Instance Similarity-based Adversarial Domain Adaptation Network for Hyperspectral Image Classification

Abstract

In recent years, deep domain adaptation (DA) technology has been widely used in cross-domain hyperspectral image (HSI) classification problems. However, most existing deep DA algorithms aim to align the distributions of domains while ignoring sample information. In order to solve this problem, this paper proposes a new instance similarity-based adversarial domain adaptation (ISADA) method. ISADA combines the ideas of adversarial learning and contrastive learning to reduce domain shift from two aspects: marginal distribution and class conditional probability distribution. The adversarial training of generator and discriminator is used to reduce the difference in marginal distribution between domains, and the multi-instance contrastive (MIC) loss is used to align the category distribution between domains. Experimental results on two cross-domain HSI classification tasks show that our proposed ISADA method outperforms some existing deep DA methods.