POLAR Classification Based On Objectoriented SVM And Spectral Clustering

Polarimetric Synthetic Aperture Radar(Pol-SAR) is a Radar Imaging Systems which work on All-weather, with multi-parameters and multi-channels,it can obtain polarimetric scattering information of targets at special wave length and visual angle.Compared with traditional Synthetic Aperture Radar(SAR),Pol-SAR contains more polarimetric scattering information of the targets,Moreover,The dimension and complexity of Pol-SAR data is relative high, So it has become a research focus in the field of Pol-SAR about how to classify the Pol-SAR data with available techniques efficiently and precisely.In order to overcome the difficulty of time complexity is too high in the process of classify Pol-SAR data with traditional methods,This article proposed a support vector based clustering method.Since Pol-SAR has too much speckle noise,which has significant negative impact on later classification,This article proposed object-oriented method which based on model of speckle noise to classify Pol-SAR data, Details are as follows:(1). This paper presents Polarimetric SAR classification based on object-oriented SVM algorithm。Traditional SVM has advantages in accuracy and speed,but it will be significant affected by Speckle noise when conduct classification,This article conduct an effective combination between area-based and pixels-based methods of Pol-SAR classification.First of all,The coherent matrixs T of the Pol-SAR are classified by SVM to get initial classification,Then overfitting the pauli features of the Pol-SAR with object-oriented method,Finally,Voting the initial classification result on the overfitting image to get a secondary classification.Since the efficient use of scattering and spatial information,This method has a series of advantages,such as unaffected by speckle noise, high accuracy and perfect boundary.(2). This paper presents Polarimetric SAR classification based on object-oriented spectral clustering algorithm.Traditional object-oriented method can conduct overfitting of image,but how to merge the overfitted image is quite a problem. Though spectral clustering has a good performance of classifing Pol-SAR data,when the data is too large,time complexity is too high and it is easy to cause memory overflow.This article proposed a method to solve this problem,Firstly,Overfitting the Pol-SAR data with object-oriented method,which do help in dimensional reducing,Then regard each bolck of the overfitted data as a object and cluster those object with spectral clustering method,Finally,Classify the data with object as unit.Since this method execute dimensionality reduction with object-oriented method, The time complexity is greatly reduced,And because of clustering with objects, Effection of speckle noise become invisible。(3). This paper presents Polarimetric SAR classification based on spectral clustering and SVM algorithm. Spectral clustering is based on spectral graph theory,Compared with traditional methods, Spectral clustering have many advantages,such as it can be conducted in any sample space,It can converge to the global optimal solution,And it is not sensitive to irregular data and so on.Because the amount of Pol-SAR data is very large,it is infeasible to solve the similarity matrix directly,this article proposed a method,which reduce dimensionality firstly,and then cluster the data.Firstly, Select a small mount of samples and train them with fast SVM for support vector,Then use spectral clustering to cluster support vector and get the cluster centers,Finally,Calculating the distance between other samples and cluster centers and finish the final classification.Compared with SVM.This method has a significant promotion in accurary,And it solved the problem of memory overflow and computational complexity is too high when the amount of data is too large.