Hyperspectral Image Classification Based On Multiple Kernel Learning

Hyperspectral remote sensing has become a major breakthrough and research focus in remote sensing due to its unique perspective and ability to discern different objects.Due to the continuous improvement of spectral and spatial resolution,some problems arise,such as high feature dimension,strong band correlation,and highly non-linearity caused by external interference.So,classification,which is an important and commonly used information acquisition technique in remote sensing,is challenged.Therefore,in the framework of multiple-kernel learning,the thesis focuses on some key problems of hyperspectral image classification,such as heterogeneous features,highly nonlinearity and strong band correlation.By mining the structure information and discriminant information contained in samples,the thesis designs some multiple-kernel learning algorithms which are more suitable for hyperspectral image classification from different aspects,such as kernel function combination,kernel structure optimization,sample feature optimization and so on.The main research work is as follows:Firstly,a multiple-kernel learning algorithm framework based on ideal kernel optimization is proposed to solve the problem of effectively utilizing the two heterogeneous features of spatial and spectral features in HSI.By constructing an objective optimization function that contains ideal kernel,the discriminant information contained in the ideal kernel can be incorporated into the combination coefficient optimization process.Then,the objective optimization function is constructed and solved by two different ideas: the similarity information contained in the kernel matrix and the signal sparse representation in signal processing filed.The experimental results show that the proposed algorithms can effectively integrate spatial and spectral features and have better classification accuracy.Secondly,a multiple-kernel learning algorithm framework based on adaptive kernel structure is proposed to solve the problem of design and optimization of combined kernel function structure,which is caused by highly non-linearity of class boundaries in HSI.In the thesis,we combine the basic idea of data dependent kernel with multiple-kernel learning,to adaptively optimize the kernel structure adaptively by mining sample information.On the basis of the conformal transformation of the combined kernel,a set of optimized expansion coefficients is solved to increase the volume of the elements at class boundaries,so as to enlarge the separability between classes and obtain a feature space with better structure.The experimental results under different number of basic kernels show that the classification accuracy of the algorithm is improved after the adaptive optimization of the kernel structure,especially in the case of a limited number of basic kernels.Finally,a multiple-kernel learning algorithm framework based on Mahalanobis distance metric is proposed to solve the problem of strong coupling features and high band correlation in HSI.In the thesis,we combine the basic Mahalanobis distance metric with multiple-kernel learning,using the learned Mahalanobis distance matrix,the samples can be mapped into a new feature space whitch is more suitable for classification.The obtained Mahalanobis distance matrix can not only remove the coupling relationship between features and eliminate the scale effect,but also make the samples from the same class are more closely,so the samples can be optimized.The experimental results show that compared with the multiple-kernel learning method using Euclidean distance kernel function,the algorithm based on Mahalanobis distance has higher classification accuracy,and reduces the number of support vectors by reducing the boundary of similar samples,so the training time and testing time of the classifier are shortened.