Hyperspectral Image Classification via Spectral-Spatial Shared Kernel Ridge Regression

We propose the kernel version of the recently introduced spectral-spatial shared linear regression (SSSLR) for hyperspectral image (HSI) classification. Original SSSLR used original data space-based shared subspace learning (SL) model and spectral-spatial-based ridge linear regression (RLR) to learn a subspace projection matrix. However, HSI data sets have multivariate attributes and are often linearly inseparable, thereby limiting the classification performance of the conventional SSSLR. Hence, we introduce a modified kernel version of SSSLR algorithm [spectral-spatial shared kernel ridge regression (SSSKRR)] in which nonlinear high-dimensional feature space-based shared SL model is included into the kernel ridge regression (KRR). Finally, an efficient singular value decomposition (SVD)-based alternating iterative algorithm is used to obtain the optimal classification results. Experiments results show that the proposed SSSKRR had superior classification performance compared to the state-of-the-art SL algorithms.