Remote sensing image fusion technology of integrated spatial improvement network and spectral adjustment network

Traditional remote sensing image fusion techniques often suffer from spectral distortion or spatial information loss. Moreover, the performance quality of deep learning-based methods is inconsistent, and they necessitate separate training processes, thereby constraining their practical applicability. A network framework for fusing different satellite remote sensing images is developed to address this issue. The framework structure mainly includes two stages. In the first stage, the research utilises Squeeze-and-Excitation Network (SEN) technology to significantly improve and optimise the geometric details of the image, enhancing the overall visual effect of the image. In the second stage, different Spectral Attention Networks (SAN) are studied to adjust the chromatic data in the SEN results. In both simulated and real data testing, the research method performs excellently on the GeoEye-1, WorldView-4, and Quick Bird datasets, outperforming the comparison models. In GeoEye-1, its spectral distortion index and no reference quality index perform the best, and its chromatic data are superior. In WorldView-4, the DS metric is the best, with an improvement of over 10% and rich geometric details. The results indicate that the research method has high image fusion quality and significant advantages in chromatic data and geometric details.