Hyperspectral Image Classification Algorithm Based On Correntropy Method

In the past decades, hyperspectral image classification has become more mature and achieved very rich research results. Recently, in the Information Theoretic Learning(ITL), Santamaria has put forward the concept of correlation entropy and summarized its basic characteristics, which have aroused wide attention in academic circles. In this thesis, a complete study of hyperspectral image classification is carried out, and a new classification algorithm of hyperspectral image is proposed based on the correlation entropy framework.Due to the high dimensions and redundancy of hyperspectral image data, so before the hyperspectral image classification task, we need to extract the feature of the hyperspectral image database as preprocessing operations. The principal component analysis algorithm is good for dimensionality reduction, and the reduced dimension of the feature space restore the 99% of the original information.When hyperspectral image classification task is carried out, the correntropy-based elastic regularization representation model is established by using the correlation entropy framework and elastic net regularization. As a measure of local similarity, the correlation entropy can also be used to deal with the non-Gaussian type and impulse noise, and can be used to deal with outliers. The elastic net regularization is capable of processing data of high dimension and less labeled samples, so the model takes the advantage of mixed norm regularization term to get the sparse coefficients of the test sample in the dictionary, the sparse coefficients possess group effect, and can be the greatest degree of approximation of the original information. Eventually we will get the sparse coefficients into the correlative entropy based classifier, through the calculation of maximum nonlinear difference between the test sample itself and its approximation sample to determine its corresponding class, the final classification results are compared with support vector machine classification results.In the experimental part, considering the actual situation of hyperspectral image data possess with less labeled samples. In our experiments, the number of training samples are 5%, 10%, 15%, 20% and 25%, that is, the given training samples number is at most a quarter of all samples, experiments are done on the three commonly used hyperspectral image database, Indian pines, Salinas and Pavia university. Final experimental results show that the proposed algorithm is better than support vector machine, for example, with 10% of the training samples, in the above three databases, classification accuracy of our model followed by 85.7203%, 94.0873%, 90.4556%, support vector machine followed by 81.0154%, 93.0344%, 89.9623%.The proposed algorithm model can handle some existing problems of hyperspectral image data, such as high dimensions, less labeled samples, the redundant information and adjacent spectral correlation is higher, at the same time with better robustness, for different hyperspectral image database the algorithm can have good performance. The model also has strong generalization ability, and can be very good for other pattern recognition tasks(such as face recognition under occlusion).In this thesis, the correlation entropy framework is combined with hyperspectral image classification, and it is a new attempt. The final experiment results show that it can have a better performance. Finally, we summarize the full article, some problems also put forward as the focus of attention of the following work.