Study On Polarimetric SAR Images Classification With Small Samples

Synthetic aperture radar(SAR)system is a multi-channel and multi parameter radar imaging system.Compared with traditional optical remote sensing,it has all-weather and all-weather advantages.Moreover,Polarimetric SAR(Pol SAR)images record four different polarizations because their multi-channel characteristics to obtain more target information and higher resolution image information,which plays an important role in remote sensing application,such as Natural disaster monitor,urban planning,classification of vegetation types,geological exploration,military exploration and ocean monitoring.In recent years,as one of the important tasks of Pol SAR data understanding and interpretation,the terrain classification of Pol SAR images has attracted more and more researchers' attention.Based on the National Natural Science Foundation of China(Pol SAR image classification based on co-training and sparse representation,No.61173092),and the National Natural Science Foundation of China(Pol SAR image classification based on generative adversarial network,No.61771379),This paper study the Pol SAR image classification method for small sample problems based on semi-supervised learning and deep learning.The main content of this paper is as follow:(1)According the problem of Pol SAR image classification of unlabeled samples,we proposed a new unsupervised classification for Pol SAR image based on improved CFSFDP.This method can accurately estimate the number of classes for Pol SAR images,and not require to set the number of terrain classes by man-made.Firstly,the Pol SAR image is initialized into many small clusters by using the complementary information from Yamaguchi decomposition and distribution characteristics of data.Secondly,the improved clustering by fast search and find of density peaks,named as improved CFSFDP algorithm,is introduced to select the appropriate category number.Finally,to improve the representation of each category,the Pol SAR data set is classified by an iterative fine tuning process based on a complex K-Wishart function.(2)According the problem of Pol SAR image classification of the small samples,a novel semi-supervised classification method based on an improved co-training process is proposed for Pol SAR image classification.In traditional semi-supervised learning,it is difficult to select the high reliability samples because that the small labeled samples can not represent the distribution characteristics of the image.In order to increase the reliability of the selected samples,using the similarity principle and superpixel method,we proposed a new samples selection method based on the pre-selection method.In this method,a pre-selected unlabeled sample set is established by using the similar information of labeled samples and superpixel segmentation method.Then,the improved co-training method is executed based on the pre-selected samples.Through the repeatedly training process,the more and more unlabeled samples is added to the training sample set.Finally,the trained classifier is used to classify the Pol SAR image.The proposed method can not only select the high reliable unlabeled samples,but also save time cost because the number of pre-selected samples is far less than the number of all unlabeled samples.(3)According the problem of multi views constraints are difficult to satisfy in co-training method,combing the Tri-training method with the semi-supervised classification method based on improved co-training,we proposed a more simply semi-supervised classification method of Pol SAR image based on the improved Tri-training method when labeled samples are few.The proposed method is mainly composed of four aspects: Firstly,we stacked several polarimetric features based on multiple target decomposition method to construct high-dimensional feature.Secondly,a new feature selection method based on the ratio of between class scatter and within class scatter is proposed to reduce the redundant feature;Third,an improved Tri-training process is executed to train three different classifiers;Finally,the Pol SAR image is classified by these classifiers,and the major voting and the confidence are utilized to decide the final class labels.(4)Through analysis of the co-training method and Tri-training,we found that the aim of independent views in co-training and three base classifiers in Tri-training is to produce the sufficient diversity.The diversity is a prerequisite for the effective operation of these two methods,which directly affects the reliability of selected samples.Therefore,from the point of increasing diversity,we proposed a new semi-supervised Pol SAR image classification based on neighborhood minimum spanning tree(NMST).The proposed method not only adopts the multi classifier cooperative training strategy of Tri-training,but also introduce the multiple feature subset.Therefore,the diversity of the proposed method contains two parts: in one part,diversity is caused by using different classifiers;in another part,diversity is produced by different feature subsets.What's more,in order to increase the reliability of selected samples,we propose a new sample selection strategy NMST,by exploiting the local spatial information and minimum spanning tree(MST).Finally,the Pol SAR image is classified by these classifiers,and the major voting and the confidence are utilized to decide the final class labels.This method can effectively avoid the effects of the superpixel segmentation method in the previous chapter,and has the stronger robustness.(5)Spatial information is an important information for Pol SAR images classification,which can well improve the classification results.In order to make full use of spatial information o Pol SAR image,a novel width convolution neural network(WCNN)based on weighted space is proposed for Pol SAR image classification when labeled samples are small.Firstly,Labeled propagation method is used to enlarge the labeled sample set.Then,weighted space is used to increase the role of central pixel and weaken the role of other neighborhood pixel in the pixel block to make the center pixel and its similar pixel to decide the label of center pixel.Thirdly,a width convolution neural network is proposed to obtain more spatial information by enlarging the width of CNN.Finally,the Pol SAR image is classified by the trained width convolution neural network.