Dual-Branch Autoencoder with Clustering Information for Hyperspectral Blind Unmixing

Deep learning methods, especially autoencoder-based methods, have become increasingly prevalent in the domain of hyperspectral blind unmixing. However, many models still have a limitation, that is, they cannot fully extract spatial and spectral information from hyperspectral images. In this paper, a dual-branch autoencoder with clustering information is proposed called DBA-CIU, which can fully utilize the spatial and spectral information of hyperspectral and add a priori clustering information in the hyperspectral image. First, we design a multi-scale convolution to extract spatial information and a band-wise split convolution to extract spectral information. Second, embedding prior clustering information of the hyperspectral image in two branches enhances information utilization. Lastly, sparse regularization is applied to the abundances obtained from the former branch. Experimental results validate the superior competitiveness of our proposed method over alternative approaches in the task of unmixing.