于 欢1,2, 张树清1, 孔 博3, 李晓峰1,2,3(1.中国科学院东北地理与农业生态研究所, 长春 130012;2.中国科学院研究生院, 北京 100049;3.中国科学院成都山地灾害与环境研究所, 成都 610041)
Optimal Segmentation Scale Selection for Object-oriented Remote Sensing Image Classification
YU Huan,1,2, ZHANG Shuqing1, KONG Bo3, LI Xiaofeng1(1.Northeast Institute of Geography and Agricultural Ecology, Chinese Academy of Science, Changchun 130012;2.Graduate School, Chinese Academy of Sciences, Beijing 100049;3.Institute of Mountain Hazards and Environment, Chinese Academy of Sciences, Chengdu 610041)
Segmentation is a basic and pivotal step of object-oriented remote sensing image classification, and the scale is a key problem of image segmentation. Aiming at the optimal segmentation scale selection for object-oriented remote sensing image classification, conformity degree between vectorial boundary lines of image region object after segmentation and true boundary lines of classification objects as criterion, through their multi-dimensions distance to define the conformity degree, the paper brought forward a new method of optimal segmentation scale selection for object-oriented remote sensing image classification-vector distance index method. Research verified the validity and applicability of this method through two experiments One experiment compared the results of optimal segmentation scale selection based on vector distance index method and ‘trial and error’ method. Results showed the vector distance index could reflect the optimal segmentation scale for seven classes exactly. The other experiment classified the segmentation result from the first experiment, through the accuracy assessment, explored the relationship of selection result based on vector distance index method and classification accuracy. Results showed water, dryland, rice, forest and resident region gained the highest accuracy on the scale that was selected by the vector distance index method, although marsh and grass didnt gain the highest accuracy on the scale that was selected by the vector distance index method, through further analysis, the classification result tallied with the practical condition. Then both two experiment results showed that this method could realize the optimal segmentation scale selection for object-oriented remote sensing image classification, which was similar to human thought, unconfined to data source, intuitionistic, comprehensible and practical. Based on the basic theory of vector distance index method, aiming at the ‘submergence’ and ‘fragmentation’ phenomenon, research brought forward a scale index, which could reflect the small or big status for a given object type, and provided a quantitative tool to assess the conflict degree between ‘submergence’ and ‘fragmentation’, then showed its significance during the process of segmentation scale selection.