Extended-Aggregated Strategy for Hyperspectral Unmixing Based on Dilated Convolution

Autoencoder unmixing is a popular deep learning-based spectral unmixing algorithm, which decomposes the mixed pixels into pure endmembers and their fractional proportions, but the existing methods cannot fully exploit the spatial correlation features of hyperspectral image (HIS). In this letter, we propose a dilated convolution extended-aggregated strategy (DEAS), which enhances the ability of autoencoder unmixing algorithms to extract spatial correlation features. This strategy constructs a module utilizing various combinations of dilated convolutions with different scales. DEAS extracts the spatial relationships within multiple ranges around each pixel. Compared with the full connection and convolution commonly used in the encoder layers, autoencoder algorithms with DEAS expand the acceptance domain of the network. Furthermore, DEAS aggregates the spatial information in different ranges to obtain the feature map fully acquiring the relationship between pixels, which improves the unmixing performance. In particular, DEAS can be inserted into the existing autoencoder unmixing algorithms to get more abundant spatial information, and the methods using DEAS can show better unmixing effects. We apply the DEAS to two autoencoder methods using full connection and convolution, respectively. Experiments indicate the competitiveness of the algorithms using this strategy in hyperspectral unmixing tasks.