Research On Multiple Kernel Learning Based Target Interpretation Technologies In Hyperspectral Imagery

Hyperspectral imagery has attracted a great variety of attention because of its fine ability to distinguish spectra. Hence it is used to sense and recognition the surface characteristics of interesting landcover. Some properties of hyperspectral imagery posses a high challenge to target interpretation technologies. Many problems have not been solved yet. This thesis starts from the ambiguity of discrimination and focuses on multiple kernel learning with its application in target interpretation. We will purchase to improve the interpretation accuracy and processing efficiency for the purpose of promoting the development of kernel methods in hyperspectral image processing.This thesis mainly studies information mining and interpretation techniques for hyperspectral imagery based on kernel machine learning theory, which include following three aspects in detail:First of all, the properties of hyperspectral image data are described in detail, and focused on analyzing the advantages and disadvantages involved with these properties, especially the impact on target interpretation. From the perspective of machine learning, it could be explained that these properties will increase the ambiguity on discriminating targets and reduce the interpretation accuracy. Then the typical single kernel learning and its improved version, i.e. spectral weighted kernel, are researched for these properties. It is denoted that there is a limitation from these methods to mining intra-class and inter-class information. And then a current direction is demonstrated from pattern sensing with single kernel to information fusion with multiple kernels.And then the basically theoretical framework of multiple kernel learning is introduced. According to the shortages of existing methods, an issue is proposed to reasonably interpret multiple kernel learning with a view of multiscale similarity measurement. Based on this point, an ensemble method is proposed on multiple kernel similarity measure which is constraint via L2 norm (noted as L2MKL), and implemented to deal with multiple-kernel learning problems. Validated via experiments on real hyperspectral images, L2MKL shows excellent performance on both classification accuracy and computational efficiency. Especially, it does not require an optimal pre-selection of scale parameters. While tackling classification problem with small sized training set, L2MKL exhibits significant advantages.Finally, multiple kernel similarity measure is extended to some other target interpretation technologies. Firstly, composite kernel method fusing spatial and spectral features on the level of kernel structure is investigated. Multiple kernel based similarity measure is embed to mine discriminative information from spatio-spectral features and improve the generalization capability of classifier in both image and spectral spaces. Secondly, semi-supervised learning based on graph Laplacian is researched. Multiscale graph similarity measure is constructed in order to modify the smooth of decision function and improve the capability of mining limited priori knowledge. Thirdly, target recognition based on support vector data description is investigated and extended. Both the compactness and generalization capability are enhanced via inducing multiple kernel mapping. All the interpretation technologies embed multiple kernel are validated effective via experiments utilizing real hyperspectral image data.