In this paper, a dual-stream convolutional autoencoding network for hyperspectral unmixing with attention mechanism (DSCU-Net) is proposed to address the issue of excessively smooth abundance maps caused by excessive incorporation of spatial correlations during pixel spectra in hyperspectral unmixing using a convolution-based autoencoding network. First, the spatial and spectral features of the hyperspectral images are extracted using a dual-stream convolution network. Second, the extracted spatial features are reweighed using a channel attention mechanism and fused with the spectral features to ensure a balance between the spatial and spectral features. Finally, the fusion features are used to reconstruct the hyperspectral image. Furthermore, these features are sent to the backbone in the unmixing network for hyperspectral unmixing. The entire unmixing network is trained by minimizing the two reconstruction errors. Additionally, experiments were conducted on two real datasets to evaluate the performance of the proposed method. The performance of the methods was also analyzed in complex scenarios. The results show that the proposed DSCU-Net can effectively overcome the fuzziness of abundance details because of the excessive introduction of spatial correlation. Moreover, the proposed method has a better unmixing performance.