Multiple kernel learning based on ideal kernel optimization for hyperspectral imagery classification

In this paper, a novel multiple-kernel learning (MKL) algorithm is proposed for classification of hyperspectral images. The goal of classification is to acquire the class label of each pixel. The land covers is linearly separable in the kernel space spanned by class labels (ideal kernel). The ideal kernel is used as the optimization objective of our proposed MKL algorithm. Linear programming (LP) and signal sparse representation (SSR) are used to find the optimal weighting coefficients of basic kernels in our proposed based on ideal kernel MKL (BoIKMKL), thus leading to two variants of the proposed method, BoIKMKL-LP and BoIKMKL-SSR, respectively. Experiments are conducted on a real hyperspectral data set, and the experimental results show that the proposed algorithms, especially for BoIKMKL-LP, achieve the outstanding performance for hyperspectral image classification with few labeled samples when compared with several state-of-the-art algorithms. To a certain extent, BoIKMKL-SSR solves the problem of basic kernels redundancy with satisfactory classification accuracy.