An Iterative Networks Based on Homogeneous and Heterogeneous Local-Global Interaction Modules for Multispectral Image Panchromatic Sharpening

Panchromatic sharpening lays the foundation for subsequent remote sensing image interpretation tasks by fusing spatially detailed panchromatic images from different sources and multispectral images with redundant spectral information to obtain multispectral images with high spatial resolution. Existing panchromatic sharpening methods can achieve good fusion results using advanced techniques such as convolutional neural networks and Transformer. However, these methods extract local and global information in a coarser way, ignoring the mutual information throughout the homogeneous and heterogeneous features, which makes the fused image distorted. To solve this problem, we propose a homogeneous and heterogeneous local-global interaction module, which focuses on extracting mutual information in homogeneous and heterogeneous features in panchromatic and multispectral images. It consists of two main parts: 1) local-global feature interaction in homogeneous regions and 2) local-global feature interaction in heterogeneous regions. This approach is capable of extracting and integrating the mutual information in homogeneous and heterogeneous regions in both panchromatic and multispectral images, thus realizing high-fidelity fusion. To validate the performance of the module, we constructed an iterative network based on the module. The proposed method surpasses state-of-the-art techniques, as demonstrated by both qualitative and quantitative assessments conducted on the standard datasets of Gaofen-2 and WorldView-3.