Abstract: This paper proposes a multi-spectral remote sensing image sharpening method based on a deep convolutional neural network and residual network. The method addresses the problems of spectral distortion in traditional remote sensing image sharpening methods and insufficient information utilization between network layers in current deep learning-based methods. The proposed method uses the depth convolution and residual network to design the depth residual module to extract the spatial and spectral features of the deep image. Additionally, residual connections between sub-blocks are established to transmit gradient information to deeper networks and avoid gradient explosion problems, making the network more efficient. Experiments are conducted on simulated and real-world multi-spectral images from WorldView-2, and the results are compared with traditional and existing deep learning-based methods. The proposed method improves the spectral distortion phenomenon and learns deeper image features to better preserve the spatial and spectral information of the image. The proposed method outperforms the deep convolutional sharpening network method in terms of various evaluation metrics, including ERGAS, SAM, SCC, UIQI, and the global fusion quality evaluation index. The proposed method improves these metrics by 24.3%, 26.7%,6.2%,4.6% and 6.3% respectively. Subjective and objective evaluations and spectral curve also indicate that the proposed method significantly improves the spatial and spectral resolution of remote sensing images, especially under complex environmental conditions.

Key words: remote sensing image pan-sharpening method, deep learning, multispectral remote sensing image, convolutional neural network, residual network