PolSAR Image Terrain Classification Based On Deep Learning And Spatial Neighborhood Information

Polarimetric synthetic aperture radar(Pol SAR)is an active microwave radar imaging technology.Because the Pol SAR data obtains the multi-channel polarization information of the observed target,it has richer polarization information and richer target information and characteristics.Pol SAR classification problem is an important research direction of remote sensing image analysis and interpretation.In recent years,the deep learning algorithm has made significant progress,new methods have been raised.Therefore,the deep learning has a wide range of applications in the field of image classification and identification.The deep learning model automatically learns more abstract features from the input data than the previous methods of manually extracting features,which can be better used to classify them.Based on the Pol SAR data and some methods and models of deep learning,this paper presents some new methods of Pol SAR terrain classification.First,Pol SAR terrain classification method based on dictionary learning and stack network.Dictionary learning is an unsupervised learning method that does not require tag information and is well matched with the problem that Pol SAR data labels are less and labels are costly.First,we select a certain percentage of non-tagged data for feature extraction,and then use K-SVD method to build a dictionary.Then,we get the dictionary after the operation as the network input layer and the middle hidden layer of the initial value instead of a random initialization through the ridge regression method.Finally,we also introduce a stack of network structure.Experiments show that the proposed method has better algorithm stability.Second,convolutional neural network(CNN)is used to Pol SAR terrain classification tasks.Most of the previous classification methods are based on a single pixel,and rarely consider the neighborhood information,and neighborhood information as an important image information on the classification has a great help.We are based on the pixels,select the surrounding area of a certain size of the adjacent area as the information of the pixel,the pixel unit of the class standard as the overall neighborhood block class.Experiments show that through the use of neighborhood information,the accuracy of classification has been improved accordingly.Third,CNN-based method is easy to select in the neighborhood when the boundary of the real object is easy to introduce too much interference information because of the complexity of the feature.The selected neighborhood area can not correctly represent the information of the center pixel.Based on this situation,we propose a regional consistency judgment method for different regional consistency.To take the idea of division,the use of different methods to deal with.Experiments show that the problem of direct use of CNN is improved after adding regional consistency judgment.