Polarimetric SAR Decision Tree Classification Based On Multi-level Feature Filtering

As an active microwave sensor,Synthetic Aperture Radar(SAR)has the advantages of high resolution,all-weather monitoring and strong ability to penetrate clouds and fog.In contrast to unipolarized SAR,Polarimetric SAR(PolSAR)can collect more rich polarization information of observed targets,so it can be widely used in agriculture,forestry,geology and other fields.In the classification of PolSAR images,many complex features such as polarization,texture and intensity are often taken account of,so it is particularly important to select them reasonably.On the basis of polarization decomposition and object-oriented segmentation,this thesis focuses on the feature selection and classification of PolSAR images.The main contents and conclusions are as follows.(1)A multi-level PolSAR feature screening algorithm based on Symmetric Uncertainty-Correction-based Feature Selection(SU-CFS)is designed.In order to solve the problem that SU or CFS can only screen single-level features,this thesis constructs a SU-CFS multi-level feature screening algorithm which combines polarization feature selection with attribute feature selection(texture and intensity features).On the basis of the polarization features screened by SU evaluation model,CFS algorithm is used to further optimize attribute features,and CART algorithm is used to optimize the selected features again.Given the initial dataset consisting of multiple features,a set of the features are selected based on SU,CFS and SU-CFS algorithms,respectively.The results show that using SU-CFS algorithm can select the optimial features and improve the classification accuracy.(2)A PolSAR classification method based on SU-CFS multi-level feature selection is proposed.In order to solve the problem of the decline of classification accuracy caused by feature redundancy or high dimension in PolSAR image classification.An object-oriented decision tree classification method based on SU-CFS multi-level feature selection is proposed,which improves the classification accuracy of PolSAR images,addresses the problem of a large number of fine broken spots in traditional classification methods,and overcomes the deficiency that many kinds of features are not reasonably selected in PolSAR classification.Taking Radarsat-2 and GF-3 satellite images as examples,the proposed classification method is compared with Wishart supervised classification,Wishart unsupervised classification,depolarization decomposition feature classification,texture feature classification,intensity feature classification,all feature classification,CFS algorithm classification and SU evaluation model classification.Comprehensive evaluation is made from qualitative,quantitative(four commonly used accuracy evaluation indexes: comparison accuracy,intersection and merging ratio,overall accuracy and Kappa coefficient)and time efficiency.The experimental results show that the classification accuracy and efficiency of the proposed method are better than other methods.This paper includes 21 figures,11 tables and 95 references.