Discriminating multiple kernel learning for joint classification of optical and LiDAR data in urban area

In this paper, two contributions are made. Firstly, we propose a discriminating multiple kernel learning (DMKL) algorithm to solve the combination coefficient of basic kernels by maximizing the separability in the kernel Hilbert space in the process of MKL. The core idea of the proposed algorithm is to find the optimal projective direction, which projects the basic kernels to a discriminating kernel, by making the minimum intraclass scatter and maximum interclass scatter. Secondly, in order to make full use of the information provided by LiDAR and optical data, a strategy for fusion of heterogeneous features is proposed. DMKL is used to excavate information of different attributes in spectral, spatial and elevation level respectively. Then, composite kernel strategy is used to make a fusion for spectral, spatial and elevation information. The experiments were carried out on a hyperspectral and a multispectral image along with LiDAR data acquired over an urban area, and the experimental results show that DMKL algorithm provides the best performance among several the state-of-the-art algorithms and the proposed strategy of fusion for heterogeneous features is effective.