Semi Supervised Classification Of Polarimetric SAR Based On Sparse Graphs

Polarimetric synthetic aperture radar(Pol SAR)has become one of the most advanced sensors in the field of remote sensing.For the classification of Pol SAR images,the traditional machine learning methods are supervised learning and unsupervised learning.Supervised learning requires a lot of labels,and the processing of labels is time-consuming and labor-intensive,while the unsupervised learning always can not obtain higher classification accuracy.Semi-supervised learning combines the advantage of the supervised learning and unsupervised learning.By only a small number of labels and large unlabeled samples it can achieve a higher classification accuracy.Therefore,the semi-supervised classification method is becoming a hotspot of research.And the semi-supervised classification based on graph is one of the main methods of semi-supervised classifications.In this thesis,three semi-supervised classification methods based on graphs are proposed with the aims of reducing time complexity and enhancing classification accuracy.1.The time complexity of the conventional graph based method is high.a semi-supervised polarimetric SAR classification based on fast auction graph is proposed.The main idea is to first use the superpixel segmentation algorithm to characterize the image data,and then improve it on the basis of the auction theory.The Nystrom method is used to estimate the graph.The similarity relation between all the sample points is estimated by using the partial sample points in the image,which greatly reduces the graph construction time.And then multiple auction graphs are constructed.The labels are spread from the labeled samples to the unlabeled samples parallelly on the auction graphs.Experimental results on the real Pol SAR data indicates that the proposed method greatly reduces the time spent in graph construction and has higher classification results.2.In view of to the nonlinear characteristics of Pol SAR data,a semi-supervised classification method based on random forest is proposed.The main idea is to combine the two KNN classifiers on the basis of the basic random forest algorithm,and use co-training to develop a semi-supervised stochastic forest algorithm.The similarity between the sample points is measured by comparing the path of the sample points in the random forest decision tree.In order to reduce the time complexity of the algorithm,the annealing algorithm is used to optimize the path.Experimental results indicate that the proposed method can achieve high classification accuracy compared with conventional methods.3.Based on the conventional KNN graph.This paper presents a semi-supervised polarimetric SAR image classification method based on fast updating graph is proposed.The main idea is to first select a small number of sample points to construct a adjacency matrix,and then add the remaining sample points to the adjacency matrix by iteration.When adding each sample point,in order to reduce the computation time,a threshold is set to standardize the quality of the connection edges of each newly added sample point until the specified number is met,followed by the next point.Combining the spatial information of the image sample points,we get a final similarity matrix.Experiments on synthesized and real Pol SAR data have yielded sound results,and the time complexity has been reduced.