Classification Of Polsar Images Based On Polarimetric Decomposition

Polarimetric SAR image classification，as an important part of the remote sensingimage processing, has become a research hot spot. Compared with the common remotesensing image, polarimetric SAR image is more valuable, due to its data characteristicswhich can provide more bases to deep levels image analysis. Based on the imagingtheory and data features of polarimetric synthetic aperture radar, some improvedmethods aimed to achieving a better classification performance and accuracy arepresented in this thesis.In terms of basic theory, polarization scattering mechanism and representation ofpolarimetric SAR image data are emphatically introduced. Some unique polarizationdecomposition means of polarimetric SAR image are researched to get decompositioncharacteristics reflecting scattering mechanism, which is an important basis forclassification.Some mainstream classification methods are introduced and the results arecompared in this thesis. H αclassification based on Cloude decomposition,classification methods based on Wishart distribution and the classification method ofpolarimetric SAR based on support vector machine (SVM) are included. In order toimprove the performance of the SVM classifier, the statistical region merging theory isintroduced into the classification, which is used for get pre-segmentation of polarimetricSAR image. Then classifying the image based on region and SVM. The result has beengreatly improved.In order to use more polarimetric characteristics and improve classificationperformance, the quotient space and granularity theory is introduced into theclassification of polarimetric SAR in this thesis. With the SVM classification method,different quotient spaces are constructed based on different characteristics. Then,quotient spaces with the same granularity are composed to achieve an expression insmaller granularity. The results of different composition rules are discussed. Comparedwith singer classifier and linear integration method, this algorithm improves theclassification accuracy both in terms of visual classification map and accuracy of statistics, which proves effectiveness of this improved algorithm.