Object-Oriented Classification and Data Mining: Mapping Urban Area with Open Source Tools

Abstract

Most of the traditional classifiers are pixel by pixel, which makes them less efficient in the separability of the objects in urban regions because of the spectral mix of the different objects in the scene, besides the frequent changes in the spatial dynamics. In Object Oriented classifiers (O.O.), also known as GEOBIA, a knowledge-based analysis is performed, extracting attributes of the segments for that the classification can be performed, that is, analyzing context information and not only in the isolated pixels. Thus, there is a tendency for classification to be more assertive. However, most of the O.O. classifiers are proprietary software, which ends up making the product generated more expensive. The purpose of this study was to evaluate the Object Oriented classification in high resolution images using the free InterImage classification software and the open source data mining Weka, which was used to obtain the thresholds for distinguishing the nodes from the semantic network. The classification was carried out in an urban expansion area of the Federal District, based on a QuickBird image, where a thematic map of soil use was generated with the purpose of evaluating the accuracy of the result through the confusion matrix, from the which were extracted accuracy index. Also, the quality of the vegetation cover was evaluated by extracting the NDVI index and was calculated the positive anomaly, with the purpose of analyzing the photosynthetically active vegetation. It was verified that about 3.14% of the total of classified green areas have high NDVI, demonstrating the health of the vegetation. The InterImage classifier presented satisfactory results, with a separability of the Global Accuracy classes of 82.5% and a Kappa coefficient of 0.806, thus promoting a significant discrimination between the different classes proposed in the study area.