In the hyperspectral image (HSI) classification, convolutional neural networks (CNNs)-based approaches often struggle with the scarcity of labeled samples. The letter proposes an HSI classification method based on data expansion and consistency regularization with small samples. Specifically, we leverage the pixel-pair feature (PPF) to expand the dataset, which facilitates the adequate tuning of CNN parameters and alleviates the issue of overfitting. In addition, a designed CNN structure is employed to extract discriminative features from the limited number of labeled PPFs and numerous unlabeled PPFs. The CNN is trained via minimizing the weighted sum of supervised and unsupervised losses, where the supervised loss is calculated through the cross-entropy function while the unsupervised loss is evaluated with the consistency regularization item. Moreover, reliable references required in the consistency regularization item are provided after making an exponential moving average (EMA) on the outputs of CNNs at different training epochs. Ultimately, we conduct experiments on three real HSI datasets, and the results show that the proposed approach gains superior classification accuracy compared to several existing CNN-based approaches.

Remote sensing image fusion aims to generate high spatial resolution hyperspectral images (HR-HSI) by integrating low spatial resolution hyperspectral images (LR-HSI) and high spatial resolution multispectral images (HR-MSI). This paper propose a HSI-MSI fusion method based on the Pyramid Swin Transformer (PSTF). The pyramid design of the PSTF facilitates the extraction of multi-scale information, and it incorporates the Spatial-Spectral Crossed Attention (SSCA) module. The SSCA module is composed of the Cross Spatial Attention (CSA) module and the Spectral Feature Integration (SFI) module. The CSA module uses a cross-shaped self-attention mechanism, enhancing modeling flexibility for different spatial scales and non-local structures. The SFI module integrates global spectral information and local spatial-spectral correlation by introducing a global memory block, efficiently extracting and preserving spectral information. Compared with the state-of-the-art (SOTA) methods, experimental results demonstrate the effectiveness of the PSTF method.

Multispectral and hyperspectral image fusion (MHF) aims to reconstruct high-resolution hyperspectral images by fusing spatial and spectral information. The block-item tensor fusion model is able to use endmember and abundance information to improve the quality of hyperspectral images. This paper implements image fusion based on a coupled non-negative block term tensor decomposition model. Firstly, the two abundance matrices are formed into a chunking matrix and L2,1-parametric is added as well, promoting structured sparsity and eliminating the scaling effect present in the model. Immediately after, the counter-scaling effect present in the model is eliminated by adding a L2-parametric number to the endmember matrix. Finally, the focus is on solving the noise/artifacts generated by the no exact estimation of rank in the model, and over-estimation of rank by coupling the chunking matrix and the endmember matrix together to reconstruct the matrix, adding L2,1-parameters to it to facilitate the elimination of chunks, and solving the problems using an extended iteratively reweighted least squares (IRLS) method. The experiments on the University of Pavia dataset show that the proposed algorithm works better compared to the state of the art methods.

Hyperspectral image super resolution (HSI-SR) problem aims to fuse a low-resolution hyperspectral image (HSI) with its corresponding multispectral image (MSI) to obtain a high-resolution hyperspectral image (HSR). However, the commonly used methods have some limitations. For example, the matrix decomposition method does not preserve the spatial or spectral information of the image well, and the tensor decomposition method has a high computational complexity. This paper proposes a method based on spectral subspace sparse tensor factorization (SSTF), which learns the spectral subspace from hyperspectral images, constrains this model using sparse tensor regularisation, transforms it to solve a convex optimisation problem, and iteratively optimises the problem using the alternating direction method of multipliers (ADMM). The computational complexity of the algorithm is effectively reduced while retaining spatial and spectral features. Compared with the state-of-the-art methods, experimental results demonstrate the effectiveness of the SSTF method.

Recently, graph learning methods have attracted much research attention, which uses first-order nearest-neighbor relation between pixels to construct adjacency graphs for capturing smooth abundance. However, the first-order nearest-neighbor information neglects the shared domain structure of pixels making the higher-order nearest-neighbor relation missing resulting in limited spatial structure learning. Additionally, the feature values in hyperspectral images vary greatly in different bands, tiny value contributions are easily neglected due to the statistical properties of the model, which hinders the learning efficiency of the approach. To address these shortcomings, a higher-order graph regularizer nonnegative matrix factorization with adaptive feature selection method is proposed. Specifically, we introduce pixel second-order nearest-neighbor relation in graph learning to capture the nearest-neighbor domains of pixels with missing connection in first-order neighborhoods and enhance the aggregation of homogeneous regions. Then, the first-order and second-order nearest-neighbor relation are combined to construct higher-order graph regularizer to better preserve the global spatial structure of the image. Additionally, adaptive weight is introduced in the data reconstruction to balance the effects of different feature values by adaptively selecting spectral features to enhance the contribution of tiny features. Finally, the objective function is designed by fusing the higher-order graph regularization, adaptive feature selection andL1/2sparse regularizer into the nonnegative matrix factorization framework, and the optimization algorithm of the model is derived using a multiplicative update rule. We conducted extensive experiments on synthetic and real datasets, and the effectiveness and superiority of the proposed method were verified by comparing it with several state-of-the-art methods.

Fusion of high spatial resolution multispectral images (HR-MSI) and low spatial resolution hyperspectral images (LR-HSI) of the same scene can effectively combine spectral and spatial information to obtain high resolution hyperspectral images (HR-HSI), but it can also cause spectral distortion. To address this problem, we propose a new fusion algorithm (NLLR) based on a combination of low-rank prior and observation model in this paper. In the proposed NLLR method, we incorporate nonlocal spatial similarity and low-rank prior into the fusion problem to better simulate the spatial and spectral features of HR-HSI. By extracting tensor blocks from the hyperspectral and multispectral images, performing a chunking clustering operation on the hyperspectral and mul-tispectral data respectively, and constraining the fusion model using low-rank regularization to transform it into solving a convex optimization problem, followed by iterative optimization of the optimization problem using the alternating direction method of multiplier (ADMM), which can achieve an accurate reconstruction. Experimental results show that NLLR can provide better fusion performance compared to state-of-the-art fusion models.

Transformer networks have shown impressive performance for hyperspectral interpretation. Nevertheless, the high-dimensional redundant spectral distribution of hyperspectral images (HSIs) hinders their validity of interaction between features from distant locations. In this letter, we propose the HSI-Mixer, a novel extremely simple convolution neural network (CNN), which is similar in spirit to Transformer to re-consider the remarkable inductive biases of convolutions. In specific, we construct a hybrid measurement-based linear projection (HMLP) to merge spectral signatures and spatial positions of an HSI cuboid. Meanwhile, according to the merging relations between spectral-spatial attributes, we establish both spectral and spatial Mixer blocks to separate features from a mixed volume to a pure one, across either spectral bands or spatial locations, respectively. Furthermore, our HSI-Mixer maintains the same-depth-and-resolution throughout the network. Experimental results on three benchmark datasets demonstrate that our proposal achieves promising performance, in contrast to other state-of-the-art methods. The codes of this work will be available at https://github.com/Blueseatear/IEEE_GRSL_2022_HSI-Mixer.

This letter proposes a two-phase strategy-based grouped collaborative representation classifier (CRC) for hyperspectral image classification. Specifically, a spectral correlation-based CRC (SCCR) is proposed in the first phase, which considers a regularization term that uses the spectral correlations between the test sample and training samples. The representation coefficient vector generated by SCCR is then transformed into class-specific group weights. In the second phase, we integrate group weights and spectral correlations into the CRC and propose a grouped CRC (GCRC). Experimental results obtained from three real hyperspectral data sets demonstrate that the proposed SCCR and GCRC can provide better classification performance over other state-of-the-art representation-based classifiers.

Hyperspectral image (HSI) is usually composed of hundreds of continuous bands, leading a challenge task for pixel-level classification owing to high-dimensional spectral features and insufficient labeled samples. In this paper, an adaptive neighborhood strategy based generative adversarial network with (AN-GAN) for semi-supervised HSI classification is proposed. The proposed AN-GAN approach firstly uses superpixel algorithm, e.g., simple linear iterative clustering (SLIC), to generate multiple spatially homogeneous regions. Furthermore, each superpixel is merged with its spectrally similar neighbor superpixels. Then, for the reconstructed superpixels, the limited labeled samples are used to train discriminator, and a large number of unlabeled samples are utilized to generate noise using sparse autoencoder and also used to train discriminator for purpose of improving discriminator performance. Experiments were conducted on both Pavia University and Indian Pines datasets, which show that AN-GAN could provide better classification performance comparing with state-of-the-art classification models.

Nonnegative matrix factorization (NMF) has been widely used in hyperspectral unmixing (HU) in recent years since it can simultaneously estimate endmember and abundance matrices. However, most existing NMF unmixing methods only impose geometric or statistical unilateral prior on endmember or abundance matrix, meanwhile ignore the synergistic effect of both priors. To overcome this problem, in this paper, we propose a novel geometric and statistical hybrid method, called the weighted Ly2sparse total variation regularized and volume prior constrained NMF (wL1/2TVVC- NMF).The proposed approach integrates the endmember volume, abundance sparsity and piecewise smoothness into the unified NMF unmixing framework, and imposes the two kinds of prior information to the matrix factors simultaneously. It not only captures the sparsity and smoothness of abundance map, but also enhances the endmember identification accuracy, and improves the stability of results and noise robustness. The optimization model is simply solved by the variable splitting and augmented Lagrangian algorithm. Several experiments were conducted to demonstrate the performance of proposed method.

Rotation forest (RoF) is a powerful ensemble classifier and has been demonstrated the outstanding performance in hyperspectral data classification. However, the classification task suffers from the class imbalanced problem which has been considered to be one of the most important challenges. The traditional construction method of RoF biases classifying the majority classes and ignores recognizing the minority classes samples. This letter proposes a novel adaptive ensemble method based on SMOTE and RoF with differentiated sampling rates (AdaSRoF) for the multiclass imbalance problem. The proposed method adaptively generates several balanced data sets with more diversity and less noise by using SMOTE and a dynamic data sampling ratio for base classifiers. The obtained results on two publicly available hyperspectral images show that the proposed method can get more diversity and better performance than support vector machine (SVM), random forest (RF), and RoF in multiclass imbalance learning.

In this letter, we propose a new weight-based rotation forest (WRoF) induction algorithm for the classification of hyperspectral image. The main idea of the new method is to guide the growth of trees adaptively via exploring the potential of important instances. The importance of a training instance is reflected by a dynamic weight function. The higher the weight of an instance, the more the next tree will have to focus on the instance. Experimental results on two real hyperspectral data sets show that the WRoF algorithm results in significant classification improvement compared with random forests and rotation forest.

Nonnegative Matrix Factorization (NMF) factorizes a nonnegative matrix into product of two positive matrixes, which is widely used in hyperspectral unmixing. However, the convergence speed of NMF is comparatively slower, and a large number of local minimum will be existed when it is directly adopted in the factorization of hyperspectral image mixed pixels. A modified hyperspectral unmixing method based on NMF is presented in this paper. The local linear embedding (LLE) algorithm is used to reduce the dimension of hyperspectral data. The sparseness and smoothness constraints are added into the cost function. The NeNMF algorithm is used in updating endmember matrix and abundance matrix for hyperspectral data. The results show that this method can achieve good result of classification.