Augmented Associative Learning-Based Domain Adaptation for Classification of Hyperspectral Remote Sensing Images

Abstract

Associative learning-based domain adaptation approach is investigated for the classification of hyperspectral remote sensing images in this article. It employs the criterion of cycle consistency to achieve features that are both domain-invariant and discriminative. Two cross-domain similarity matrices based on network-generated features and probability predictions are introduced in the two-step transition procedure. The associative learning with feature and prediction-based similarity metrics is referred to as augmented associative learning (AAL). The AAL-based domain adaptation network does not require target labeled information and can achieve unsupervised classification of the target image. The experimental results using Hyperion and AVIRIS hyperspectral data demonstrated the efficiency of the proposed approach.