HCKBoost: Hybridized composite kernel boosting with extreme learning machines for hyperspectral image classification

Abstract

Utilization of contextual information on the hyperspectral image (HSI) analysis is an important fact. On the other hand, multiple kernels (MKs) and hybrid kernels (HKs) in connection with kernel methods have significant impact on the classification process. Activation of spatial information via composite kernels (CKs) and exploiting hidden features of the spectral information via MKs and HKs have been shown great successes on hyperspectral images separately. In this work, it is aimed to aggregate composite and hybrid kernels to obtain high classification success with a boosting based community learner. Spatial and spectral hybrid kernels are constructed using weighted convex combination approach with respect to individual success of the predefined kernels. Composite kernel formation is realized with certain proportions of the obtained spatial and spectral HKs. Computationally fast and effective extreme learning machine (ELM) classification algorithm is adopted. Since, main objective is to obtain optimal kernel during ensemble formation operation, unlike the standard MKL methods, proposed method disposes off the complex optimization processes and allows multi-class classification. Pavia University, Indian Pines, and Salinas hyperspectral scenes that have ground truth information are used for simulations. Hybridized composite kernels (HCK) are constructed using Gaussian, polynomial, and logarithmic kernel functions with various parameters and then obtained results are presented comparatively along with the state-of-the-art MKL, CK, sparse representation, and single kernel based methods.