Advances in Hyperspectral Image Processing Techniques

13Model‐Inspired Deep Neural Networks for Hyperspectral Unmixing

Yuntao Qian¹, Fengchao Xiong², Minchao Ye³, and Jun Zhou⁴

¹College of Computer Science, Zhejiang University, Hangzhou, China

²School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China

³College of Information Engineering, China Jiliang University, Hangzhou, China

⁴School of Information and Communication Technology, Griffith University, Nathan, QLD, Australia

Hyperspectral unmixing is a significant tool to learn the material constitution and distribution of a scene. The traditional model‐based unmixing approaches rely on the definition of physical mechanism, introduce mathematical descriptions of the underlying mixing process, including various linear or nonlinear spectral mixture models, and then use well‐designed iterative optimization algorithms to solve these models. Unfortunately, defining an accurate model is always difficult due to the complicated reflection interaction mechanism and imaging environment. Inferencing an optimal solution is also difficult due to the exhaustive hyperparameter selection and time‐consuming optimization procedure. Data‐driven learning‐based approaches motivated by the great success of deep learning are becoming a new trend in hyperspectral unmixing. Various supervised and unsupervised learning‐based unmixing networks have been proposed, typically including convolutional neural networks (NNs) and encoder–decoder networks. These approaches ...