Research On The Classification And Building Extraction Of High Resolution Image Based On Object-oriented And Class Rules

Based on the continuous development of space remote sensing technology, the spatialresolution, spectral resolution and temporal resolution of remote sensing images has beengreatly increased. Satellite remote sensing technique of high resolving power gets a quickdevelopment in recent year has reached to an unprecedented altitude.Compared with themassive high-resolution image data, how to efficiently interpret and process a large number ofhigh-resolution remote sensing data which have rich space feature,geometry feature andtexture feature,has become an important and difficult research. The paper is based on twostudy points to improve classification accuracy with Worldview-2remote sensing images inSouth Tibet. Firstly, the study point is the best segmentation scale method, such asmean-variance method、maximum area method and scale correlation method. Through theanalysis of three scale selection methods, the result shows that alone using mean-variancemethod and maximum area method can not make sure the best segmentation scale of landfeatures, and the scale correlation method can not fully show the inherent land features scale.Thus, I integrate three method to select the best scale segmentation for every land features andset up multi-level network image object hierarchy. The second study point is to discover andmine land features knowledge and build classification rules. Through collection and analysiscomprehensively and typically study area land features samples, we study the change ofspectral(NDVI、NDWI), shape(Length/Width、Width、Rectangular Fit、Curvature/Length、stddev curvature and density) and custom features enhanced parameters. The paper shows thatthe combination of feature rules can maximize the difference in the extraction processing andobtain efficient and accurate classification of high resolution remote sensing images. At thesame time, we use identical training samples, checking samples, segmentation scaleparameters and object features to classify and evaluate precision with support vector machineand nearest neighbor classifier based on image objects, and compare three classification mapsand precision. The experimental results show that object-oriented and rules classificationmethod can make classification more visual interpretation and get more precision inclassifying high resolution images. And the overall accuracy and Kappa coefficient ofobject-oriented and rules classification method is97.38%and0.9673.It is high6.23%and0.078than object-oriented SVM classification method.Also it is high7.96%and0.0996thanobject-oriented KNN classification method. The building producer accuracy and useraccuracy is high18.39%and3.98%than object-oriented SVM classification method, and ishigh21.27%and14.97%than object-oriented KNN classification method.