DEEP LEARNING IN HYPERSPECTRAL UNMIXING: A REVIEW

In remote sensing, hyperspectral unmixing is very challenging inverse ill-posed problem which does not have closed-form solution. Since more than three decades, several researchers and practitioners have proposed remarkable algorithmic approaches for estimating endmembers and abundances from the data. Typically they construct a forward model (function) and then invert it for the unknowns. Meanwhile a few researchers have reported the use of artificial neural networks for unmixing. With the growth in computer technology, recently the researchers have started exploring deep learning to handle this critical issue. In deep neural networks, the underlying function is learned either by supervised training or by learning the hidden structures from the data. In this paper (within space constraint), we review the learning-based approaches and provide deep learning perspective to the spectral unmixing. We discuss spectral unmixing using three deep learning architectures, viz., autoencoders, convolutional neural network, and generative model. Further, we highlight unattended research problems in the unmixing and brief about possible use of deep learning. Finally, we attempt to discuss some of the practical challenges to be addressed and possible opportunities.