Classification Of Polarimetric SAR Data By The Combination Of Support Vector Machine Classifier And Decision Tree Classifier

Synthetic aperture radar (SAR) plays an important role in the field of remote sensing because of its all-day, all-weather, penetrate fog area unique advantages. Now with the rapid development of the SAR hardware technology, as well as the improvement of the theory, research on SAR data is becoming a hotspot in research of remote sensing. Most SAR are equipped with horizontal polarization antenna (H) and vertical polarization antenna (V), according to the emission and receiving polarization pulse polarization pulse mode can be divided into HH polarization (horizontal polarization transmitting and receiving) of horizontal polarization, VV polarization (vertical polarization transmitting and receiving) of vertical polarization, HV polarization (horizontal polarization launch and vertical polarization receive), the VH polarization (vertical polarization launch and horizontal polarization receive). According to the polarization modes, the SAR can be divided into single polarization SAR, double polarization SAR and full polarization SAR. Usually called full polarization SAR polarization SAR (polarimetric synthetic aperture radar, PolSAR). PolSAR contains HH polarization, HV polarization, the VH and VV polarization. Because of PolSAR records in the form of scattering matrix return wave of a variety of polarization mode object so you can get more feature information by PolSAR data.Research on object-oriented PolSAR data classification is introduced in this paper. In recent years, the classification of the object-oriented method is accepted by more and more people, especially for PolSAR which is seriously troubled by speckle noise. Object-oriented classification is equivalent to multi-look processing for objects within the data to reduce the speckle noise on the influence of a single pixel. That's accuracy is significantly higher than that of the pixel-based classification method. Object-oriented PolSAR data classification generally includes constructing polarization characteristics, image segmentation, and selecting classifier.Polarization characteristics used for PolSAR data classification mainly comes from a variety of polarization decomposition. Decomposition based on Kennaugh matrix includes Huynen decomposition and Barnes decomposition. Decomposition based on the characteristic vector includes the Cloude decomposition, decomposition of Holm, and van Zyl decomposition. Coherent decomposition includes the Pauli decomposition, Krogager decomposition and Touzi decomposition. Decomposition based on the decomposition model includes Freeman-Durden decomposition, Yamaguchi decomposition, and Neumann decomposition. In this paper, characteristics for experiments are from the polarization of the total power and coherent matrix. Classification experiment using polarization characteristics was carried out in section2.3of this paper, and the accuracy of evaluation and analysis is made for the classification results. Object-oriented classification method must segment the image to obtain each object. Watershed segmentation method was used to acquire the object in this paper. Improvement for graient pre-process of watershed segmentation is made to strengthen the segmentation results.Each channel of the RGB image is compared with the threshold value of the gradient to the final gradient value of the whole image. In addition, the watershed segmentation generally exist the over-segmentation problem. The initial segmented regions need to be merged, while mergering the region requires a given threshold to terminate the merging process. Aiming at this problem, this paper proposes an adaptive threshold value method to determine the region merging color distance threshold.This article first analyzes the distribution of each feature color distance of a polarization data, and find that the color of each feature distance approximate to a normal distribution. Also in ideal conditions, the color of the neighboring areas of the same object should be the same, namely the color distance should be0, so the color of the actual data of the same object in a nonzero distance can be regarded as a kind of error. This paper obtains probable error of every object through theory of errors. The probable error of each feature is though as the appropriate threshold of each feature. Finally, weighted approach for all appropriate thresholds is used to obtain a threshold of all features.Support vector machine classifier has high classification accuracy. However, polarization characteristics are so many that support vector machine classifier can not adaptive to choose the appropriate features, which results in the decrease of classification efficiency. In this paper, the classification method of the combination of the decision tree classifier and support vector machine classifier is proposed. Firstly, decision tree classifier can choose the best for the classification of polarization characteristics through mining the potential relationship between data, and then the polarization characteristics are used to train support vector machine classifier.The proposed method not only ensures the precision of the classification results, but also raises the efficiency of classification.This paper implements a set of object oriented PolSAR data classification process. Watershed segmentation was improved, and an adaptive threshold method was proposed to get the threshold to merge the segmented regions. Finally, the method combing the decision tree classifier with the support vector machine classifier was proposed to resolve the ineffient of the support vector machine classifier.