Adaptive Nonlocal Spatial–Spectral Kernel for Hyperspectral Imagery Classification

Window-based technique is an effective and widely used strategy to exploit spatial-spectral features for hyperspectral image (HSI) classification, which demonstrates outstanding performance in recent research works, such as composite kernels (CKs), joint sparse representation classification (JSRC), etc. However, the utilization of square-shaped window is liable to misclassify the pixels located around the boundaries of class and leads to over-smoothed classification performance. In this paper, in order to alleviate this problem and sufficiently utilize the rich spectral signatures, we propose an adaptive nonlocal spatial-spectral kernel (ANSSK) to reflect complex manifolds by simultaneously measuring the similarities of different homogeneous patches in kernel feature space rather than pair-wise similarities in CK methods. Meanwhile, an adaptive nonlocal (AN) strategy is proposed to automatically assign an adaptive threshold for acquiring homogeneous patches based on the estimation of available training samples, which provides a data-driven and local-based strategy for dealing with the problem caused by window-based techniques. Experimental results on three real HSIs demonstrate the effectiveness of the proposed methods, and the proposed ANSSK can be easily embedded with different classifiers, such as support vector machine (SVM), representation-based classifiers, etc.