Cross-Domain Association Mining Based Generative Adversarial Network for Pansharpening

Multispectral (MS) pansharpening can improve the spatial resolution of MS images, which plays an increasingly important role in agriculture and environmental monitoring. Existing neural network-based methods tend to focus on global features of images, without considering the inherent relationships between similar substances in MS images. However, there is a high probability that different substances at the junction mix with each other, which leads to spectral distortion in the final pansharpened image. In this article, we propose a cross-domain association mining-based generative adversarial network for pansharpening, which consists of a spectral fidelity generator and dual discriminators. In our spectral fidelity generator, the cross-region similarity attention module is designed to establish dependencies between similar substances at different positions in the image, thereby leveraging the similar spectral features to generate pansharpened images with better spectral preservation. To mine the potential relationship between the MS image domain and the panchromatic image domain, we pretrain a spatial information extraction network. The network is then transferred to the dual-discriminator architecture to obtain the spatial information of the pansharpened images more accurately and prevent the loss of spatial details. The experimental results show that our method outperforms several state-of-the-art pansharpening methods in both quantitative and qualitative evaluations.