Font Size: a A A

Classification Of Polarimetric SAR Imagery Based On Statistical Distribution And Spatial Information

Posted on:2022-07-16Degree:MasterType:Thesis
Country:ChinaCandidate:Z YangFull Text:PDF
GTID:2518306605471414Subject:Master of Engineering
Abstract/Summary:PDF Full Text Request
PolSAR image classification is one of the most important tasks in image interpretation.It has application value in both civil and military fields.In recent years,it has received more attention and research from scholars.Due to the interference of speckle noise,the existed methods can not make sure that the pixels at the edge of a category are classified correctly as well as that pixels at a region of the same category are consistent.Therefore,how to overcome the interference of speckle noise,extract effective features and classify are the key problems of PolSAR image classification.In general,supervised classification is easier to get better classification results than unsupervised classification,but the traditional supervised classification method based on statistical model has higher requirements for labeled samples,and the acquisition of a large number of labeled samples requires manpower and material resources.Based on the statistical distribution and spatial information,this paper studies the PolSAR image classification by solving the two problems of a small number of training samples and speckle noise interference.The main work is divided into the following two parts:(1)In order to reduce the interference of speckle noise,PolSAR image classification is realized by combining PolSAR statistical characteristics with spatial information.At the same time,in order to solve the problem of few training samples,a semi supervised active learning method is designed to obtain a reliable initial classification.The proposed method consists of two layers.The first layer is based on semi supervised active learning method,which trains SVM classifier with a small number of labeled samples to get a reliable initial classification.Different from other methods,only unlabeled samples are selected in the neighborhood of labeled samples,but a new sample selection strategy is developed for labeled and unlabeled superpixels from the perspective of information and confidence.The pseudo labeled samples with rich information and high confidence are selected from the whole image to join the training set to improve the classification accuracy.In order to get more accurate classification results,the second layer design is based on the statistical distribution and spatial information of PolSAR,WG??MRF is used to correct the results of the first level classification,and an accurate edge acquisition algorithm is designed.The complete edge is added to the Markov random field,and the boundary information and region smoothing are considered.The classification results of four groups of measured data show that the method is effective.(2)With the improvement of the resolution of PolSAR image,speckle noise affects the effect of pixel based classification algorithm,and the calculation of each pixel label also increases the amount of calculation.In order to overcome the above problems,an object-oriented small sample SGTE?MWGC is proposed in PolSAR image classification.firstly,PolSAR image is segmented by statistical region merging(SRM),and then image blocks with only one class of training samples are identified to expand the training sample set.The WG? distribution used in the third chapter is introduced into a new hybrid model,that is,the WG? hybrid model is used to model and classify each class,a group of components is used to represent the probability distribution of a single class.Then,the MWGC model is used to classify PolSAR images at the regional level.This method effectively reduces the amount of calculation in classification,and uses the spatial information between the segmented image blocks to obtain good classification results in the case of a small number of labeled samples.
Keywords/Search Tags:PolSAR image classification, WG? distribution, spatial information, mixture WG? distribution, object-oriented, Semi-supervised
PDF Full Text Request
Related items