Spectral-Spatial Hyperspectral Unmixing Using Double-Constraints Convolutional Autoencoder.

Recently, the development of deep learning (DL) opens a new path for achieving hyperspectral unmixing. The DL-based unmixing methods show many advantages compared with traditional unmixing methods. However, most of these methods only consider the spectral information and ignore the spatial distribution, which might lead to to the poor smoothness of the estimated abundance. To deal with this problem, we proposed a double-constraints convolution autoencoder (DCAE). The newly proposed method uses the convolutional network to capture the spectral-spatial information from hyperspectral data. By utilizing the output of the Gaussian function can distinguish between two spectral similar but different materials. Meanwhile, this strategy can improve the abundance smoothness. In the new network structure, we adopt a softmax as the activation function of the abundances constrain the abundance estimation. In addition, we add L 1/2 regularization in the softmax layer to avoid overfitting when completing nodes embedding(NE). Experiments with real datasets verify the effectiveness and competitiveness of our DCAE.