Dual Attention Based Convolutional Neural Network with Scattering Transform Features for Hyperspectral Unmixing

Recently, deep learning has shown powerful learning ability. However, convolutional neural networks still have limitations, and it is difficult for classical deep CNN models to obtain global information and long-term dependencies of the input image (feature map) in the same layer network structure. Although the local perceptual field of the same-layer convolutional structure can be increased to some extent by the superposition of multiple convolutional modules, such a simple stacked structure often results in disadvantages such as computational inefficiency and difficulty in optimization of deep CNN models, and inspired by this, we combine scattering transform with attention-base convolutional neural network to extract deep features, while fully considering the characteristics of the endmember extracted from hyperspectral images, and propose dual attention-based convolutional neural network (DACNN). Here, the endmember bundle is added to the unmixing process to reduce the effect of spectral variability in hyperspectral unmixing, and in addition the scattering transform feature with the attention mechanism allows the network to focus on the deep features, making the unmixing process more interpretative and accurate. Experiments on two different real data sets demonstrate the superiority of DACNN over several advanced unmixing algorithms.