Study on hyperspectral unmixing method based on deep generative model

With the development of unmanned aerial vehicles, aviation, and aerospace Earth observation systems, the available data of hyperspectral imaging has been further expanded. Due to the limitations of spatial resolution of hyperspectral images and the complex diversity of land distribution, mixed pixels are widely present, which seriously restricts the quantitative application of hyperspectral imaging. Spectral unmixing technology is an effective tool for improving the level of quantitative application of hyperspectral remote sensing. It can identify the constituent materials and obtain their spatial distribution characteristics in the observed scene. Currently, artificial intelligence algorithms represented by deep learning have provided a new research paradigm for hyperspectral image unmixing and have achieved notable results. Most existing studies ignore the spatial information of the imagery and are not suitable for unknown nonlinear mixing scenarios. In view of this, this paper proposes a novel spectral unmixing strategy that first constructs a synthetic hyperspectral dataset with abundance labels based on Variational Autoencoder Networks, and then carries out supervised training based on convolutional networks to obtain the abundance information of the image. The implementation of this project aims to explore a dual-branch neural network with strong generalization abilities for hyperspectral unmixing, which has significant implications for improving the refined application of hyperspectral remote sensing.