Terrain Classification Of Polarimetric SAR Images Based On CRF Model

Polarimetric SAR image terrian classification is the hot spot of SAR image automatic interpretation,which has important applications in geological prospecting and environmental monitoring.However,polarimetric SAR image classification also faces many challenges,on the one hand,with the development of polarization SAR technology,the acquired polarization SAR image resolution is getting higher and higher,so that image modeling,image characterization is facing new problems.On the other hand,the uncertainty and vagueness of the image increase the difficulty of the polarization SAR image classification.CRF model is a discriminant model for posterior probability modeling directly.It does not need to observe the polarimetric SAR image,and the model can fuse the multifeatures and the context information of the fusion image to suppress the uncertainty and fuzziness of the image.In this paper,CRF model is applied to classification of polarimetric SAR images.The main work and innovation are as follows:1.The concept of CRF model and the characteristics of polarimetric SAR image classification are introduced.The experimental results show that the extracted features are distinguishable for polarimetric SAR images.2.CRF models often use stacked features to exploit multiple features,which can easily lead to overfitting and characteristic interference.To solve the above problems,this paper proposes a combinatorial conditional random field model for classification of polarimetric SAR images.The combined model obtains several different child classifiers by training the same CRF model separately,and then fuse the child classifiers with the fusion formula.Experiments show that the model can effectively solve the over-fitting and characteristic interference.3.On the one hand,to suppress the speckle noise,we need to filter the polarimetric SAR image,which leads to the decrease of the classification accuracy.On the other hand,we can not make full use of the region information of the image.To solve this problem,this paper proposes a classification model of polarimetric SAR images based on superpixel and combined conditional random fields.Firstly,the super pixel is generated by the modified superpixel segmentation algorithm based on iterative edge refinement.Then,the superpixel feature value is obtained by adaptive threshold selection algorithm.Finally,the super pixel is classified as the basic unit of the combined CRF model.Experiments show that the combination of superpixel combination of random field model can effectively improve the classification accuracy.4.It is a commonly used method to calculate the superpixel eigenvalue as the superpixel eigenvalue of the mean value of the eigenvalues of all the pixels in the superpixel.However,abnormal points such as speckle noise included in the super-pixel result in a typical drop in the characteristic value of the super-pixel.To solve this problem,an adaptive threshold sample selection algorithm is proposed in this paper.The algorithm first calculates the similarity of the pixels and obtains the adaptive threshold by similarity.Then,the pixel points with similarity less than the threshold are removed.Finally,the average of the remaining pixels' eigenvalues is the most super pixel eigenvalue.