| Full polarization Synthetic Aperture Radar(PolSAR),as a kind of active remote sensor,combines the strong penetrability of microwave band and the advantage of multi-polarization Synthetic,cannot depend on the sun light and can get image data all-weather and all-time due to the ability to penetrate the clouds and the rain.The way of Multi-polarization synthetic,makes PolSAR image with more details to express target information which derived from different target decom position.These features can be used to discriminate different land cover types and make PolSAR image with better monitoring and classify capability than the traditional optical image.PolSAR image classification is one of the most basic applications of PolSAR data,which is of great significance for the research and subsequent application of PolSAR data.Traditional PolSAR image classification methods,mainly based on a single type of target decomposition method,the dimension of feature used in the process of PolSAR image classification process is relatively less and cannot make full use of the abundant feature of PolSAR image,which is the one of the most important characteristic of PolSAR data,the coherent speckle noise also has a great influence on classification results.With the development of deep learning,a great number of excellent deep learning model was used in image classification,such as DBN,AlexNet and ResNet,its classification methods based on learning the feature of image,can effectively extract more essence feature of image.In the learning process of sample image,the model automatically adjusts parameters in order to better simulate the real image.A large number of experiments in optical and SAR image show that the image classification methods based on deep learning is more accurate than that of traditional classification.The PolSAR image classification method based on deep learning model,can make full use of the multi-dimensional feature of PolSAR data,it’s helpful to improve the feature utilization rate of PolSAR data.However,not all the features are suitable for PolSAR image classification,some irrelative features may reduce the classification accuracy.In order to make full use of features and reduce the redundancy of PolSAR image at the same time,reduce the impact of inherent speckle noise and improve the classification accuracy.This paper integrates the super pixel and feature set optimization,introduce the deep Residual Network(ResNet)in PolSAR classification,the main work of the paper is state as follows:(1)Select the classification feature set.In this paper,we analyze the experimental images by multiple target decomposition methods and filter the target decomposition characteristics to filter out the noise characteristics of the target objects,so as to reduce the interference of classification independent features to the classification model,reduce the feature redundancy and make classification feature sets have better robustness.(2)Compare various radar vegetation indices.Because the experimental data contain multiple vegetation types,this paper introduces radar vegetation index characteristics to participate in image classification.By comparing and analyzing the sensitivity differences between three kinds of common used radar vegetation indices for different planting types,this paper selects the radar vegetation index based on Freeman decomposition and the radar vegetation index based on H/a/Alpha decomposition.(3)In the data preprocessing stage,the super pixel segmentation is introduced to reduce the influence of speckle noise.In this paper,the SLIC super pixel segmentation method is used to divide the PolSAR image into a specified number of super pixel blocks,and each super pixel block is classified as the classification object,and the PolSAR data is classified by the pixel neighborhood information.Experimental results show that pretreatment of experimental data with super pixel segmentation can reduce the influence of speckle noise,and further improve the classification results.(4)PolSAR image classification method based on ResNet model,we build a ResNet model to learn the characters of sample image which include pixel-based image and based on super pixel image.The experimental result shows that the classification accuracy of proposed method is higher than that of traditional classification method,and the classification method based on hyper pixel block and ResNet model can further improve the classification results. |
