Attention-based dual residual network-based for multi-spectral pan-sharpening

Pansharpening belongs to an important part in the field of remote sensing image fusion, which refers to the fusion of low spatial resolution multispectral (MS) images with high spatial resolution (PAN) images to finally obtain high resolution multispectral (HRMS) images. The existing deep learning-based pan-sharpening methods have achieved better results compared with the traditional methods, but there are still two problems: spectral distortion and loss of spatial detail information. We propose an end-to-end attention-based dual-residual multi-stage remote sensing image fusion network (ADRPN) for MS image and PAN image pan-sharpening based on an in-depth study of spectral information of MS images and spatial information fusion of PAN images, which consists of three main stages, each resembling an encoder-decoder, and cross-stage fusion to achieve channel domain feature stitching. The first two stages are the feature extraction stage, where the features of MS images and PAN images are extracted using the residual module, and the different features learned are used to guide the training of individual networks using CRFB (Cross residual feature block). In the third stage (image reconstruction), we use the CA (Coordinate attention) and SE (Squeeze-and-Excitation) attention mechanism to enable the network to more precisely locate the region of interest by the precise location information obtained, which allows the features extracted in the first two stages to be better fused with the original image, thus reducing the occurrence of spectral distortion and loss of spatial detail information. Qualitative and quantitative analyses of real and simulated data from the benchmark datasets QuickBird (QB), GF-2, and WorldView-2 (WV2) show that the method can better preserve the spectral and spatial detail information and obtain high-quality HRMS images.