Two-Stage Domain Alignment Single-Source Domain Generalization Network for Cross-Scene Hyperspectral Images Classification

Due to differences in acquisition, domain shift is prevalent between different hyperspectral scene, and conventional classification models usually have poor generalization performance in cross-scene case. Domain generalization allows models trained in the source domain to be applied directly to the unseen domain, thus it is an important way to address cross-scene classification problems. Most of the existing methods eliminate domain shift by aligning the output space. However, due to the possible loss of information in the output space, enforcing consistency of distributions in the output space alone is not enough to ensure that distributions between different domains are correctly aligned. To alleviate the above problems, we propose a two-stage domain alignment single-source domain generalization network for cross-scene hyperspectral images classification based on generative domain adversarial networks. First, to reduce the negative impact of noise in data cube on pseudo domain generation, a spectral learning branch is introduced to guide generator training. Meanwhile, to enhance the ability of the discriminator to extract joint spatial-spectral features, a pyramid feature fusion network with double projection heads is proposed. Finally, combining the double projection head of the discriminator, we propose a two-stage domain alignment loss function to better eliminate domain shift. Several experiments on three public datasets demonstrate that the proposed method has competitive advantage in classification. The code will be available at:
https://github.com/XiaozhenWang-NUAA/TSDAnet
.