A Spatial-Spectral Transformer Network for Remote Sensing Pansharpening

Pansharpening fuses low-resolution multispectral and high-resolution panchromatic images to generate high-resolution multispectral images. Traditional methods are prone to spectral distortion, while deep learning-based methods, despite enhancing spatial feature learning, still exhibit limitations in preserving spectral information. To address these issues, this paper proposes a novel spatial-spectral transformer network (S2TPAN), which introduces the spatial-spectral transformer (SSTG) block. The SSTG comprises spatial self-attention (SSA) and spectral self-attention with guidance (SSAG), designed to capture spatial similarities and spectral correlations in the input images, respectively. Additionally, SSAG incorporates a global spectral guidance library to enhance spectral fidelity. Experiments on multiple datasets demonstrate that S2TPAN outperforms state-of-the-art methods in both reduced-resolution and full-resolution experiments, effectively improving the fusion quality of spectral and spatial information.