Due to factors such as low spatial resolution, microscopic material mixing, and multiple scattering, hyperspectral images generally have problems with mixed pixels. This paper proposes two network structures under the framework of deep learning, which can be well applied to hyperspectral unmixing: 1) network architecture based on spectral information, the architecture uses a fully connected neural network and the spectral vector is used as an input for unmixing; 2) network architecture based on spatial-spectral information, the architecture further combines the convolutional neural networks to fuse the spatial information and spectral information of the hyperspectral image for unmixing. Experiments on simulated dataset and real dataset show the efficiency of our approach.