A Multi-Tasks Autoencoder Hyperspectral Unmixing Model With Information Gain Based On Graph Network

Hyperspectral unmixing is a hyperspectral image processing technique that decomposes mixed pixels into pure spectral components (endmember) in percentage proportions (abundance). In the context of hyperspectral image processing and analysis, it is worth noting that complex mixed terrain scenes, such as urban areas, can pose significant challenges to achieving accurate unmixing results. This is primarily due to the high levels of mixing and low information content inherent in such scenes. To address this issue, we introduce additional urban auxiliary information and propose a multi-task autoencoder unmixing model based on graph networks. Our framework first utilizes urban auxiliary data to enhance the information of hyperspectral images. Specifically, it performs superpixel segmentation on complex urban environments and selects different autoencoder unmixing methods based on the segmentation results. Secondly, we employ graph networks to retrieve similar blocks in the image and explore additional information present in hyperspectral images. In our experiments, we specifically considered a hyperspectral unmixing dataset with urban auxiliary data. The results demonstrate that our method achieves good performance even in complex urban environments.