Deep learning-based methods have demonstrated promising performance in hyperspectral image (HSI) classification. However, the black-box nature of deep learning poses a significant challenge in designing effective network architectures for HSI classification. To overcome this issue, this article presents a representation model-inspired interpretable deep unfolding network (NSR-Net). First, we formulate a deep-constrained nonnegative sparse representation (NSR) model with enhanced generalization ability to address the limitations of the prior-constrained NSR, i.e., its reliance on manual priors and specific assumptions. Second, the solving process for deep-constrained NSR is unfolded into a deep network, with each component of the network corresponding directly to a specific step. Finally, following the principle of representation model-based classification, a subdictionary reconstruction module (SDRM) is designed to determine the class label. In SDRM, each subdictionary is learned through a context-integrated training process, resulting in superior discriminative capability. In addition, to better guide NSR-Net optimization, we introduce a new composite loss function, which consists of constraint loss and residual loss, aiming to effectively recover representation coefficients and reconstruct data from the subdictionary. Experiments conducted on four distinct HSI datasets illustrate the superiority and generalization performance of the proposed method compared with advanced representation model-based and deep learning-based methods, with overall accuracy (OA) improvements of 0.72%–9.80%, 1.39%–8.09%, 0.40%–5.85%, and 0.56%–6.84% for Indian Pines, Salinas, LongKou, and Loukia, respectively. The source code will be available at:https://github.com/ZhaohuiXue/NSR-Net.

Hyperspectral image classification (HSIC) often struggles due to the scarcity of labeled samples. Knowledge distillation (KD), including self-distillation (SD) where a model learns from its own predictions, has emerged as a promising solution. However, existing distillation methods in HSIC face a “granularity mismatch” problem as they rely on coarse, patch-level data for fine-grained, pixel-level classification, which introduces label noise and causes misclassification. To overcome this issue, we propose central spectral self-distillation (CSSD), a framework that isolates pure spectral information at the patch center and leverages it for SD. CSSD consists of three main components. First, the backbone network separates spectral and spatial feature processing to extract pure central spectral features. Second, a spectral refiner module enhances these spectral features before integrating spatial context. Finally, an SD loss aligns the final predictions with the central spectral guidance, ensuring granularity matching at the pixel level. The experimental results on five hyperspectral datasets demonstrate the effectiveness of CSSD under few-shot conditions. The source code will be available online athttps://github.com/ZhaohuiXue/CSSD.

Transformer-based deep learning (DL) methods have gradually been advocated for remote sensing (RS) image semantic segmentation due to the great global modeling capability. Nevertheless, Transformer-based DL methods have not yet been sufficiently explored on the large-scale hyperspectral image (HSI) semantic segmentation. Current algorithms lack a comprehensive consideration of the impact of positional encoding (PE) interpolation when constructing Transformer-based decoders. Moreover, existing segmentation heads usually directly concatenate multiscale features to achieve segmentation, which ignores the inherent semantic differences between different features. To address the above issues, a U-shaped multimixed Transformer network (UM2Former) is proposed for large-scale HSI semantic segmentation. First, a weight encoder consisting of two modules, the overlap-down and the channel-weight, is built to extract hierarchical discriminative spectral-spatial features and decrease spectral redundancy. Second, the proposed multimixed Transformer block (MMTB) develops a PE-free module, spatial-feature-retention attention (SFRA) mechanism, in which “multimixed” represents the global dependency modeling of each pixel with the retented average spatial characteristics of different locations in the input feature maps. Finally, a linear fuse segmentation head (LFSH) is designed to align semantic information among multiscale feature maps and achieve accurate segmentation. Experiments were conducted in single cities and the entire large-scale WHU-OHS HSI dataset. The segmentation results indicated that the proposed method achieved higher accuracy compared to the existing semantic segmentation methods, with performance improvements of 17.80% and 4.16% in terms of intersection over union (mIoU) and overall accuracy (OA), respectively. The source code will be available athttps://github.com/ZhaohuiXue/UM2Former.

Deep learning (DL) has shown great potential for hyperspectral image (HSI) classification. However, DL models easily get trapped into overfitting due to limited training samples. To overcome this issue, a novel spectral–spatial triplet network (S2TNet) is proposed for few-shot HSI classification. First, a lightweight spectral–spatial network (SSN) composed of 1-D and 2-D convolution is introduced to extract spectral–spatial features. Second, a hard sample selection strategy is proposed by integrating classification and contrast training to deal with unbalanced positive and negative samples in traditional triplet networks. Third, an enhanced triplet loss function is proposed by considering the relationship between positive and negative sample pairs to ensure the distance between homogeneous samples is smaller than that of heterogeneous samples, which effectively improves the discrimination ability of the model. Experiments conducted on two widely used hyperspectral datasets demonstrate that S2TNet significantly outperforms other related methods, with 0.81%–16.83% and 1.40%–13.83% improvements (under 20 labeled samples per class for training) in terms of overall accuracy (OA) in Indian Pine (IP) data and University of Pavia (PU), respectively.

Unlike other deep learning (DL) models, Transformer has the ability to extract long-range dependency features from hyperspectral image (HSI) data. Masked autoencoder (MAE), which is based on Transformer architecture, employs a “mask-reconstruction” strategy for training, allowing the model to be effective for downstream tasks. However, existing MAE-based methods only apply spectral or spatial masking to HSI and reconstruct them for feature learning, which is too simplistic and insufficient for the model to learn robust features. Additionally, the issue of lacking labeled samples in HSI and the primary objective of MAE to reduce the reliance on labeled samples are often overlooked. To address these issues, we are inspired by diffusion-based representation learning and propose diffusion-enhanced MAE (DEMAE) for HSI classification with few labeled samples. First, an asymmetric encoder–decoder framework is constructed as the backbone by stacking both conditional and standard Transformer blocks. Second, we devise an auxiliary task aimed at simultaneous denoising and reconstruction, facilitating heuristic feature learning from HSI data. Third, the encoder of DEMAE is isolated for training with few labeled samples. Finally, the encoder is used for classification, and a novel signal-to-noise ratio enhanced (SNR-Enhanced) loss function is introduced to regularize the model training process. The performance of DEMAE is evaluated on four benchmark datasets, demonstrating its superiority in classification accuracy and mapping capabilities on unlabeled areas compared to existing state-of-the-art methods with few labeled samples. The source code will be available online athttps://github.com/ZhaohuiXue/DEMAE.

Transformer has shown great potential in extracting global features, and it can achieve better classification performance with a large number of training samples compared with other deep learning (DL) models. However, most of the existing Transformer-based models for hyperspectral image (HSI) classification simply use the multihead self-attention and channel multilayer perceptron (MLP) modules that contain many parameters to learn, resulting in poor performance in a few-shot learning scenario. To overcome the above issue, a lightweight self-pooling Transformer (SPFormer) is proposed for few-shot HSI classification. First, a one-layer autoencoder based on self-supervised learning is built to reduce the dimensionality of HSI. Second, two parameter-free modules, channel shuffle for multihead self-pooling with sparse mapping (CSSM-MHSP) and central token mixer (CTM), are proposed for mapping spectral features to higher dimensions and promoting information interaction between pixels, respectively. Third, a lightweight channel embedding is designed to extract deep spectral features. Finally, a fully connected layer is used for classification. The classification performance of the proposed method is evaluated on four benchmark datasets, showing its superiority in classification accuracy, generalization performance, and model complexity compared to existing state-of-the-art methods with limited training samples.

Hyperspectral coastal wetland classification requires an extensive quantity of labeled samples, which are hard to acquire. Therefore, a novel semi-supervised dynamic ensemble learning (SSDEL) framework is proposed to overcome the limitations of labeled samples in wetland hyperspectral classification. Firstly, a collaborative relationship is established between labeled and unlabeled samples in the sample augmentation stage. Based on this relationship, unlabeled samples were assigned to the region to which the most similar samples belonged. Then, multiple classifiers are trained using labeled samples and predict unlabeled samples in the same region to obtain higher confidence pseudo-label results. Secondly, based on the assumption that different classifiers should produce similar classification results for a specific target sample, an objective function is designed to unify the classification behavior of multiple classifiers. The representation coefficients of multiple classifiers in the same region are constrained by optimizing the objective function through thel2norm. Finally, a complete SSDEL framework is constructed by applying consistency learning again to the augmented samples. The proposed method is evaluated using three wetland hyperspectral images of China, and the experiments results demonstrate its effectiveness.

Low-rank and sparse representation (LRSR) model has gained popularity in hyperspectral image (HSI) classification. However, most existing LRSR models are limited by the highly nonlinear correlation of hyperspectral data, which leads to poor subspace segmentation performance. Furthermore, current LRSR methods usually directly used labeled samples to build the dictionary, whereas low discriminative labeled samples may degrade the representation ability of the dictionary. To solve the above issues, we propose a novel structure-prior-constrained low-rank and sparse representation with discriminative incremental dictionary (SPCLSR-DID) method for HSI classification. First, global and local data structures are maintained by low-rank and sparsity constraints, while a structural prior constraint is introduced to explore the intrinsic spectral-spatial structural information of HSI, improving the subspace segmentation ability of the model. Second, a discriminative incremental dictionary (DID) method is presented to find reliable and discriminative augmented atoms to improve the completeness and representation power of the dictionary. In DID, the incremental dictionary size is controllable to suit different tasks. Finally, the class label of each target sample is determined by jointly considering contextual information within a certain local range, which ensures the accuracy and smoothness of the classification map. Experimental results based on four popular hyperspectral datasets demonstrate that the proposed SPCLSR-DID method significantly outperforms other related comparison methods in terms of classification accuracy and generalization performance.

A novel approach is proposed in this study that combines superpixel (SP) segmentation and multiclassifier ensemble learning (EL) to address the limited availability of labeled samples in coastal wetland land cover classification. First, the SP segmentation technique is employed to partition unknown samples into multiple homogeneous regions, thereby facilitating the effective capture of spatial information pertaining to land cover. Subsequently, a multiclassifier EL strategy is employed within these regions to process the samples, effectively leading to a reduction in classification errors and an improvement in accuracy. To enhance the performance of semi-supervised learning (SSL), a sample iteration selection metric is introduced to optimize the training samples based on the consistency of sample types within homogeneous regions and the results obtained from the multiclassifier ensemble, thus enhancing the reliability of pseudo-labels. Additionally, multiscale SP segmentation is utilized to augment the ensemble strategy for samples in order to reduce the necessity for hyperparameter adjustments and increase the automation and reliability of the model. Overall, the accuracy of coastal wetland classification is improved by this approach while simultaneously mitigating the complexity of SSL in terms of hyperparameter tuning. The effectiveness of the proposed approach has been assessed through experiments conducted on three GF-5 hyperspectral images of coastal wetlands in China. In particular, the proposed methods provide superior performance compared with the state-of-the-art classification methods.

Graph convolutional network (GCN) combined with convolutional neural network (CNN) exhibits significant potential in hyperspectral image (HSI) classification. Hypergraph convolutional network (HGCN) can address the limitations of GCN-based methods in representing high-order nonlinear relationships among multiple nodes. However, the existing pixel-based HGCN methods mostly adopt partial pixels for hypergraph modeling, thus limiting the representation of the global structure. Additionally, both pixel-based and superpixel-based HGCN methods only utilize the k-nearest neighbors (kNN) for hyperedge representation, thereby ignoring the rich topological information in HSI segmentation regions. To tackle these issues, we propose an HGCN with multiple hyperedges fusion (HGCN-MHF) for HSI classification with limited samples. First, we introduce a CNN branch for capturing spatial and spectral pixel-level features with different receptive fields, which begins with denoising and spectral transformation, followed by cross multiscale convolution (CMC). Second, we design an HGCN branch for extracting superpixel-level features guided by diversified high-order hypergraph structures, which incorporates a multiple hyperedges fusion (MHF) module followed by hypergraph convolution (HGC). Finally, a score-weighted feature fusion (SWFF) strategy is proposed to balance and promote the feature fusion of the two branches. Experimental results on four benchmark HSI datasets demonstrate that HGCN-MHF outperforms other state-of-the-art methods, with improvements in terms of overall accuracy (OA) around 3.50%–20.47% (Indian Pines), 2.85%–19.83% (University of Pavia), 2.31%–8.75% (Salinas), and 3.13%–20.03% (WHU-Hi-HongHu) under five labeled samples per class.

Spectral shifts between source and target domains (TDs) pose significant challenges in cross-domain hyperspectral image classification (HSIC). Current methods often struggle to balance mitigating these shifts while preserving crucial TD information, which limits their ability to leverage spectral priors and domain-specific characteristics for accurate classification. Our work proposes a novel knowledge swap net (KSN) for few-shot cross-domain HSIC. KSN tackles the challenge by enabling effective knowledge transfer between homogeneous (spectral) and heterogeneous (domain-specific) feature spaces through a two-step knowledge swap strategy: leveraging homogeneous knowledge distillation (Homo-KD) for transferring spectral knowledge and heterogeneous meta-learning (Hetero-ML) for model refinement with TD feedback. In addition, we develop a relative distance difference (RDD) loss function to improve feature discriminability under few-shot conditions. Experiments conducted on four target datasets demonstrate the superiority of KSN. Notably, KSN achieves a remarkable overall accuracy (OA) of 82.56% on the Houston University (HU) 2013 dataset, surpassing other leading methods by 3.83%–8.98%. The source code will be available online:https://github.com/ZhaohuiXue/KSN.

In recent years, hyperspectral anomaly detection methods based on representation models has attracted much attention. However, when the dictionary is polluted by anomalous pixels, their performance is greatly affected. To adjust the contributions of different dictionary atoms, traditional methods usually predefine a distance weighting matrix and impose it on the dictionary matrix or coefficient vector, which may not be accurate enough. To solve this problem, a self-paced probabilistic collaborative representation detector (SP-ProCRD) is proposed in this article. It assigns weights for each atom loss term according to the probability that the pixel under test (PUT) belongs to the same class as each dictionary atom. Unlike the predefined weight matrix approach, a self-paced learning (SPL) strategy is used for iterative optimization, so that dictionary atoms participate in the representation from "good" to "bad" ones when solving the model. The representation residuals are utilized to accelerate the convergence. The proposed model can optimally represent each PUT using similar dictionary atoms and minimize the negative impact caused by anomalous atoms contained in the dictionary. In terms of weighting for SPL, an adaptive weighting scheme based on the polynomial self-paced (SP) regularizer is proposed to address the generalization issues of most previous weighting schemes. This scheme improves the generalization and automation of the model. Experimental results reveal that the proposed method produces more accurate result than existing methods and runs efficiently.

Graph convolution networks (GCNs) have shown great potentials for few-shot hyperspectral image (HSI) classification. Mainstream GCNs construct graph according to single scale segmentation, which usually ignores subtle adjacency relation between small regions, leading to unreliable initial local graph. To overcome the above issue, we propose a differentiated-scale restricted GCN (DSR-GCN) for HSI classification. Firstly, we propose a differentiated-scale graph construction method considering both the subtle and relative wider range spectral-spatial relation. Secondly, restricted fusion loss is designed to restrict the fusion of features extracted with differentiated-scale GCN branches. Finally, we design a lightweight spatial-spectral siamese network to remedy local pixel-level features. The proposed DSR-GCN can better model spatial structure with a reliable and refined graph, and it can capture more discriminate features in few-shot learning (FSL) scenario. Extensive experiments conducted on four benchmark data sets demonstrate that DSR-GCN outperforms the other deep learning methods in terms of classification accuracy and generalization performance, with the improvements in terms of OA around 6.20%~23.41% (Indian Pines), 4.45%~ 16.48% (University of Pavia), 4.25%~11.85% (Salinas), and 2.0%~17.23% (University of Houston) under 5 labeled samples per class.

Monitoring mangroves is critical to protect the coastal ecosystems, and deep learning has gained great popularity in mapping mangroves using remote sensing. However, mangroves are usually submerged by cyclical tide since they are grown in land–sea interface places, resulting in some drawbacks of existing mangroves mapping models. On one side, the correlations between the vegetation index (VI) and the water index (WI) time series of mangroves are not fully considered. On another side, existing models rarely explored the local differences between mangroves and other land covers. Considering the above two aspects, we propose a novel two-stream translating long short-term memory network (TSTLN) for mangroves mapping. First, we construct multivariate time series (MTS) by compositing VI and WI based on Sentinel-2 time-series data. Second, we build a two-stream architecture and design a Siamese translating (ST) module in both streams. In the global stream, MTS is embedded into the ST module directly to get global features, whereas, in the local stream, a depthwise convolutional self-attention (DCA) module is conceived to capture local information first, and then, local features are further learned by the ST module. Finally, a fully connected layer and softmax are used to classify the representations extracted from the two streams. Experiments conducted over the Maowei Sea, the Dongzhai Port, and the Quanzhou Bay in 2019 demonstrate that: 1) TSTLN outperforms other methods, with improved OA of 0.49%–3.89%, 1.35%–6.85%, and 0.73%–3.65% in the three areas, respectively; 2) two-stream architecture, ST module, and DCA module all contribute to the good performance of TSTLN; and 3) TSTLN maintains higher accuracy with few parameters and less running time compared to other counterparts.

Recently, few-shot learning (FSL) has exhibited great potentials in hyperspectral image (HSI) classification due to its promising performance under few training samples. Although existing FSL methods have achieved great success, some limitations can still be witnessed. On the one hand, current methods mainly rely on the single metric to identify, which cannot effectively represent the class distribution with few labeled samples. On the other hand, existing methods usually only use the last deep feature of feature extractor, which may lead to the under-utilization of scarce labeled samples. To overcome the above issues, a novel multistage relation network with dual-metric (DM-MRN) is proposed for few-shot HSI classification. Firstly, a sample recombination strategy is designed to increase the variety of classification tasks in training period. Secondly, an embedding module is employed to extract deep features of the input image patches. Thirdly, we propose two relation modules: image-to-class (I2C) block and image-to-image (I2I) block. I2C block is designed to compute I2C-level relation score between second-order features, and I2I block is conceived to generate I2I-level relation score between first-order features. Finally, DM-MRN is constructed by integrating one embedding module, two I2C blocks, and one I2I block. In addition, an adaptive weighting strategy is designed to fuse the obtained relation scores, and classification can be achieved by assigning each query sample to the class with the highest value of the fused relation score. Extensive experiments carried out on five popular HSI data sets demonstrate that the proposed method outperforms other traditional and advanced models under few training samples in terms of classification accuracy and generalization performance, i.e., the performance improvement in terms of OA is around 0.30%-27.98% under 10 labeled samples per class.

Low-rank and sparse representation (LRSR) has gained popularity in hyperspectral image (HSI) classification. However, existing LRSR models usually treat HSI as a two-dimensional matrix, which may destroy the original 3D intrinsic structure of HSI. Moreover, the dictionary consisting of only training samples lacks completeness and may be suboptimal for representation. To overcome the above issues, we propose an incremental dictionary learning-driven tensor low-rank and sparse representation (TLRSR-IDL) model for HSI classification. First, we represent HSI as a third-order tensor to retain its original 3D intrinsic structure by using the TLRSR model, which also combines both sparsity and low rankness to maintain global and local data structures. Second, we design an optimal reconstruction within regularized neighborhood (ORRN) method to exploit spectral-spatial information by avoiding the interference of heterogeneous samples in the neighborhood. Finally, an incremental dictionary learning (IDL) scheme is designed to iteratively introduce augmented samples into the dictionary, and the final classification map is produced by feeding back the last round of the incremental dictionary into the TLRSR-IDL model. The main innovative contribution lies in that the proposed IDL scheme can leverage supervised and unsupervised information, which greatly enhances traditional LRSR and TLRSR models. Experimental results based on three popular hyperspectral datasets demonstrate that the proposed method outperforms other related counterparts in terms of classification accuracy and generalization performance, with OA improvements of 0.97%-16.83%, 1.25%-6.89%, and 0.85%-6.67% for Indian Pines, Pavia University, and Salinas, respectively.

Deep learning (DL) has shown great potentials for hyperspectral image (HSI) classification due to its powerful ability of nonlinear modeling and end-to-end optimization. However, DL models are easily get trapped into overfitting due to limited training labels since the labeling process is time-consuming and laborious in real classification scenario. To overcome this issue, we propose a novel spectral-spatial siamese network (S3Net) for few-shot HSI classification. Firstly, a lightweight spectral-spatial network (SSN) composed of 1-D and 2-D convolution is proposed to extract spectral-spatial features. Secondly, S3Net is constructed by two SSNs in dual branches, which can augment training set by feeding sample pairs into each branch, and thus enhancing the model separability. To provide more features for the model, differentiated patches are fed into each branch, where negative samples are random selected to avoid redundancy. Finally, a weighted contrastive loss is designed to promote the model to fit in the right direction by focusing on sample pairs that are hardly to be identified. Moreover, another adaptive cross entropy loss is conceived to learn the fusion ratio of the two branches. Experiments based on three commonly used HSI data sets demonstrate that S3Net outperforms traditional and state-of-the-art DL-based HSI classification methods under few-shot training scenario. In addition, the weighted contrastive loss and the adaptive cross entropy loss jointly improve the discrimination power of the model.

Soil moisture (SM) is critical in maintaining the balance of Earth’s environment and climate system. Existing machine learning-based SM estimation methods mainly belong to a single model, which may be unstable and probably lacks adaptability when switching to other sites. In addition, thein situobservations are usually inadequate, deteriorating the generalization performance of a single model. To overcome the above issues, we design two novel ensemble learning models based on the Gaussian process regression (GPR) for SM estimation. One is bagging embedded with GPR (BAGGPR), which is a parallel algorithm designed by randomly selecting multiple data subsets to train an ensemble of GPR models. The other is gradient boosting (BOOST) embedded with GPR (GBGPR), which is a sequential algorithm built by iteratively learning the residual between the previous prediction and its true value. In GBGPR, we also propose an improved Huber loss function by applying square loss on different residuals. The proposed methods greatly enhance the stability, adaptability, and generalization performance of a single GPR model. Experiments are conducted in the continental U.S. (CONUS) between April 2015 and March 2016, where BAGGPR and GBGPR are tested to estimate SM based on 11 multisource remote sensing features. The results demonstrate that our proposed methods outperform other state-of-the-art models, including thirteen single models and three ensemble models, with R being 0.8958 and 0.9047, and root mean square error (RMSE) being 0.0513 cm3/cm3and 0.0490 cm3/cm3, respectively, for BAGGPR and GBGPR. Moreover, the proposed methods can well capture spatial dynamics, and they have higher consistent with thein situmeasurements, better generalization performance in terms of training data, and higher adaptability to differentin situnetworks.

Transfer learning (TL) offers an effective way to reduce the demand for labeled samples in remote sensing image classification. However, existing TL methods have some limitations. Simple linear TL methods cannot align the source and target domains well when the data shift is complicated, while nonlinear methods consist of many learnable parameters and often need many labeled samples. To overcome these issues, we design a novel grouped subspace linear semantic alignment (G-SLSA) algorithm, which consists of four main ingredients. First, inspired by the linear supervised transfer learning (LSTL) approach, we propose subspace linear semantic alignment (SLSA) aiming to reduce the demand for labeled samples. Second, considering the heterogeneity of class-level data shift, we extend SLSA to G-SLSA through a grouped alignment strategy, which can reduce the data shift by decomposing a multiclass transfer learning task into a set of binary subtasks. Third, considering the demand of subtasks fusion on posterior probabilities, we propose a robust posterior probability estimation method for the binary generalized learning vector quantization (GLVQ) that is used in G-SLSA. Finally, a pairwise coupling (PWC) method is applied to fuse the results of each subtask. Experimental results conducted on three popular hyperspectral datasets demonstrate that G-SLSA outperforms other traditional and state-of-the-art deep learning (DL) methods, with an OA of 80.78±4.28% for PC-UP transfer learning scenario when 5 samples per class are available in the target domain.

In this letter, we propose a novel weighted sparse graph regularization (SGR) (WSGR) model by incorporating the spatial information both in pre and postprocessing stages. In the preprocessing stage, we generate adaptive patch features to represent the spatial information. Traditional patch feature extraction methods rarely consider the quality of the neighboring pixels within a patch since the central pixel and its neighbors may belong to different classes. To solve this issue, a Gaussian-kernel-based weighting scheme that can adaptively model the neighborhood information is introduced into the original SGR model. Experimental results, conducted with two popular hyperspectral data sets, indicate that the proposed method outperforms other related methods with an overall accuracy higher than 90% when only 20 labeled samples per class are used for training.

Deep learning (DL) has exhibited huge potentials for hyperspectral image (HSI) classification due to its powerful nonlinear modeling and end-to-end optimization characteristics. Although the superior performance of DL-based methods has been witnessed, some limitations can still be found. On the one hand, existing DL frameworks usually resorted to first-order statistical features, whereas they rarely considered second-order or higher order statistical features. On the other hand, the optimization of complex hyperparameters (e.g., the layer number and convolutional kernel size) is time-consuming and a very tough task, making the designed DL framework unexplainable. To overcome these challenges, we propose a novel attention-based second-order pooling network (A-SPN). First, a first-order feature operator is designed to model the spectral–spatial information of HSI. Second, an attention-based second-order pooling (A-SOP) operator is designed to model discriminative and representative features. Finally, a fully connected layer with softmax loss is used for classification. The proposed framework can obtain second-order statistical features in an end-to-end manner. In addition, A-SPN is free of complex hyperparameters tuning, making it more explainable and easily equipped for classification tasks. Experimental results based on three common hyperspectral data sets demonstrate that A-SPN outperforms other traditional and state-of-the-art DL-based HSI classification methods in terms of generalization performance with limited training samples, classification accuracy, convergence rate, and computational complexity.

Frequency spectrum has been proven to have the potential in hyperspectral image classification and ground object recognition. The characteristics of the frequency spectrum, such as dc component, descent rate, and spectrum oscillation, are different from each other; thus, based on the discrepancy in the frequency spectrum, 14 frequency spectrum features, including frequency spectrum integration area, spectral centroid (Ckand Ck-log), spectral rolloff (Ct), spectral flux, spectral gradient of peaks, and valley, number of crosspoint, and first three peaks and valleys position, are proposed. To evaluate the performance of the proposed features, two commonly used hyperspectral images were taken as experimental data sets. Then, we employed three frequently used classification methods to perform the experiment based on spectral-only and frequency-spectral features. The results show that the proposed features can distinctly prompt the classification accuracies by combining the original spectral features.

Recently, deep learning (DL) is of great interest in hyperspectral image (HSI) classification. Although many effective frameworks exist in the literature, the generally limited availability of training samples poses great challenges in applying DL to HSI classification. In this paper, we present a novel DL framework, namely, semisupervised stacked autoencoders (Semi-SAEs) with cotraining, for HSI classification. First, two SAEs are pretrained based on the hyperspectral features and the spatial features, respectively. Second, fine-tuning is alternatively conducted for the two SAEs in a semisupervised cotraining fashion, where the initial training set is enlarged by designing an effective region growing method. Finally, the classification probabilities obtained by the two SAEs are fused using a Markov random field model solved by iterated conditional modes. Experimental results based on three popular hyperspectral data sets demonstrate that the proposed method outperforms other state-of-the-art DL methods.

Active learning (AL) is a promising solution to hyperspectral image classification with very few initial labeled samples. Although previous AL heuristics have exhibited encouraging results, some challenges are still open. On the one hand, traditional AL heuristics measured uncertainty only in feature domain (i.e., spectral or spectral-spatial features) with a pixelwise manner, which ignores the spatial uncertainty. On the other hand, traditional batch-mode AL methods rarely considered spatial homogeneity, since they selected a batch of samples from the candidates, which will induce redundancy unavoidably. To overcome these issues, we first propose an enhanced uncertainty measure considering the neighborhood information. We then propose to use simple linear iterative clustering for generating superpixels, where the selected batch samples are constrained to be from different superpixels, which improves the diversity of the selected samples. The experimental results with two popular hyperspectral data sets indicate that the proposed methods can significantly improve the classification accuracy compared with the traditional methods.

Recently, multifeature learning in collaborative representation classification (CRC) for hyperspectral images has generated promising performance. In this paper, two novel multifeature learning algorithms that update dictionary directly and indirectly are proposed. In order to offer the complementarity of multifeature, four different types of features-global feature (i.e., Gabor feature), local feature (i.e., local binary pattern), shape feature (i.e., extended multiattribute profiles), and spectral feature-are adopted in this paper. Under the hypothesis that most of the features should share the same coding pattern in CRC, this paper proposes to learn proper dictionaries for each feature until obtaining stable codes in a linear classifier. Furthermore, to avoid the explicit mapping of infinite-dimensional dictionaries in a nonlinear kernelized classifier, an indirect approach to construct the transformation matrix from original dictionaries to learn new dictionaries is developed. Three real hyperspectral images acquired from different sensors are adopted for performance evaluation. The experimental results demonstrate that the proposed methods can provide superior performance compared with those of the state-of-the-art classifiers.

Regularization has appeared explicitly in hyperspectral image (HSI) classification community, which serves as a promising paradigm for leveraging labeled and unlabeled information, computer's automation and user's interaction, spectral and spatial information, and so on. Graph-based regularization is capable of modeling the nonlinear structures embedded in high-dimensional space, with the great potential for HSI classification. However, traditional methods exhibit low capacity when facing noisy and large-scale data, thus posing a big challenge for their successful use in this community. In this paper, we present two novel sparse graph regularization methods, SGR and SGR with total variation (TV-SGR). In SGR, the labels of large unknown data are propagated based on the fraction matrix and the prediction function, where the fraction matrix is obtained using an effective sparse representation (SR) algorithm with respect to the dictionary, and the prediction function is estimated by optimizing a typical graph-based regularization problem. In contrast, TV-SGR is an extension of SGR by considering spatial information modeled by total variation in SR. Propagating the prediction function from dictionary to large unknown data using the fraction matrix is the essence of the paradigm. SGR and TV-SGR can be equipped with semisupervised learning, active learning, and spectral-spatial classification with large flexibility. The experimental results with two popular hyperspectral data sets indicate that the proposed methods outperform some state-of-the-art approaches in terms of computational efficacy, classification accuracy, and robustness to noise.

In this research, a novel sparse graph regularization (SGR) method was presented, aiming at robust crop mapping using hyperspectral imagery with very few in situ data. The core of SGR lies in propagating labels from known data to unknown, which is triggered by: 1) the fraction matrix generated for the large unknown data by using an effective sparse representation algorithm with respect to the few training data serving as the dictionary; 2) the prediction function estimated for the few training data by formulating a regularization model based on sparse graph. Then, the labels of large unknown data can be obtained by maximizing the posterior probability distribution based on the two ingredients. The study area is located at Zhangye oasis in the middle reaches of Heihe watershed, Gansu, China, where eight crop types were mapped with Compact Airborne Spectrographic Imager (CASI) and Shortwave Infrared Airborne Spectrogrpahic Imager (SASI) hyperspectral data. Experimental results demonstrate that the proposed method significantly outperforms other classifiers, with an overall accuracy of 87.43% and a kappa value of 0.827 (5 labeled samples per class), which are respectively, 9%-30% and 0.1-0.3 higher than other counterparts.

Sparse graph embedding (SGE) is a promising technique useful for the nonlinear feature extraction (FE) of hyperspectral images (HSIs). However, such images exhibit spatial variability and spectral multimodality, presenting challenges to existing FE methods, including SGE. To address this issue, this paper presents two novel SGE methods for HSI classification. One method, which is termed simultaneous SGE (SSGE), is designed to consider the spatial variability of spectral signatures by using a simultaneous sparse representation (SSR) model integrated with a shape-adaptive neighborhood building approach. In addition, a sparse graph is constructed via matrix computation based on sparse codes. Then, low-dimensional features are produced by employing linear graph embedding (LGE) based on the constructed sparse graph. The other method, which is termed simultaneous sparse multimanifold learning (SSMML), is proposed to handle the multimodality of an HSI. In SSMML, multiple views are generated to represent different modalities. Then, multiview-oriented submanifolds are produced by adopting SSGE, and they are further integrated via coregularization. SSGE is capable of modeling both local and global data structures. Furthermore, SSMML serves as a prototype that can model multimodal data structures. The proposed methods are evaluated by using sparse multinomial logistic regression for HSI classification. Experimental results with two popular hyperspectral data sets validate the good performance of the two methods in producing more representative low-dimensional features and yielding superior classification results compared with other related approaches.

This paper presents a new spectral-spatial classification method for hyperspectral images via morphological component analysis-based image separation rationale in sparse representation. The method consists of three main steps. First, the high-dimensional spectral domain of hyperspectral images is reduced into a low-dimensional feature domain by using minimum noise fraction (MNF). Second, the proposed separation method is acted on each features to generate the morphological components (MCs), i.e., the content and texture components. To this end, the dictionaries for these two components are built by using local curvelet and Gabor wavelet transforms within the randomly chosen image partitions. Then, sparse coding of one of the MCs and update of the associated dictionary are sequentially performed with the other one fixed. To better direct the separation process, an undecimated Haar wavelet with soft threshold is performed for the content component to make it smooth. This process is repeated until some stopping criterion is met. Finally, a support vector machine is adopted to obtain the classification maps based on the MCs. The experimental results with hyperspectral images collected by the National Aeronautics and Space Administration Jet Propulsion Laboratory's Airborne Visible/Infrared Imaging Spectrometer and the Reflective Optics Spectrographic Imaging System indicate that the proposed scheme provides better performance when compared with other widely used methods.