A remote sensing image sharpening method is proposed that combines cross stage local network (CSPNet) and parameter free attention (SimAM) to address the issues of uneven spectral distribution and missing spatial details in pansharpening of remote sensing images. Firstly, CSPNet is introduced into the backbone structure to replace the residual blocks in feature extraction with ordinary convolution and skip connections, in order to alleviate gradient redundancy and improve model learning power. Secondly, add SimAM blocks to directly derive 3D weights from the features, and then reverse optimize the extracted features to enable the model to extract deeper level feature information. Finally, design a learnable subtraction parameter to control the subtraction weights, in order to highlight the edge information of the fused image. The experimental results show that the proposed method can not only improve the gradient redundancy of the model, but also further enhance the spatial spectral resolution of the fused image.