ABSTRACT
Convolutional autoencoder, which can well model the spatial correlation of the data, have been widely applied to spectral unmixing task and achieved desirable performance. However, the fixed geometric structure of convolution kernels makes it difficult to capture global context. To address this issue, strategies such as dilated convolution or transformer are often employed, but this may result in minor loss of local details. Therefore, we propose a collaborative unmixing network with a multi-scale pyramid structure to capture both global and local features simultaneously. To integrate features from different scales in the unmixing process, we employ a cross-stream fusion feature strategy, which not only promote collaborative representations but also capture long-range dependencies while preserving local details. Meanwhile, we also design the residual spectral attention mechanism to refine the features from different scales and facilitate their fine-grained fusion. In the proposed network, each convolutional stream undergoes effective collaborative training using a convolutional autoencoder structure. The collaborative strategies include cross-stream feature fusion mechanism and alternating training strategy with weight sharing for endmember information. Experiments over three real hyperspectral datasets indicate the effectiveness of our method compared to other unmixing techniques.
Acknowledgement
Thanks to the providers of the dataset. The data presented in this study are openly available at http://rslab.ut.ac.ir/data.
Disclosure statement
No potential conflict of interest was reported by the author(s).