Deep learning-based hyperspectral unmixing is getting increasing attention in the field of remote sensing, aiming at endmember extraction and abundance estimation at pixel scale. However, many existing deep learning-based unmixing methods are generally based on linear mixing model which ignore the complex nonlinear interactions of the light scattered. Furthermore, these methods often treat all spectral bands indiscriminately, failing to capture the characteristic difference between inter-class endmembers which is very unfavorable for endmember separation. To address these issues, we present BU-Net, a novel approach for hyperspectral unmixing based on the generalized bilinear mixing model (GBM), which is a two-stream stacked autoencoder architecture designed to enhance inter-class separability. In the encoder, we employ 3D convolutions with multiple receptive field to extract multiscale spatial and spectral features simultaneously. Addtionally, we design a novel band selection based on inter-class separability (BSICS), which identifies bands with inter-class separability (BICS) and the obtained bands are taken as an additional stream for improving performance. In the decoder, BU-Net develops a two-stream structure consisting of linear and bilinear components. This design matches the components and constraints in GBM theoretically. To further enhance separability between endmembers during training, we use the spectral angle distance between BICS and its reconstruction as a loss regularization term. Moreover, we utilize materials' representative pixels obtained in the process of BSICS to initialize endmembers, which offers effective guidance for modeling the spectral properties, ultimately leading to improved unmixing results. Experimental results on synthetic and real hyperspectral datasets show that our method outperforms state-of-the-art methods. This novel approach not only addresses the limitations of linear mixing models but also leverages the power of deep learning to enhance the accuracy of endmember extraction and abundance estimation in remote sensing applications. The codes are available at the website: https://github.com/1songwei/BU-Net .

Hyperspectral images (HSI) have spectral variability, which leads to spectral dependence in adjacent and non-adjacent regions, and this dependence is essential for the classification of regions with mixed pixels. Current neural architecture search (NAS) methods have achieved significant advantages in HSI classification, but these methods cannot capture spectral dependence in non-adjacent regions because only use feedforward connections. Meanwhile, the cost of the search process in NAS is proportional to the scale of the search space, which limits the expansion of the search space. To address these issues, we propose a sparse-coding-inspired feedback neural architecture search (SCIF-NAS) method for HSI classification. Firstly, we view HSI samples as sequences and introduce a feedback mechanism in NAS to model the spectral dependence of non-adjacent regions to mitigate the effects of spectral variation. Secondly, we design several feedforward operations according to the characteristics of HSI, to form the search space together with feedback operations. Meanwhile, a sparse-coding-inspired NAS accelerated strategy is introduced to alleviate the search time burden caused by the expansion of search space. Thirdly, we integrate center loss with cross-entropy loss to construct a hybrid loss function that helps to obtain a better classification boundary. Finally, we conduct experiments on three popular HSI benchmarks, which show that SCIF-NAS outperforms the state-of-the-art methods in HSI classification.