Scale Issues In Object-oriented Image Analysis

Entity, pattern and process show inherent hierarchy structure in nature. Scale issues correspond with hierarchy become one of important themes in many fields. The scale issues have increasingly attracted the attention of a number of scientists in nature and society sciences. It is widely acknowledged that it is a basic subject in remote sensing applications. Image information with different scale displays distinct spatial structure, so image analysis with single scale could not meet the heterogeneity and dynamic of pattern and process. Object-oriented image analysis creates meaningful objects and builds a hierarchy levels close to surface character using multi-scale segmentation. Different geography processes could be represented in corresponding image object levels. Object-oriented image analysis has realized multi-scale analysis of spatial pattern and process. In object-oriented image analysis "scale" is not the pixel size, but the segmentation scale threshold. Scale is different for each remote sensing image analysis. To show the variety of pattern and process correctly, it is important to understand image objects scale effect in scaling courses, and to select the optimal scale for special class abstraction.Based on methods and approaches in pixel-based image analysis, this thesis develops several results which focus on scaling and scale choose in object-oriented image analysis.Two aspects of optimal scale are exploited in object-oriented image analysis. One is the biggest resolution to extract the smallest class; the other is the optimal segmentation scale for different classes.Developed a new method for selecting optimal segmentation scale: mean variance method. From the graph many classes' optimal segmentation scales could be got. It proved that for different classes their optimal segmentation scale are different. It break the limitation of one optimal scale choose. There is one disadvantage for this method. The result is good for high-resolution images, but is inefficient for middle resolution images.Developed another new method for selecting optimal segmentation scale: objects max-area method. The graph that shows the relationship of objects max-area andsegmentation scale is ladder like increased. Each flat is the appreciate scale range for one class. The result proved that optimal scale is a range, not a point.Presented three conceptions of object: pure object, mix object and false object. Analysis in one image level that has most pure objects will get high classification accuracy. The optimal segmentation scale means in that image level pure object is most and mix object is lest.Multi-scale segmentation is a good scaling method in object-oriented image analysis. Objects in different image levels are the aggregation of the initial pixels with the appointed scale threshold. This scaling method could avoid losing much meaningful information.Mean, StdDev, heterogeneity coefficient and fraction dimension are the indices for evaluating the scaling method in object-oriented image analysis. The change trend of these indices corresponding to segmentation scales is used to judge the validity.In this thesis, object-oriented image analysis is introduced fully. Object-oriented approach is a current for image analysis today. Image segmentation and information extraction are two main technologies.A snapshot of object-oriented image analysis and optimal scale choose is presented to detect vegetation information in Daqing. The result of vegetation information extraction is promising and the precision of classification is higher than other conventional processes. It is obvious that this image analysis approach offers a satisfying solution to extract information quickly and efficiently.