| In recent years,with the development of radar polarization measurement technology and synthetic aperture radar(SAR)imaging technology,polarimetric SAR(PolSAR)which combines the advantages of both technologies has emerged.PolSAR has gradually replaced traditional low resolution and single-polarization SAR,and has become the main direction of the development of modern radar system.Effective PolSAR image interpretation technology is the basis of the successful application of PolSAR.However,compared to mature PolSAR imaging technology and system design,the development of PolSAR image interpretation technology is relatively backward,which seriously restricts the application of PolSAR,so the research of PolSAR image interpretation technology is of much significance in the improvement of the application efficiency of PolSAR system.An important part of image interpretation is land-use classification.Pixel-based classification method only uses the characteristics of pixel itself in classification.Although the pixel-based classification method can retain the detail informations of images better,but for PolSAR images,speckle noises will cause the errors between the measured value and the true value of single pixel.The errors will result in the appearances of the small areas and the isolated pixels in classification result,which will increase the difficulty of image interpretation.Object-oriented classification method enables the acquisition of a variety of textural and spatial features for improving the accuracy of classification by delineating objects from images.Furthermore,image objects are less affected by the speckle noises in PolSAR images than in pixels.Therefore,the research on the application of object-oriented technology in PolSAR image classification is of great significance to promote the development of PolSAR image classification.Aiming at the main problems about the object-oriented classification of PolSAR images currently,taking care of the complexity of the existing theories and problems,with the latest theories of polarimetry,image processing,computer vision,data mining and pattern recognition,this thesis made an deep study on the correlative problems with the strategy of “polarimetric decomposition---image segmentation---feature selection---classification---multiple classifiers combination”.The study site was located in the southeastern part of Changchun City,Jilin Province.A RADARSAT-2 Fine Quad-Pol image was selected as data source for this study.The main work and the research results in this thesis can be summarized as follows:1.In this thesis,many classical polarimetric decomposition methods were used to extract polarimetric parameters for classification support.These decomposition methods are the Pauli,Krogager,Huynen,Barnes1,Barnes2,Cloude,H/A/α,Freeman2,Freeman3,Yamaguchi3,Yamaguchi4,Neumann,Touzi,Holm1,Holm2,and Van Zyl methods.A total of 61 polarimetric parameters were extracted from the RADARSAT-2 image using these polarimetric decomposition methods.All the extracted polarimetric parameters were synthesized to form a multi-channel image with 61 layers on which subsequent processing steps were performed.2.According to the characteristics of PolSAR images,a new PolSAR image segmentation algorithm based on the fusion of edge information and region information was proposed.In order to test the performance of the proposed segmentation algorithm,two subareas of the Radarsat2 image were selected for segmentation experiments,and the results of the experiments were compared with the results obtained by FNEA algorithm in eCognition software.The results demonstrated that the proposed segmentation algorithm was very suitable for the Radarsat2 image,and it also had a strong antinoise ability,which is very important for the analysis and the interpretation of PolSAR data.3.The availability of numerous features with object-oriented classification renders the selection of optimal features.In this thesis,a hybrid feature selection algorithm which combined filter approach and wrapper approach was proposed.In the filter approach,ReliefF algorithm was employed to filter out the features which had the lower relevance with land-use classes.The wrapper approach used genetic algorithm(GA)as a search method and classification accuracy as an evaluator to search for an optimal feature subset from the selected features.In order to verify the effectiveness of the proposed feature selection algorithm,a series of comparisons between the proposed algorithm and the 3 other feature selection algorithms were made.The results of the comparisons indicated that the proposed hybrid feature selection algorithm could be effectively applied to object-oriented land-use classification.4.In addition to support vector machine(SVM)and decision tree(DT)algorithms which are commonly used in the object-oriented classification of PolSAR images,this thesis also used random forest(RF),Bayes,and k-nearest neighbor(kNN)algorithms for the land-use classification.The results showed that the ranking of the 5 algorithms in the order of highest to lowest classification accuracy was: SVM>RF>DT>kNN>Bayes;The classification that used SVM algorithm obtained the highest overall accuracy and Kappa value of 92.60% and 0.9022;The overall accuracy and the Kappa value of the classification using Bayes algorithm were 87.57% and 0.8376,which were lowest among the 5 algorithms.5.Multiple classifiers combination method was introduced into the object-oriented classification of PolSAR images.To begin with,Q statistic,correlation coefficient,and entropy were employed to measure the diversities of the combinations which are composed of different classifiers.Then,the 6 combinations which had larger diversities were selected to carry out the experiments of multiple classifiers combination.The results showed that the overall accuracy and the Kappa value of the 6 combinations were improved through multiple classifiers combination method,which proves the effectiveness of multiple classifiers combination method in improving the object-oriented classification accuracy of PolSAR images. |
PDF Full Text Request