The Wetland Research With Remote Sensing Image Based On Object-oriented Classification In Xixi, Hangzhou

Wetland called "the kidney of the Earth" is the most productive ecosystem in the nature andone of the most important human settlements. But due to unreasonable exploitation of wetlandresources, the world is facing a great loss of wetland, as the value of wetland to society has beenrecognized; much effort has been made to preserve the wetland ecosystem.With the test site of Xixi Wetland in Hang Zhou and raw data of QuickBird, the research is onthe identification and extraction classification methods of wetland. Using multi-level andmulti-scale segmentation method, and then based on the analysis of spectral information of groundfeatures,shape feature,texture,vegetation index, and so on, the research constructs feature spaceabout the characteristics of the test site and optimizes them. Comparing to conventionalclassification methods, it takes full advantages of spectral feature,texture,position,shape feature,and so on, avoids isolated point noise, which results from the enhancement of interiorheterogeneity of the same object in the classification of high resolution images, and improves theaccuracy of classification. At the same time, in order to resolve the defect of unoptimizing offuzzy classification in object-oriented classification software, it uses knowledge-based decisiontree methodology. The research primarily generates a set of object-oriented wetland classificationmethodology.The core of the study and conclusions are as follows:(1) The objects of Xixi Wetland in QuickBird are individually qualified by field surrey andvisual interpretation. The features of ground objects are consistent in spectral information, and thediscrimination between constructions and water or vegetations is clear, particularly in NIR. Eachobject has the difference of spectral response feature, but has different signification in differentbands.(2) Using two-scale segmentation. vegetation,water and constructions in the image are Firstlyclassified. Then, vegetations more easily confused are extracted by mask, which are segmented bydifferent scales. According to regional combination algorithm, the distribution of fractal isachieved, i.e., the generating of object Multi-scale segmentation avoids wrong classification ofdifferent types and helps to improve the accuracy of classification. (3) After the objects segmentation, we choose mean, standard deviation and NDVI fromvegetations which is difficult to classify, and construct new feature space. Then the feature spaceis optimized in decision tree to reduce the calculation and improve the classification efficiency.(4) In combination with knowledge-based decision tree methodology, object-orientedmethodology can increase the accuracy of classification of wetland and work well in the automaticextraction of wetland. Many feature sets demonstrate the feature of wetland in different angles.The algorithm CART was used to extract optimum feature subset of wetland classification andcorresponding knowledge rules; then, vegetation types were recognized by using the positivereasoning mechanism. The accuracy of classification is clearly higher than conventionalclassification methods.