Adjacent Superpixel-Based Multiscale Spatial-Spectral Kernel for Hyperspectral Classification

Abstract

The kernel-based spatial-spectral approach has been widely used for hyperspectral image (HSI) classification in recent years, where composite kernel (CK) and spatial-spectral kernel (SSK) are the most representative methods. Unlike CK, SSK measures the similarity of two clusters in kernel space to capture the hidden manifold in HSI, which has proven to be powerful to handle the well-known sample selection bias problem. However, the cluster feature in kernel space is always handled by a spatial filter which treats a cluster as a collection of individually independent pixels. This is neither efficient nor robust. Therefore, in this paper, we propose a method which treats each cluster as the combination of several subclusters. On the basis of the assumption that each manifold within a local region is a homeomorphic subspace in Euclidean space, we consider computing subcluster-based spatial-spectral features in the original input space and measure the similarities between subclusters to construct the proposed generalized SSK (GSSK). Moreover, due to the powerful ability to preserve the object boundaries, the adjacent superpixels (ASs) under an over-segmentation condition seem to be perfect candidates that serve as the subclusters. Therefore, we propose AS-based GSSK (ASGSSK) for HSI classification. Furthermore, the superpixel-based multiscale strategy is utilized in the framework of ASGSSK (termed ASMGSSK) to further explore the multiscale structures of HSI for improving the classification performance and sidestepping the selection of an optimal superpixel scale. Compared with the classic SSK method, the proposed ASMGSSK achieves 2.2% and 3.8% higher overall accuracy on average, with only 25% and 20% computational consumption for the Indian Pines and University of Pavia datasets, respectively. Experiments on two real HSI datasets demonstrate the superiority of the proposed method.