Spectral-Spatial Hyperspectral Unmixing Using Deep Double-Constraints Convolutional Network

‍ ‍Hyperspectral images play an important role in the environment， military， agriculture， etc. Imaging spectrometers can acquire the continuous spectrum of each pixel in the image while acquiring images. However， due to the limitation of sensor spatial resolution and the complex diversity of ground feature distribution， many pixels may be mixed by several materials （mixed pixels） in hyperspectral images. However， these mixed pixels seriously restrict the application scope of hyperspectral remote sensing. Nowadays， hyperspectral unmixing is the most effective analytical way to deal with the mixed pixels problem. Recently， the development of deep learning has brought a significant impact on hyperspectral remote sensing and has fostered many deep learning-based unmixing algorithms. Autoencoder is an unsupervised learning tool commonly used in deep learning， and it has been widely used in the research of hyperspectral unmixing methods for constructing deep networks under its good scalability. Existing deep learning-based unmixing methods show significant advantages in hidden information mining and can usually achieve more accurate results. However， most of these methods only consider the spectral information of the ground features and ignore the spatial distribution pattern， which leads to the poor smoothness of the estimated abundance from the complex scenes and makes it challenging to meet the practical needs of engineering applications. This paper proposes a deep double-constraints convolution network （DDCCN） that can further explore and utilize spatial information to improve unmixing performances and accuracies. The new proposed method uses a convolutional network to obtain a priori information. By utilizing the properties of the Gaussian kernel， the method may better distinguish different materials and assign weights between the central image pixels. Meanwhile， this strategy can improve the smoothness of abundance. In the new network structure， we adopt the softmax as the activation function of the abundance counterpart layer to constrain the abundance estimation. In addition， the L_1/2 regularization is used in the softmax to avoid overfitting the nodes. Moreover， the L_1/2 regularization can further enhance the sparsity of the abundances. To evaluate the effectiveness and advantages of our method， we use a series of hyperspectral data （including synthetic data sets and real images） to test our approach， compared with other state-of-the-art unmixing methods. From the comparisons with other methods， it is observed that the proposed DDCCN demonstrates competitive performance. The work can provide new technical support and theoretical reference for dealing with the mixed pixels problem.