Hyperspectral unmixing of autoencoder based on attention and total variation

In recent years, methods based on autoencoders (AE) in deep learning have received extensive attention for hyperspectral unmixing. The purpose of hyperspectral unmixing is to estimate terminal members and their respective abundances. This is similar to the learning process of an autoencoder, which is trained to find a set of low-dimensional hidden layers and combine them with their corresponding weights to reduce the reconstruction error. Therefore, AE is well-suited to solving the problem of unsupervised hyperspectral unmixing. Aiming at the problems of being unrobust to noise and the unmixing accuracy to be further improved, this paper proposes a convolutional autoencoder unmixing network (CAA-Net) based on attention mechanism. First, an attention mechanism is introduced to improve the unmixing performance. Then, a total variation regularization term is introduced to exploit spatial information and facilitate piecewise smoothness of abundance maps. The paper conducts experiments on the Samson dataset and Jasper dataset, and compares with other classical methods to obtain higher accuracy.