Classification Of Hyperspectral Data Based On Multi-feature Combination By Multiple Kernel Boosting

The advantages of hyperspectral remote sensing, which has high spectral resolution, numerous number of bands, and the unity of graphics and spectral, opened the prelude to another revolution. Hyperspectral remote sensing classification technology is a effective geographic information technology of data mining.it is a further development of remote sensing data classification based on traditional technology, which classification technique is unique. When the sensor resolution is too low, a pixel mixed with a variety of surface features which covering mutual intermixing images caused the "synonyms spectrum", "with spectrum foreign body" phenomenon.The presence will cause uncertainty of the hyperspectral image classification, the classification uncertainties faced the biggest challenge in the remote sensing classification. We must understand the nature of these uncertainties, appropriate handling of these uncertainties, and establish a reliable and strong robust hyperspectral image classification to improve multi-class terrain classification accuracy in hyperspectral image classification. And the redundancy numerous bands brought hyperspectral remote sensing image data cause huge image data preprocessing difficult high correlation between the band brings an increase in the number of training samples and hyperspectral remote sensing data classification model using conventional estimation parameters especially suffering problems when statistical classification model; therefore design an appropriate and strong robustness suitable for hyperspectral data classification is necessary.This paper presents a new framework for the combination of multiple features by multiple kernel learning for hyperspectral image classification, which achieve the optimal combination of heterogeneous features. We construct a new family of multiple kernel learning which achieve optimal classification in large number of features and classifier when combining the spectral and spatial information contained in hyperspectral data, making full use of the complementary information. The proposed method aims at finding an optimal combination of hyper-plane regarding specific features that can best classify from a large pool of classifiers and image attributes. Rather than combining all features in the same kernel space, we distribute all features to a group of different kernels. The classification performance of features on specific kernel-based classifiers can thus be thoroughly evaluated, which leads to a more discriminative ability to the final decision function. Moreover, we utilize the boosting technique for efficiently selecting good hyper-planes from support vector machines(SVM) to reduce the computation load that is different from related methods based on global optimization.