Polarimetric Radarsat-2 Image Classification Based On Target Decomposition Theorems In Polarimetry

Multi-polarity is one of the main trends in the development of synthetic aperture radar, multi-polarization data can provide richer information about polarization compared to conventional single polarization data. The polarization information is another important source of information in addition to frequency and space, which can reflect the different scattering characteristics of the target in fullest extent in order to achieve identification and classification of the objectives.In this paper, a study is made using fully polarimetric Radarsat-2 data to scattering mechanism and classification of the main land cover in Shandong Yucheng area. First, polarization characteristics of the image are extracted using model-based polarimetric target decomposition, and then analyze and compare the effect of the target in the study area by three decomposition methods, and then choose the best decomposition method based on SVM classification experiment. On this basis, combining with other polarization characteristics such as scattering angle, polarization entropy, the total power, a classification is made using two modified SVM classifier, and then analyze and compare the effect and accuracy of the two classification methods. The main conclusions are as follows:(1) An3, Arii3, Singh4 three kinds of feature decomposition method can target Shandong Yucheng region for effective characterization. Separate from a certain kind of scattering mechanism, the dihedral angle scattering for urban areas have better representation ability, scattering body for rural area better representation ability. On natural surfaces such as arable land and water, results of three decomposition method is basically the same. However, in urban and rural areas such artificial surfaces, the presence of certain results of three decomposition of differences; this difference in urban areas show a more significant. Three kinds of decomposition are effectively reducing the appearance of negative power, An3 Singh4 decomposition decomposition and avoid over-estimate the scattering body.(2) Polarization parameters feature four categories in line with its goal of substantially the theoretical value. The total power of the four categories in descending order of the feature value of urban, rural, farmland, water, which is consistent with the theory. However, due to the characteristics of the study area, there are also situations Polarization parameters anomalies. Angle scattering four kinds of feature values in descending order of urban, rural, water, arable land, and in theory, the scattering angle is greater than the farmland water bodies. The reason is that the study area, no large area of water, the river is narrow, susceptible to both sides of the feature. Polarization entropy four kinds of feature values in descending order of rural, urban, water, arable land, and theoretically, the city polarized entropy is generally greater than in rural areas. The former reason is that the rate of planting trees in the study area is relatively high in rural areas, which affects the complexity of rural scattering mechanism.(3) An3-SVM, Arii3-SVM, the overall accuracy Singh4-SVM classification methods in the order of 82.2543%,80.9294%,83.2448%. From the confusion matrix, the misclassification is the worst mistake would be rural residents into urban residents, the three methods misclassification rates were 25.73%,30.62%,24.98%, this misclassification directly affect the overall accuracy. Given the highest classification accuracy Singh4-SVM method, better reflect the Singh4 decomposition of the study area to characterize the effect of scattering mechanisms.(4) Since the target model based on the decomposition of inherent limitations in the characterization of the existence of the feature scattering characteristics shortage by increasing polarization characteristics and improved classifiers can make up for this shortfall, proved by increasing scattering angle, polarization entropy, the total power, coherent polarization matrix diagonal elements and other information, and the use of F-SVM and MK-SVM classifier, the classification accuracy is improved to some extent. Overall accuracy is increased respectively, by 83.2448 to 86.6513 percent, up from 83.2448 to 86.5963%, increased by 3.4065%,3.3515%.