Abstract: In recent years, hyperspectral unmixing (HU) has been a crucial preprocess in hyperspectral imagery analysis. Current learning-based methods of HU are built on the autoencoder framework.