Recently, deep learning-based algorithms have been widely used for classification of hyperspectral images (HSIs) by extracting invariant and abstract features. In our conference paper presented at IEEE International Geoscience and Remote Sensing Symposium 2018, 1-D-capsule network (CapsNet) and 2-D-CapsNet were proposed and validated for HSI feature extraction and classification. To further improve the classification performance, the robust 3-D-CapsNet architecture is proposed in this article by following our previous work, which introduces the maximum correntropy criterion to address the noise and outliers problem, generating a robust and better generalization model. As such, discriminative features can be extracted even if some samples are corrupted more or less. In addition, a novel dual channel framework based on robust CapsNet is further proposed to fuse the hyperspectral data and light detection and ranging-derived elevation data for classification. Three widely used hyperspectral datasets are employed to demonstrate the superiority of our proposed deep learning models.

Recently, deep learning has been widely applied in hyperspectral image (HSI) classification since it can extract high-level spatial–spectral features. However, deep learning methods are restricted due to the lack of sufficient annotated samples. To address this problem, this letter proposes a novel generative adversarial network (GAN) for HSI classification that can generate artificial samples for data augmentation to improve the HSI classification performance with few training samples. In the proposed network, a new discriminator is designed by exploiting capsule network (CapsNet) and convolutional long short-term memory (ConvLSTM), which extracts the low-level features and combines them together with local space sequence information to form the high-level contextual features. In addition, a structured sparse $L_{2,1}$ constraint is imposed on sample generation to control the modes of data being generated and achieve more stable training. The experimental results on two real HSI data sets show that the proposed method can obtain better classification performance than the several state-of-the-art deep classification methods.

In this paper, we propose two novel classification frameworks for hyperspectral image (HSI) based on capsule network (CapsNet), which could address the drawbacks of convolutional neural network (CNN) and problem of limited training samples by introducing affine transformation matrix. Specifically, the proposed framework first performs the classification of HSI based on spectral information. Second, considering the importance of spatial information for HSI processing, we integrate the spatial and spectral information into the proposed framework to further improve the classification performance. Experimental results on real HSI data demonstrate the effectiveness of the proposed framework.

Due to the great benefit of rich spectral information, hyperspectral images (HSIs) have been successfully applied in many fields. However, some problems of concern also limit their further applications, such as high dimension and expensive labeling. To address these issues, an unsupervised latent low-rank projection learning with graph regularization (LatLRPL) method is presented for feature extraction and classification of HSIs in this paper, in which discriminative features can be extracted from the view of latent space by decomposing the latent low-rank matrix into two different matrices, also benefiting from the preservation of intrinsic subspace structures by the graph regularization. Different from the graph embedding-based methods that need two phases to obtain the low-dimensional projections, one step is enough for LatLRPL by constructing the integrated projection learning model, reducing the complexity and simultaneously improving the robustness. To improve the performance, a simple but effective strategy is exploited by conducting the local weighted average on the pixels in a sliding window for HSIs. Experiments on the Indian Pines and Pavia University datasets demonstrate the superiority of the proposed LatLRPL method.

Hyperspectral unmixing is significant for advancing remote sensing applications, aiming at extracting the spectra of pure materials (called endmembers) and obtaining their proportions (called fractional abundances) from an observed hyperspectral image (HSI). Nonnegative matrix factorization (NMF) is a popular technique for hyperspectral unmixing; however, it does not effectively preserve the spatial and spectral correlation of HSIs and fully characterize spectral variability. To overcome these limitations, we propose a novel superpixel-based autoencoder-like nonnegative tensor factorization (SANTF) model for hyperspectral unmixing. Specifically, drawing inspiration from the architecture of autoencoder, an autoencoder-like nonnegative tensor factorization (ANTF) model is constructed to directly project the hyperspectral data into abundance space. To further exploit the local spatial information, superpixel strategy is incorporated into ANTF framework, thereby building the superpixel-based ANTF (SANTF) model. Note that the superpixel cubes are jointly factorized instead of being represented as average. Meanwhile, each superpixel cube generates one endmember matrix to capture spectral variability. Subsequently, a robust double weighted endmember (rDWE) constraint is designed to obtain the consensus endmember matrix adaptively. Moreover, tensor nuclear norm (TNN) constraint is employed to enhance the low-rank characteristic of abundance tensor. Experimental results on both synthetic and real HSI datasets, compared with several unmixing methods, demonstrate that the proposed SANTF method can achieve superior unmixing performance.

Low-rank representation (LRR) is widely utilized in image feature extraction, as it can reveal the underlying correlation structure of data. However, the subspace learning methods based on LRR suffer from the problems of lacking robustness and discriminability. To address these issues, this paper proposes a new robust feature extraction method named the weighted Schatten p-norm minimization via low-rank discriminative embedding regression (WSNM-LRDER) method. This method works by integrating weighted Schatten p-norm and linear embedding regression into the LRR model. In WSNM-LRDER, the weighted Schatten p-norm is adopted to relax the low-rank function, which can discover the underlying structural information of the image, to enhance the robustness of projection learning. In order to improve the discriminability of the learned projection, an embedding regression regularization is constructed to make full use of prior information. The experimental results on three hyperspectral images datasets show that the proposed WSNM-LRDER achieves better performance than some advanced feature extraction methods. In particular, the proposed method yielded increases of more than 1.2%, 1.1%, and 2% in the overall accuracy (OA) for the Kennedy Space Center, Salinas, and Houston datasets, respectively, when comparing with the comparative methods.