부산대학교 도서관

The deep learning methods have started showing promising results for spectral unmixing. We observe that many of them need direct supervision in the form of unmixed components, which is scarcely available in practice. At the same time, some require extensive data for supervised training in order to handle high implicit noise. In this letter, we propose a two-stage fully connected self-supervised deep learning network for alleviating these practical issues in performing blind hyperspectral unmixing. Given the data, the first stage (inverse model) jointly estimates the endmembers and abundances, whereas the second stage (forward model) learns the physics of hyperspectral image acquisition. The central idea is to reconstruct the hyperspectral input vector using estimated endmembers and abundances at the inverse model, which best presents the input vector’s underlying physics to the forward model. The network is jointly optimized against a two-stage loss function (in measurement domain) during the training, and we decouple the second stage at the time of inference. AdamW is used to optimize the loss function, while ReLU with a dropout of 0.3 is employed to avoid possible overfitting. The proposed network requires only the sensed hyperspectral data and learns the unmixing function from data itself, even at a lower signal-to-noise ratio (SNR). Experiments are conducted on synthetic data at different SNRs and two real benchmark hyperspectral data. The efficacy of the proposed model is evaluated and compared with the state of the art both qualitatively and quantitatively.