PolSAR Classification On A Small Scale Sample

Polarimetric SAR classification problem is an important research content and key technology in polarized SAR image interpretation,which is of great theoretical and practical value in civil and military fields.In the current PolSAR classification method,the supervised method is usually better than the unsupervised method,but the use of supervised methods to obtain better classification results requires a large number of training samples.However,for PolSAR,marking sample is often expensive.Therefore,the use of as few markers as possible to obtain high classification accuracy in PolSAR terrain classification has a very important significance.At the same time,with the development of deep learning theory,some methods for PolSAR classification based on deep learning have been proposed and have achieved good results.However,compared with traditional machine learning methods,this method combined with the deep learning theory requires a large amount of samples.In view of this problem,this paper starts from the reality of small samples of PolSAR data.The total work as follows:First,combining with color feature for PolSAR classification using DBN network.Based on the deep feature extraction of the deep network model DBN,the pseudo SAR feature corresponding to the PolSAR data is added to the input.Based on the Pauli decomposition,the color distribution of the image region is described by HSV,and the CSD is introduced into the image area to consider the relationship between the color of the image area.The combination of the above-mentioned color feature and DBN can obtain a better feature representing,thereby effectively improving the classification effect.Experiments show that this method can improve the classification performance when the number of samples is small(1%).Second,the PolSAR feature classification based on transfer learning.PolSAR technology has been developed for a long time.the amount of data is very much,but the public is very little.If the existing polarized SAR data can be "reused",it is helpful to classify the new difficult polarized SAR data.The only way to transfer is to apply the knowledge or patterns learned in a field or task to different but related areas or problems.This paper analyzes the influence of parameter migration on the performance improvement of the classification.The experimental results show that the transfer strategy is helpful to the depth of the learning sample.Third,PolSAR classification based on class RBM.RBM is a probabilistic model that can only generate preprocessing processes or as initial models of some other models rather than being independently a complete supervised classifier.In this paper,we discuss how to provide an independent classification framework based on RBM,mainly through the introduction of RBM variants-class RBM to adapt to the classification problem.Experiments show that proper combination of discrimination and generation of training can achieve better classification performance.