Comparative Experimentation Study On Information Extraction Of Object-Oriented Based On Quick Bird Image

Compared with the low or middle resolution image, the high resolution remote sensing image has the richer structure and texture information, but using traditional classification technology based on pixel spectrum can't obtain the ideal effect, the thematic map produced by traditional way has lots of "pepper and salt" noises which result in the lose of integrity, and "different things with same spectrum" and "different spectrum with same thing" can't be distinguished. This urgently requires people to put research on classification of the high-resolution remote sensing, in order to satisfy the increasing application and study requirement of high-resolution remote sensing images.Object-oriented method used eCognition, which is professional object-oriented remote sensing classification software developed by Definiens Imaging GmbH in Germany, and extracted information through multi-segmentation, establishing of class hierarchy structure and semantic structure, classification role and reference and so on.In this paper, selected the typical area of Quick bird image as study area, which size is 1135*1135, as followed is the work of this paper:1) During the process of multi-scale segmentation, adding texture to put multi-scale segmentation experiments to the study area, and get different segmentation parameters which are suit to different features. For example, segmentation scale of 50 is suit to extract building, vegetation and shadow; segmentation scale of 75 is suit to extract subordinate road; segmentation scale of 90 is satisfactory for extracting main road.2) Multi-scale segmentation is the basis of information extraction, object-oriented information extraction used the method of fuzzy classification, this experiment realized information extraction through two forms-the nearest neighbor and member function, the extraction precision by member function is higher than the nearest neighbor method for the study area.3) Compared extraction result of pixel-based (Maximum Likelihood) to extraction result of object-oriented. After contrasting we found that the whole accuracy of object-oriented method is 89.32%, higher 10.1% than accuracy of traditional classification (79.2691%). And using the method of object-oriented can distinguish different forms of the same features, for example bright building and dark building, tile building, asbestos building, main roads and subordinate roads.