigm, for the determination of multiscale training sets (ROIs). This methodology is based on the following hypothesis: the objects selected in an unsupervised way and characterized by certain attributes provide meaningful and reliable training sets for supervised classification. The proposed methodology allows the determination of regions at different scales, adapting the training set (size and number) to land cover characteristics. In order to show the potential and validity of this methodology, regions of interest have been used as input patterns to a nonparametric classifier (Decision Tree) and the results have been compared with classification results obtained when the classifier is trained with a set of training patterns obtained manually. Some initial experiments show that the proposed methodology provides classification results with comparable quality, and in most cases better, than that obtained when the ROIs are manually selected. Moreover, this methodology eliminates routine tasks and operator involvement is limited to make decisions, reducing time, cost and subjectivity in the selection of the ROIs. Therefore, it is expected that a more thorough study of the selection criteria and attributes used for ROIs characterization, will improve the quality of ROIs in terms of the accuracy of the classification results, maintaining the advantages already mentioned.