Cross-Domain Hyperspectral Image Classification Based on Generative Adversarial Networks

When classifying hyperspectral images, the limited number of samples will make classifier training difficult. Cross-domain information can help solve the problem of insufficient training samples. Cross-domain classification problem is discussed in this paper. In the problem, one scene with sufficient labeled samples is called source scene, and the other with limited training samples is called target scene. However, due to the changes in the imaging condition, the source scene and the target scene usually contain different feature distributions. This paper proposes a heterogeneous transfer learning method for cross-domain hyperspectral image classification based on generative adversarial networks (GANs). The method consists of two submodules: classification submodule composed of convolutional neural networks (CNNs), and scene alignment submodule that helps in reducing the domain shift between different datasets with a generator and a discriminator. The experimental results on two cross-domain hyperspectral image datasets reveal the excellent competitiveness of the proposed method.