This is published under the terms of CC BY-NC.

Downloaded: 355 times

Abstract

The pansharpening task is to fuse low-resolution multispectral (LRMS) images and high-resolution panchromatic (PAN) images to generate high-resolution multispectral images. Most of the existing methods do not preserve spatial and spectral details well, which is due to ignoring the difference in resolution between the two images. To address this issue, we propose a novel fusion network (ESAFormer) that effectively enhances the spatial and spectral information representation. In the proposed model, a hybrid multiresolution structure of CNN and Transformer is deployed to allow the features of LRMS images and PAN images to fuse progressively. Subsequently, the enhanced spatial attention module is adopted to preserve spatial details and long-range information. Extensive experimental results indicate that the proposed method is superior to existing SOTA methods on World-View2 and IKONOS datasets.