Research On Some Key Techniques For Fully Polarimetric SAR Image Classification

With the development of airborne and satellite remote sensing, the application requirements are getting more and more urgent, the conflict between large quantity of remote sensing data and relatively low information processing capability becomes more and more apparent. Machine learning method has been widely used on large-scale data processing, signal processing, feature extraction and remote sensing image interpretation and other areas for its advancement, intelligence and other technical characteristics. Fully polarimetric SAR is a hot research area in remote sensing field now. It has a lot of advantages, for example, it is not affected by time, and is able to image earth day and night, it can resist the influence of weather conditions, and can image on almost all kinds of special weather conditions. Recently, fully polarimetric SAR has received wide attention. Information processing techniques relevant to it are developing continuously, such as imaging, filter, feature extraction and classification algorithms.This dissertation focuses on some key issues on feature selection, classification of small training sets, multi-feature integration, multiple classifier combination and object-oriented classification which are important in fully polarimetric SAR image classification domain. The main content and conclusion can be summarized as follows:1) Common polarimetric target decomposition algorithm, including H/Alpha/A decomposition, Freeman-Durden decomposition, Yamaguchi decomposition, Pauli decomposition, Krogager decomposition, MCSM decomposition, Vanzyl decomposition, are compared by classification experiments using frequently-used remote sensing image classification algorithms such as minimum distance classification algorithm, Mahalanobis distance classification algorithm, neural network classification algorithm, support vector machine classification algorithm. The experimental results indicate that Pauli decomposition can extract the polarimetric features which are more helpful for improving the accuracy of common used classification algorithms.2) For classification of small training sets, a new semi-supervised learning and classification algorithm is proposed by combining Wishart distance and image segmentation. Pixels within the object created by segmentation are considered as candidates, and Wishart classification algorithm is used to preliminarily classify and sort the candidates to get reliable samples in the object. The reliable samples are then added to the training set to fulfill the semi-supervised learning and classification. Classification experiments are conducted to validate the efficiency of the new proposed semi-supervised learning algorithm.3) Improve conventional support vector machine from kernel function and apply it on the classification of fully polarimetric SAR images. Wavelet support vector machine is imported to the domain of classification for fully polarimetric SAR images. Experiments are conducted to verify that wavelet support vector machine can get the higher classification accuracy than common support vector machine. A simple multiple kernel learning algorithm is proposed by constructing textural kernel on texture information, and polarimetric kernel on polarimetric information created by Pauli decomposition, and combining textural kernel and polarimetric kernel to form multiple kernel. Multiple kernel support vector machine is constructed using this simple multiple kernel learning algorithm. Classification experiments using single kernel and multiple kernel support vector machine are conducted and the results show that multiple kernel support vector machine composed by same or different kernel function can get higher classification accuracy than single kernel support vector machine.4) To make full use of the diversity among different classifiers and improve the classification accuracy, a new classification algorithm based on ensemble learning is proposed. Wishart-KNN classification algorithm, Wishart classification algorithm are used to classify polarimetric coherency matrix information to form two classifiers, kernel-KNN classification algorithm is used to classify texture information to form another classifier. An improved classifier dynamic selection algorithm is used to conduct the ensemble learning. The experimental results turn out that this new proposed algorithm can improve the classification accuracy on the basis of single classifier.5) For problem of low classification accuracy of object-oriented classification algorithm, an improved object-oriented classification algorithm is proposed. Firstly select samples at random within the object created by segmentation to form the sample set of the object, secondly classify the sample set and count the proportion of all classes, if dominant class does not exist, it means heterogeneity of this object is high, so the object is classified pixel by pixel. Otherwise, this dominant class is considered as the class of the object. Thirdly the results are combined together. Classification experiments show that this improved algorithm can make full use of the advantages of object-oriented and pixel-based classification algorithms, and can gain more homogeneous and high-precision classification result.