Robust transfer joint matching distributions in semi-supervised domain adaptation for hyperspectral images classification

This paper presents a novel Semi-Supervised Domain Adaptation (SSDA) method for hyperspectral image classification. Although, SSDA methods are useful when the number of the training samples are limited, but still encounter some problems. First, the traditional SSDA methods based on kernel prediction model consider a predefined kernel for both domains without using the target samples into classifier structure, which makes the challenges for classification of the noisy and complex dataset in the target domain. Second, the previous SSDA methods only measured and decreased the Maximum Mean Discrepancy (MMD) between source and target domains in order to decrease the distribution discrepancy, which ignores the discriminative information in both domains. To solve these issues, we propose a Robust Transfer Joint Matching Distribution (RTJMD) based on both the classification error and the distribution discrepancy minimization principle. We present a generalized multi-kernel model by incorporating two Fredholm integral to find an optimal kernel. Then, we propose a Regularized Extended Maximum Distribution Discrepancy (REMDD) metric in Reproduced Kernel Hilbert Space (RKHS), which considers both the extended maximum mean discrepancy and the extended maximum variance discrepancy with Multiple Kernel Learning (MKL) between two domains. The experimental results with two benchmark datasets show that the proposed RTJMD improves the classification accuracy and generalization capabilities compared to conventional SSDA approaches even in noisy and complex cases.