Ideal Regularized Kernel Subspace Alignment for Unsupervised Domain Adaptation in Hyperspectral Image Classification

Abstract

This paper proposes a novel unsupervised domain adaption (DA) method called ideal regularized kernel subspace alignment (IRKSA) for hyperspectral image (HSI) classification. It first uses nonlinear projection to map the original source and target data into kernel space, then incorporates source labels into the source and target kernels by the ideal regularization strategy. In the next, the subspace alignment method is performed on the ideal regularized kernels to diminish the difference between source and target kernels. Finally, a classifier built on the source kernelized subspace data can be used to predict the target data. The proposed IRKSA method exploits both the sample similarity and label similarity and makes the resulting kernel more appropriate for DA tasks. Experimental results show that the performance of IRKSA is better than some classical unsupervised DA methods for the HSI classification.