Endmember Abundance Prediction in Hyperspectral Unmixing: the Impact of Endmember Extraction Algorithms and Self-Attention in Autoencoders

Recently, deep learning techniques have gained popularity in hyperspectral unmixing (HU). The autoencoder framework, which optimizes the pixel reconstruction loss, is the most commonly employed method. However, in order to ensure that the encoder's output accurately reflects the endmember abundances, guidance is necessary during the training process. In this paper, the impact on the predicted endmember abundances, when the encoder is trained with decoder guidance, where the decoder's weights are set using different endmember extraction algorithms (EEAs) namely VCA, N-FINDR, PPI and ATGP. Furthermore, a multi-head self-attention layer is introduced to the encoder to capture the spatial correlation of neighboring pixels. After that, the performance with and without the attention layer is examined. These methods are evaluated for two real datasets. The results indicate that the accuracy of the endmember abundances is highly dependent on the choice of endmember extraction algorithm (EEA) used.