关键词: 多光谱图像, 全色图像, 全色锐化, 多尺度残差注意力网络, 损失函数

Abstract: According to the panchromatic sharpening method for remote sensing image fails to consider the multi-scale directional information and the correlation between channels of the source image, which will lead to loss of spatial information and spectral distortion. A new panchromatic sharpening method of remote sensing image based on multi-scale residual attention network is proposed.Firstly, low-resolutionmulti-spectralimages after bicubic interpolation upsampling were cascaded with high-resolution panchromatic images, then the 5-channel image obtained serves as the input. Secondly, three parallel multi-scale residual attention networks are designed to different extract features of source images from spatial and channel. Each network includes shallow feature extraction, deep feature extraction, feature fusion and feature reconstruction processes, andoutputs of three networks are jump-connected to obtain the final fusion image.Spectral angle mappper, mean absolute error and geometric gradient are added with weights as a new loss function to train parameters, improving the information representation ability of the network. At last, WorldView-2 and WorldView-3 are selected to perform comparative experiments with other 7 fusion methods, and the performance of network structure and loss function of the model are verified on WorldView-3 data set .The experimental results show that this method has significant advantages in the preservation of spectral information and spatial information.