Nowadays, object saliency detection has attracted considerable attention due to the vivid description of images. The traditional saliency detection methods are faced with the challenges of sparse boundary, fractured contour and internal non-uniform density. That will result in losing texture and detailed information of images, which makes a bad contribution to the subsequent object detection. Therefore, we propose a new approach based on fuzzy region competition and a Gaussian mixture segment model via a modified line segment detector (FGM-MLSD) for airport saliency detection in remote sensing images. First, we adopt fuzzy region competition, combining a Gaussian mixture model to segment the input images and obtain the airport candidate regions. After segmentation, a modified line segment detector (LSD) is used for extracting line features, which enhances the connection between broken lines and greatly improves the detected line quality. Then we can acquire the saliency map of the airport region. At last, we fuse the above saliency maps with the binarization map obtained by the Otsu method, aiming to eliminate the false alarm. Finally, abundant experiments are conducted, and the testing results reveal that the neoteric method can clearly and accurately extract the airport region in the remote sensing images and effectively improve the accuracy of saliency detection.

Transformer-based semantic segmentation methods have demonstrated outstanding performance by leveraging global self-attention to effectively capture long-range dependence. However, there still exist two issues in existing works: (1) Most of them utilize the full-rank weight matrix to support the self-attention mechanism and feed-forward network in modelling long-range dependence between patches/pixels, resulting in a high computational cost during both training and inference. (2) Most of them ignore information interactions between high-level semantics and low-level structures during the image resolution recovery, which leads to the performance degradation in segmenting objects with complex boundaries. To tackle these challenges, a lightweight Kolmogorov-Arnold Transformer model (LKAFormer) is proposed for the image semantic segmentation, containing a two-stream lightweight Transformer encoder and a graph feature pyramid aggregation KAN-decoder. The former constructs a hierarchical feature cross-scale fusion pipeline to obtain sufficient semantics containing comprehensive multi-scale information via setting coarse-grained and fine-grained streams with different-size patches of images. In that pipeline, feature lightweight focusing modules model complex and long-range dependence across patches/pixels to refine image semantics with less computational costs by lightweight multi-head self-attention and lightweight feed-forward network designs. The latter leverages the learnable nonlinear transformation mechanism of the Kolmogorov-Arnold Transformer architecture to adaptively capture spatial structure dependence of distinct sub-regions of images. And then, it jointly performs the intra-scale graph fusion and cross-scale graph fusion during the image resolution recovery to enhance information interactions between high-level semantics and low-level structures, which achieves the robust boundary localization and texture refinement of segmentation objects. Finally, plentiful experiments are conducted on three challenging datasets, and the results show LKAFormer sets a new baseline in the image segmentation task in comparison with 11 methods.

Recently, graph neural networks have attracted great attention and achieved outstanding success in hyperspectral image (HSI) classification. However, most existing methods rely on pairwise relationship, neglecting more complex higher-order interactions, which limits the learning of deeply embedded features. Additionally, their dependence on predefined graph structures restricts dynamic node information aggregation. To solve above problems, this paper proposes a CNN enhanced hypergraph attention network (CEHGAT). In order to reveal the high-order interaction in HSI, a hypergraph attention network branch is developed to learn the dynamic connection of hyperedges through the attention mechanism to reveal more representative node embeddings. Then, the CNN enhanced branch uses two multi-scale convolutional blocks to enhance the spatial-spectral features. Finally, the features captured by two branches are fused to realize the complementary advantages of superpixel-level and pixel-level features. Experiments on three benchmark HSI datasets demonstrate that CEHGAT outperforms other state-of-the-art methods with limited labeled samples.

Hyperspectral image (HSI) classification holds great potential for practical applications, but its widespread adoption is limited by the high cost of manual annotation. While unsupervised domain adaptation (UDA) offers a solution by transferring knowledge from labeled source domains (SDs) to unlabeled target domains (TDs), existing methods primarily focus on statistical-level distribution alignment, neglecting instance-level variations in TD data. In addition, for the interfering information such as noise and redundancy that are prevalent in HSI, there are few methods to consider processing the original data at the point level. To overcome these limitations, we propose a mask-guided and confidence-driven UDA (MCUDA) method. It introduces point-level learnable masks to dynamically optimize the input HSI data cube, effectively suppressing interference and enhancing domain-invariant feature extraction. It also proposes a pseudolabel sample set generation strategy based on the idea of confident learning, which takes into account the instance-level differences and domain-related information of TD data. Comprehensive experiments on two cross-scene datasets demonstrate that MCUDA outperforms existing UDA methods, achieving superior classification accuracy.

In recent years, with the continuous advancement of Earth observation technologies, the joint utilization of hyperspectral imagery (HSI) and light detection and ranging (LiDAR) for classifying complex land cover types has garnered widespread attention in the field of multi-source remote sensing. However, the diversity of land cover increases the complexity of the spatial and spectral structures in remote sensing data, creating challenges for extracting discriminative features effectively. Moreover, the heterogeneity of multi-source remote sensing data poses challenges for existing methods to effectively integrate complementary information from various data sources. To address these challenges, a multi-scale split-recombination cooperative fusion network (MSRCFNet) is proposed for joint land cover classification using HSI and LiDAR data. It consists of three main components: multiple parallel multi-scale hierarchical inverted-pyramid (MHIP) modules, a cross-modal cooperative fusion module (CCFM), and a classification module. The MHIP module comprises multiple convolutional split-recombination blocks (CSRBs) at various scales, designed to extract and fuse discriminative multi-scale features. CCFM leverages spatial-scale consistency and the self-attention mechanism to model global relationships between different modalities, and then applies the cross-attention mechanism to effectively integrate complementary information from heterogeneous data. The classification module transforms the fused features into the final classification results. Experimental results on three publicly available HSI-LiDAR datasets demonstrate the superiority of the proposed network over state-of-the-art methods.

The joint classification of hyperspectral images (HSI) and LiDAR data is an important research direction in the field of remote sensing image processing. However, existing methods are mostly limited to inadequate feature extraction or simple serial fusion strategies, failing to fully exploit the deep cross-modal relationships between LiDAR and HSI data. Meanwhile, global modeling based on Transformers, despite capturing long-range dependencies, faces difficulties in handling high-dimensional remote sensing data due to quadratic computational complexity, leading to constrained global-local feature collaboration. To address these issues, this paper proposes a novel hierarchical multi-scale Mamba network (HMMamba) for joint classification of HSI and LiDAR data. First, the method adopts a hierarchical design to obtain rich multi-level feature representations through multi-scale feature extraction. Second, it utilizes adaptive weight allocation to achieve dynamic cross-modal feature integration. Finally, it constructs a Fused Feature enhancement Mamba Block (FFMB), which integrates a dual-branch attention mechanism and selective state space modeling to achieve efficient global context modeling under linear computational complexity Extensive experiments on four public datasets (Houston2013, MUUFL, Trento, and Augsburg) demonstrate that HMMamba significantly outperforms existing state-of-the-art methods in terms of classification performance. Specifically, on Augsburg dataset, the OA improvement is 3.24% compared to the suboptimal method, fully validating the superior performance of the proposed method. The codelink of the proposed method is https://github.com/leiyeqi/HMMamba.

Multi-source remote sensing image classification based on Mamba has received increasing attention. However, existing methods do not consider the non-causal characteristics of visual data, which leads to insufficient extraction of global features by the model. To alleviate this problem, a novel network called TBi-Mamba is proposed for joint classification of hyperspectral and LiDAR data. Firstly, a cross-modal knowledge search module (CMKS) is designed, which effectively captures local features in different modalities through multi-scale feature extraction and interaction between multi-modal data. Secondly, a triple bidirectional sequence scanning mamba module (TBi-M) is proposed, which comprehensively considers multimodal information from the perspective of bidirectional sequence scanning, and introduces Mamba to accurately model global dependencies. Finally, a mixed feature reconstruction module (MFRM) is constructed. This module constructs an auxiliary loss function by reconstructing images of different modalities, providing more comprehensive supervision information and thus improving the performance of the model. The proposed method was evaluated on three publicly available datasets, and experimental results fully demonstrated that the classification performance of the proposed method is superior to that of some state-of-the-art (SOTA) methods.

Vision Transformer (ViT) has been thoroughly explored in hyperspectral image (HSI) classification (HIC). Nevertheless, current ViT-based approaches still acquire discriminative features, resulting in relatively limited generalization capabilities when confronted with the challenges posed by high intraclass variances and interclass similarities commonly observed in HSI data. Moreover, most of these methods fail to adequately emphasize the significance of the central pixel in HIC. To address the aforementioned challenges, we introduce a masked ViT (MViT) for HIC. First and foremost, MViT endeavors for the first time to introduce the masking operations in supervised models to learn more robust patterned features instead of distinguishable features, thereby bestowing it with outstanding generalization performance. Secondly, during the training phase of MViT, when conducting the random masking operations on the embedded features, we deliberately retain the embedding corresponding to the central pixel to guarantee the effectiveness of the model and emphasize the importance of the central pixel in HIC. Finally, MViT will deactivate the masking operations during the testing phase and utilize all the embedded features to accomplish the classification task, thereby enabling the model to fully exploit its recognition capabilities. On top of that, MViT is an extremely lightweight model, and by introducing the masking operations during the training phase, its training speed becomes unprecedentedly rapid. Experiments conducted on four publicly accessible datasets demonstrate that MViT can consistently achieve excellent or even the optimal classification results in comparison with the most advanced methods. The source code was powered by Jupyter and released at https://github.com/swiftest/MViT.

Band selection plays a crucial role as a preprocessing step in hyperspectral image processing tasks. Currently, most band selection methods tend to fuse spatial spectral features, directly from the original spectral domain, without fully exploring the distinct information embedded in different frequency components of the spectrum. In this study, a greedy band selection strategy based on dual-frequency collaborative feature fusion (GSDFC) is proposed for unsupervised band selection of hyperspectral images. Firstly, a dual-frequency collaborative feature fusion module (DFCFFM) is designed to extract and fuse the spatial spectral features of two frequency domain components separately. Secondly, to further enhance the effectiveness of DFCFFM and learn important features in low-frequency components adaptively, a grouping spatial attention module (GSAM) is introduced, which effectively captures global and local spatial dependencies. Finally, a greedy selection strategy is constructed by utilizing the mutual information (MI) of the fused features and the Pearson correlation coefficient (PCC) of the original data to select the optimal subset of bands. The proposed GSDFC band selection method is validated by classification. Extensive experiments have shown that, compared with some advanced methods, the proposed GSDFC method can achieve the best classification performance on three public datasets: Indian Pines, Pavia University, and Houston 2013, which fully demonstrates the effectiveness of the proposed GSDFC method. The related code will be released at https://github.com/Zengzx716/GSDFC.

The rapid development of sensor and multimodal technology has provided more possibilities for multisource remote sensing image classification. However, some existing joint classification methods are limited to single-level feature fusion and fail to fully explore the deep correlation between cross-level features, thus limiting the effective interaction and complementarity of information between different modal data. To alleviate this issue, this article proposes a hierarchical multimodal feature aggregation-based multihead axial attention transformer (HMAT) for joint classification of hyperspectral and light detection and ranging (LiDAR) data. First, a hierarchical multimodal feature aggregation module (HMFA) is proposed to more effectively fuse spatial–spectral features of hyperspectral images (HSIs) and elevation features of LiDAR data and generate more discriminative low-dimensional feature representations. Second, a pyramid-inverted pyramid convolution module (PIP) is designed. Through the complementary feature extraction structure, PIP can more fully capture the multiscale local features in the fused feature map of hyperspectral and LiDAR data. Finally, a multihead axial attention (MHAA) component is constructed to capture information at different scales in the fused feature maps, thereby accurately modeling global dependencies. The proposed HMAT has been extensively tested on three publicly available datasets. The experimental results demonstrate that the classification performance of the proposed method outperforms that of several state-of-the-art methods.

In recent years, deep learning methods, especially convolutional neural networks (CNNs), have been gradually applied to the field of hyperspectral image (HIS) classification. Because the receptive fields of standard convolution are regular, fixed, and limited, CNNs usually only tend to focus on local formations, which cannot fully reflect the complex information in HSIs. To address the above issue, a novel HSI classification method based on pyramid feature extraction with deformable–dilated convolution (PD2C) is proposed. First, a pyramid feature extraction (PFE) model based on a multiscale double-branch module with deformable–dilated convolution (MDBD2) and a deformable downsampling module is proposed to extract local features. Second, transformer is used to extract global features. On this basis, complex information is well utilized for classification. Experiments on three public datasets show that the proposed PD2C method achieves optimal classification results compared with other state-of-the-art classification methods.

The classification of land cover material based on hyperspectral image (HSI) has important research significance. Owing to the high cost of obtaining labeled samples and insufficient training samples, the research of cross-scene HSI classification (CS-HSIC) is receiving more and more attention. At present, the performance of the feature extraction module of CS-HSIC is relatively poor, and the number of training parameters is usually large. To fill the shortcomings of domain generalization (DG) methods and reduce the number of parameters, we propose a lightweight DG network with an attention-assisted cascaded bottleneck (ACB), and it adopts a lightweight bottleneck and multiattention design. This model is adept at extracting domain invariant information contained in the source domain (SD), and it may be flexibly embedded into other models. The experimental results show that our network has good classification accuracy and DG ability when the number of training samples is a little small. As a feature extraction subnetwork, it can improve the performance of the original model or reduce the required resources. The code will be available athttps://github.com/zhulongyu1234/ACB.

Remote sensing image compression networks aim to enhance the similarity between the input image and the reconstructed image. The current network rarely considers the potential relationship between the compression features of different levels and the reconstruction features of the corresponding levels, which limits the improvement of remote sensing image compression performance. In this article, a concept of multilevel domain similarity is first proposed, which fully develops the multilevel domain similarity between the encoding and decoding processes to improve the quality of reconstructed images. On this basis, a multilevel domain similarity enhancement guided network (MDSNet) is proposed for remote sensing image compression. First, an efficient compression baseline network (BaselineA) was proposed, which realizes efficient image compression with low computational complexity. Second, a multilevel domain similarity enhancement module (MDEM) was designed, which improved the quality of the reconstructed image by enhancing the multilevel domain similarity. Third, a global information-enhanced attention module (GIE-AM) was constructed to enhance channel features and global features. Finally, under the guidance of the total loss (LossTotal), which is constructed by the proposed MDEM loss (MDEM-Loss), an effective compression was implemented by the whole network for remote sensing image compression. Experimental results show that compared with some advanced compression models, the proposed MDSNet can significantly improve compression performance with lower computational complexity. In addition, the reconstructed images obtained by the proposed method can provide better classification performance, which further proves that the proposed MDSNet helps to preserve more important features of remote sensing images during the compression process.

In recent years, convolutional neural networks (CNNs) have achieved great success in hyperspectral image (HSI) classification tasks. CNNs focus more on the local features of HSIs. The recently emerging Transformer network has shown great interest in the global features of HSIs. However, existing Transformer networks only consider single-scale feature extraction and do not combine the advantages of multiscale feature extraction and Transformer global feature extraction. To address this issue, this article proposes a dual-branch multiscale Transformer (DBMST) for HSI classification. First, a large-size spectral convolution kernel is utilized for the spectral dimension of the hyperspectral cube for downsampling feature extraction. Next, a channel shrink soft split module (CS3M) is proposed, which not only solves the problem of missing local information in large-scale tokens but also extracts shallow features and performs dimensionality reduction on channels. Then, considering the different dimensions of features extracted at different scales in two branches, a pooled activation fusion module (PAFM) is carefully designed. Finally, the proposed DBMST is evaluated on three commonly used HSI datasets. The experimental results show that DBMST achieves better classification performance compared to other advanced networks, demonstrating the effectiveness of the proposed method in HSI classification.

In recent years, transformer has attracted the attention of many researchers in the field of remote sensing due to its ability to model global information. However, it is difficult to extract local features such as textures and edges of images, thereby limiting the performance of transformer-based hyperspectral image classification (HSIC). Currently, most existing transformer models for HSIC improve their performance by combining the powerful feature extraction ability of convolution, which also introduces a large number of trainable parameters and increases model complexity. To address this issue, this article proposes a dual attention transformer for attention head interaction (DAHIT) for HSIC. First, a spatial local bias module (SLBM) was designed in the spatial branch, which introduces local priors to extract local features effectively without introducing numerous trainable parameters. Then, an attention head interaction module (AHIM) was proposed, which can make the interaction of information obtained by different attention heads. Finally, a diagonal mask multiscale dual attention module (DAM) was constructed in the spectral branch to enhance the attention to the correlation among different spectral bands through diagonal masks and then to extract features at different scales through feature vectors. Through a series of experiments, the proposed DAHIT is evaluated on four commonly used HSI datasets. The experimental results show that compared with other advanced methods, the proposed DAHIT method exhibits excellent classification performance, demonstrating the effectiveness of the proposed method in HSIC.

Convolutional neural networks (CNNs) have been widely used in hyperspectral image (HSI) classification due to their ability to extract image features effectively. However, under the condition of limited samples, the modeling ability of CNNs for the relationships among samples is limited. At present, research on the classification of HSIs with a small number of samples remains an important challenge in the field of HSI processing. Recently, graph convolutional networks (GCNs) have been applied in HSI classification tasks. In this article, a multihop graph rectifies attention and spectral overlap grouping convolutional fusion network (MRSGFN) for HSI classification is proposed. In the graph convolution branch, a multihop graph rectify attention (MHRA) is designed to weight and correct the features extracted by graph convolution. In the convolutional branch, to solve the problem of dimensionality disaster caused by high spectral dimension with a small number of samples, a spectral intra group inter group feature extraction module (SI2FEM) based on spectral overlap grouping is constructed. In order to better fuse the features extracted from CNNs and GCNs, a Gaussian weighted fusion module (GWFM) is elaborately designed in this article. The features extracted by different branches are assigned different weights by GWFM through a 2-D Gaussian map and then fused. Numerous experiments were conducted on three common datasets and showed that the classification performance of the proposed MRSGFN is superior to other advanced methods.

As the most prevalent self-supervised representation learning (SSRL) model, the masked auto-encoder (MAE) has been gradually investigated in semi-supervised hyperspectral image classification (SHIC). However, the majority of the current approaches augment MAE merely from the application perspective or by introducing a weak regularization term, and do not comprehensively consider the challenges posed by the high intraclass variances and interclass similarities that often appear in hyperspectral image (HSI) data. In this article, we present a regularized MAE (RMAE) to address the aforementioned problems. Specifically, within the framework of MAE, we introduce a self-designed induced transformer block, using a small number of visible patches to learn the embeddings of patches with larger receptive fields. The learned embeddings are used to reconstruct the corresponding patches and an induced reconstruction loss is calculated. This strategy creates a much harder task for masked image modeling (MIM), and the induced transformer block is lightweight and imposes negligible computational burden overhead the underlying MAE framework. In addition, by rethinking the masking operations, we develop a masked convolutional neural network (MCNN), uncovering the principle of MAE and affirming the efficacy of RMAE. Finally, we present two metrics: the mean intraclass distance, and the mean interclass distance. Based on the metrics we give two criteria to evaluate the performance of an SSRL model, providing a new coordinate for the research in SSRL-based SHIC. Experiments conducted on four publicly accessible datasets show that RMAE outperforms state-of-the-art methods. The source code was powered by Jupyter and released athttps://github.com/swiftest/RMAE.

For hyperspectral image classification (HSIC), labeling samples is challenging and expensive due to high dimensionality and massive data, which limits the accuracy and stability of classification. To alleviate this problem, a greedy strategy guided graph self-attention network (GS-GraphSAT) is proposed. First, a graph self-attention (GSA) mechanism is designed by combining a multihead self-attention (MHSA) mechanism with the graph attention network (GAT), which can simultaneously consider the direct and indirect relationships between nodes and deeply analyze the intrinsic characteristics of nodes. Second, a multiattention fusion (MAF) module is developed, which utilizes multiscale convolution kernels and attention mechanisms to significantly enhance the network’s ability to extract local features from images at the pixel level, thereby further enriching the hierarchy and diversity of features. Finally, a greedy training strategy (GTS) is proposed. During the training process, GTS accurately determines the optimal time to supplement samples by analyzing the changes in losses, thereby achieving a significant improvement in network classification performance with limited samples. Extensive experiments were conducted on four challenging datasets. The results demonstrate that the proposed method significantly outperforms other state-of-the-art methods in terms of classification accuracy and robustness. The performance improvement of overall accuracy (OA) can reach up to 1.70% in Houston 2013 (HT). The codes of this work will be available athttps://github.com/Isee-max/IEEE_TGRS_GS-GraphSATfor reproduction.

Convolutional neural networks (CNNs) have a strong capacity to extract deep-level features from data. However, the standard convolution (SC) only considers the intensity-information and ignores the spatial gradient-information. Since spatial difference features are more robust to illumination invariance, this letter proposes a Multi-Scale Central Differential Convolutional (MSCDC) network for hyperspectral anomaly detection. Specifically, we use Central Difference Convolution (CDC) to combine intensity- and gradient-information. This solution improves the representation ability of HSIs and enhances the difference between the background and the anomalies. Furthermore, to fully utilize local spatial information and adapt to targets with different sizes, CDC kernels of three different sizes are used to capture high-, mid- and low-level features, respectively. Finally, a SC is used to fuse multi-scale features and obtain more reliable spatial information. Compared with five popular hyperspectral anomaly detection methods on four real-world HSI datasets, the proposed MSCDC exhibits excellent performances.

Deep learning techniques have shown increasing potential in change detection (CD) in hyperspectral images (HSIs). However, most deep learning-based existing methods for HSI CD follow a patch-based local learning framework and concentrate on binary CD. In this letter, we propose an end-to-end semantic feature fused global learning (SFFGL) framework for HSI multiclass change detection (MCD). In SFFGL, a global spatial-wise fully convolutional network (FCN), which introduces a spatial attention mechanism (PAM) between encoder and decoder, is designed to effectively exploit the global spatial information from the whole HSIs and achieve patch-free inference. PAM can adaptively extract global spatial-wise feature representation. In the model training stage, a global hierarchical (GH) sampling strategy is introduced to obtain diverse gradients during backpropagation for more robust performance. The semantic-spatial feature fusion (S2F2) unit is designed to effectively fuse the enhanced spatial context information in the encoder and the semantic information in the decoder. More importantly, a semantic feature enhancement module (SEM) is proposed to weaken the influence of the unchanged regional background on the change regional foreground, thus further improving the accuracy. Experimental results on two benchmark HSI datasets demonstrate the effectiveness of the proposed SFFGL.

Hyperspectral images have been playing an important role in the field of ground object classification because of their rich spatial and spectral information. Aiming at how to extract complex feature information from hyperspectral images, we propose a new feature fusion network model(DAFFN) with dual attention mechanism, which is mainly used to capture more accurate global-local context attention features. The model extracts global context attention features using self-attention mechanism and local context attention features using cross - attention mechanism. Considering the problem that position information is easily lost during the conversion of attention mechanism, we propose a position self-calibration module that can be flexibly embedded into two attention modules. In addition, in order to better integrate global and local features, we also designed a multi-scale global and local feature fusion module (MSGL), which preserves more representative features with less communication costs by aggregating global and local attention features. We have carried out experiments on three commonly used hyperspectral datasets, and the classification results show that our model can achieve high classification accuracy even in the case of a limited number of samples.

Hyperspectral image (HSI) classification is one of the basic tasks of remote sensing image processing, which is to predict the label of each HSI pixel. Convolution neural network (CNN) and graph convolution neural network (GCN) have become the current research focus due to their outstanding performance in the field of HSI classification in recent years. However, GCN is a transductive learning method, which needs all nodes to participate in the training process to get the node embedding. Graph sample and aggregation (GraphSAGE) is an important branch of graph neural network, which can flexibly aggregate new neighbor nodes in non-Euclidean data of any structure, and capture long-range contextual relationships. Superpixel-based GraphSAGE can not only integrate the global spatial relationship of data, but also further reduce its computing cost. CNN can extract pixel-level features in a small area, and our center attention module (CAM) and center weighted convolution (CW-Conv) can also improve the feature extraction ability of CNN by enhancing the dominant position of target pixels. In order to make full use of the advantages of CNN and GraphSAGE, we propose a center weighted convolution and GraphSAGE (CW-SAGE) cooperative network for HSI classification. Specifically, graph simple and aggregate branch is constructed by superpixel-based encoder and decoder modules, then pixel-level features are extracted by a central attention convolutional neural network. Finally, the features of the two branches are spliced together for feature fusion. We conduct experiments on three hyperspectral datasets and compare the results with other current state-of-the-art methods. A series of experiments demonstrate the advantages of our method.

In recent years, pixel-wise hyperspectral image (HSI) classification has received growing attention in the field of remote sensing. Plenty of spectral-spatial convolutional neural network (CNN) methods with diverse attention mechanisms have been proposed for HSI classification due to the attention mechanisms being able to provide more flexibility over standard convolutional blocks. However, it remains a challenge to effectively extract multi-scale features of high-resolution HSI in a real-world complex environment. In this paper, we propose a pyramidal multi-scale spectral-spatial convolutional network with polarized self-attention for pixel-wise HSI classification. It contains three stages: channel-wise feature extraction network, spatial-wise feature extraction network, and classification network, which are used to extract spectral features, extract spatial features, and generate classification results, respectively. Pyramidal convolutional blocks and polarized attention blocks are combined to extract spectral and spatial features of HSI. Furthermore, residual aggregation and one-shot aggregation are employed to better converge the network. Experimental results on several public HSI datasets demonstrate that the proposed network outperforms other related methods.

In the past, deep learning (DL) technologies have been widely used in hyperspectral image classification tasks. Among them, convolutional neural networks (CNNs) use fixed size receptive field (RF) to obtain spectral and spatial features of hyperspectral images (HSIs), showing great feature extraction capabilities, which are one of the most popular DL frameworks. However, the convolution using local extraction and global parameter sharing mechanism pays more attention to spatial content information, which changes the spectral sequence information in the learned features. In addition, CNN is difficult to describe the long-distance correlation between HSI pixels and bands. To solve these problems, a spectral-spatial fusion Transformer network (S2FTNet) is proposed for the classification of hyperspectral images. Specifically, S2FTNet adopts the Transformer framework to build a spatial Transformer module (SpaFormer) and a spectral Transformer module (SpeFormer) to capture image spatial and spectral long-distance dependencies. In addition, an adaptive spectral-spatial fusion mechanism (AS2FM) is proposed to effectively fuse the obtained advanced high-level semantic features. Finally, a large number of experiments were carried out on four datasets, Indian Pines, Pavia, Salinas and WHU-Hi-LongKou, which verified that the proposed S2FTNet can provide better classification performance than other the state-of-the-art networks.

The emergence of convolutional neural networks (CNNs) has greatly promoted the development of hyperspectral image classification (HSIC). However, some serious problems are the lack of label samples in hyperspectral images (HSIs), and the spectral characteristics of different objects in HSIs are sometimes similar among classes. These problems hinder the improvement of HSIC performance. To this end, in this article, a positive feedback spatial-spectral correlation network based on spectral interclass slicing (PFSSC_SICS) is proposed. First, a spectral interclass slicing (SICS) strategy is designed, which can remove similar spectral signature between classes and reduce the impact of similar spectral signature of different classes on HSIC performance. Second, in order to solve the impact of the lack of labeled samples on HSIC, a positive feedback (PF) mechanism and a spatial-spectral correlation (SSC) module are introduced to extract deeper and more features. Finally, the experimental results show that the classification performance of the PFSSC_SICS is far exceed than that of some state-of-the-art methods.

In recent years, convolutional neural networks (CNNs) have been widely used in hyperspectral image classification (HSIC). However, the size of the convolutional kernel in CNNs is fixed, which makes it difficult to capture the dependence of long-range feature information. In addition, the extracted features often contain a large amount of redundant information. In order to alleviate these issues, a feature complementary attention network based on adaptive knowledge filtering (FCAN_AKF) is proposed in this paper. First, in order to alleviate the problem that CNNs are difficult to capture the dependence between close-range and long-range spectral features due to the limited receptive field, a non-local band regrouping (NBR) strategy is designed. NBR enables CNN to capture nonlocal spectral features in a limited receptive field to establish the interdependence between close-range and long-range spectral features. In addition, the non-local features extracted after using NBR and the local features of the original hyperspectral image are integrated to achieve complementation between non-local features and local features. Then, in order to eliminate the interference of redundant information on the network, a dual pyramid spectral spatial attention (DPSSA) module is proposed and used to capture spectral spatial attention. Next, an adaptive knowledge filter (AKF) is designed, which can adaptively further filter out redundant information and enhance feature information that is beneficial for classification. Finally, extensive experiments were conducted on three challenging datasets, demonstrating that the proposed method has stronger competitiveness compared to some state-of-the-art HSIC methods.

For hyperspectral image classification (HIC) tasks, most uncertainty-based active learning (AL) methods only consider the uncertainty, without considering the diversity of actively selected samples and the budget of expert labeling. In this paper, we propose a collaborative active learning (CAL) framework to address this problem. The proposed framework consists of two well-designed base classifiers and an ingenious CAL scheme that takes into account both the uncertainty and diversity of actively selected samples and the cost of expert annotation. Specifically, get benefit from the capsule networks’ ability to accurately identify and locate features, we design two improved capsule networks. For these two networks, we call the first CapsViT (Capsule Vision Transformer), which introduces Vision Transformer (ViT) into the capsule network (CapsNet) to learn the global relationship between the capsule features. We call the second CapsGLOM (Capsule GLOM), the basic structure of this network is derived from the GLOM system proposed by Geoffrey Hinton, we learn from the way CapsNet constructs the primary capsules to improve its implementation details. CapsViT and CapsGLOM are used as the two base classifiers in the proposed CAL framework to select the most informative samples according to the CAL scheme under the premise of fully considering the cost of expert annotation. Experimental results on four benchmark hyperspectral image data sets show that our proposed CAL framework can achieve satisfactory classification results. At the same time, compared with other advanced deep models, our proposed CapsViT and CapsGLOM are also competitive in the supervised HIC tasks. The source code can be available online (https://github.com/swiftest/CAL).

Poisson noise is one of the significant sources of noise present in hyperspectral imagery (HSI). In most of the existing denoising methods, Poisson noise is first transformed into Gaussian noise through the Anscombe transform and then removed. However, the use of Anscombe transform can give rise to transform errors that affect the final denoising results. In addition, blurs often contaminate the HSI during the imaging procedure, which makes it more difficult to remove the Poisson noise. In view of the above problems, under the maximum a posteriori (MAP) model, we propose a Poissonian blurred HSI denoising based on variable splitting and penalty technique (named as VSPT) to directly remove the Poissonian blurred HSI noise without using the Anscombe transform. By finding the minimum value of the negative logarithmic Poisson log-likelihood combined with the total variation (TV), the proposed method transforms the problem into two subproblems, which are easier to solve: 1) a TV regularized deconvolution problem and 2) an ordinary convex optimization problem. The experimental results show that the proposed VSPT method can effectively remove Poisson noise in HSI contaminated by blurs during the imaging procedure.

Autoencoder (AE) has been widely used in the field of hyperspectral anomaly detection. It is assumed that the background can be reconstructed well, but the anomalies cannot. Hence, the pixels with larger reconstruction error are considered as anomalies. However, owing to the strong nonlinear representation ability of AE, it is difficult to distinguish between background and anomalies. To address this problem, we propose a Residual Self-Attention-based AutoEncoder (RSAAE) for hyperspectral anomaly detection. RSAAE consists of dense residual self-attention modules, an encoder, and a decoder. First, a novel residual self-attention module is designed, which can effectively extract the main features and weaken the ability of subsequent network to reconstruct anomalies, as well as preserve the original features to avoid the deterioration of network performance after the use of dense self-attention modules. Furthermore, inspired by manifold learning, we assume that the background is low-rank in the original space, and has the same property in the latent space after dimensionality reduction. We proposed a low-rank loss function to constrain the latent space, thereby suppressing anomaly reconstruction. Experiments on four real hyperspectral image (HSI) datasets showed that the proposed RSAAE method can produce more accurate detection results than eight popular methods.

The very high spectral resolution in hyperspectral images (HSIs) presents an opportunity to detect subtle land-cover changes. However, availability of HSIs acquired from different platforms requires the development of change detection (CD) methods for HSIs capable to process images with different spatial resolutions. In this paper, we propose an end-to-end sub-pixel convolution-based residual network (SPCNet) to detect changes between high resolution (HR) and low resolution (LR) HSIs. First, an efficient sub-pixel convolution layer is introduced to upscale the LR feature maps into the HR one. Then, the super resolution (SR) block is designed to generate more discriminative representations in sub-pixel-based LR images. Moreover, the sub-pixel-based feature of LR image and pixel-based feature of HR image are concatenated as an input to the designed ResNet for HSI CD. Experimental results on two HSI datasets demonstrate the effectiveness of the proposed SPCNet.

In recent years, deep learning methods have made great progress in hyperspectral image classification. However, the current models obtain deep-seated global feature by deepening the number of network layers, ignore the local neighborhood environment. To solve this problem, we propose a double-branch network model, which extracts spatial and spectral local context feature respectively, fuses different levels of feature adaptively and has strong ability to aggregate context information. Specifically, this model uses self-calibrated convolution to extract spatial information and multi-scale dense convolution to extract spectral information. We propose that local context feature extraction (LCFM) module makes full use of the background information and mutual information between positions of each pixel to obtain local context feature information. In addition, focal-loss is used to solve the problem of different classification difficulty of each sample. Simulation results showed that the proposed module not only achieves high classification accuracy, but also greatly reduces the amount of calculation and parameters.

Hyperspectral image classification methods based on deep learning and attention mechanism have been extensively studied in recent years because of their superior performance. However, the currently applied spatial attention mechanism and channel attention mechanism are separated from each other. For this reason, we propose a new dual-triple attention network (DTAN), which realizes the high-precision classification of hyperspectral images based on capturing cross-dimensional interactive information. Specifically, DTAN is divided into two branches to extract the spectral information and spatial information of the hyperspectral image, which are called the spectral branch and spatial branch. While applying the channel attention mechanism to the spectral unit, the cross-dimensional interaction between the channel dimension and the spatial dimension is constructed. When the spatial attention mechanism is applied to the spatial branch, the correlation with the channel domain is also considered. Moreover, we introduce an efficient channel attention (ECA) module into the DenseNet, which allows the DenseNet to achieve partial cross-channel interaction. A series of experiments proved that DTAN has a significant advantage compared to other models when the training samples are minimal.

Hyperspectral anomaly detection is one of the most important applications in the field of hyperspectral image (HSI) processing. However, hyperspectral anomaly detectors still face several challenges, including the limited use of spatial information and the unavoidable anomaly pollution problem. To cope with the above problems, we propose a hyperspectral anomaly detector, termed COPCRD, which enhances the prevailing collaborative-representation-based detector (CRD) using COPula-based Outlier Detection (COPOD) for background purification and guided filter for spatial difference enhancement. COPCRD mainly solves the anomaly pollution problem and further considers the spatial information of hyperspectral data to enhance discrimination of backgrounds and anomalies. Experimental results on four hyperspectral datasets reveal that the proposed method is more accurate than four state-of-the-art anomaly detectors.

Compared with traditional hyperspectral image classification methods, the classification model based on the deep convolutional neural network (DCNN) can achieve higher precision classification. However, the increase in classification accuracy has led to explosive growth in model complexity. In this letter, we proposed a more lightweight and efficient residual structure to alleviate this problem to replace the standard residual structure. This structure uses the “divide and conquer” idea to reduce the number of model parameters and calculations. In addition, the structure introduces a self-attention mechanism so that the input feature map and output feature map can be adaptively fused, and the feature extraction ability of the residual structure is further enhanced. The experimental results reveal that the residual structure we proposed can significantly reduce the complexity of the model and maintain a high classification accuracy, even surpassing the current mainstream classification model.

Convolutional neural networks (CNNs) have exhibited extraordinary achievements in hyperspectral image (HSI) classification due to their detailed representation of features. However, the improvement of classification accuracy often leads to an evident increase in the complexity of the model, which makes it challenging for the model with the state-of-the-art performance to be applied in the actual scene. Considering MobileNetV3 as a lightweight feature extractor, this article proposes a model suitable for HSI classification based on MobileNetV3. To decrease the problem of massive redundant calculations in the existing spatial attention module, this article proposes a more concise and efficient spatial attention module based on the visual feature maps experiment. Besides, multiclass focal-loss is applied to solve the problem that the difficulty of classification varies for each sample. The experimental results demonstrate that in the case of using very few training sets, the proposed model can tremendously reduce the number of calculations and parameters while maintaining high accuracy.

In recent years, convolutional neural networks (CNNs) have achieved excellent performance in hyperspectral image classification and have been widely used. However, the convolution kernel used in traditional CNN has the limitation of single scale which is not conducive to the improvement of hyperspectral classification performance. In addition, training a classification network of high-dimensional data based on limited labeled samples is still one of the challenges of hyperspectral image classification. To solve the above problems, a hyperspectral image classification method based on expansion convolution network (ECNet) is proposed. The expansion convolution injects holes into the standard convolution kernel to expand the receptive field (RF), so as to extract more context features. Because the shallow features of hyperspectral images contain more location and detail information, while the deep features contain stronger semantic information, in order to further enhance the correlation between deep and shallow information, inspired by ResNet, a similar feedback block (SFB) is introduced on the basis of ECNet, and the deep features and shallow features are fused through this feedback mechanism. Thus, an improved version of ECNet method is obtained, which is called FECNet. This study was tested on four commonly used hyperspectral data sets (i.e. Indian Pine (IP), Pavia University (UP), Kennedy Space Center (KSC), Salinas Valley (SV)) and on a higher resolution and complexly distributed land cover data set (University of Houston (HT)). The experimental results show that the proposed method has better classification performance than some state-of-the art methods, which shows that FECNet has a certain potential in hyperspectral image classification.

Hyperspectral images (HSIs) always contain abundant spectral and spatial information. Most of the existing deep learning-based hyperspectral anomaly detection methods consider spectral differences between the background and anomalies, and the local spatial information is usually ignored. To make complete use of the spatial-spectral information, this paper proposed a Local Spatial-Spectral information-integrated Semi-supervised Two-stream Network (LS3T-Net) for hyperspectral anomaly detection. The two-stream network comprises an adaptive convolution and fully connected network and a variational autoencoder (VAE). The adaptive convolution and fully connected network is used to extract the local spatial features of patches, while the VAE is trained to learn spectral information close to the background pixels. Furthermore, the detection maps from the two-stream network are incorporated through a process combining the benefits of spatial learning and spectral learning. This enhances the ability to separate the background and anomalies and suppress the false alarm. The experimental results for six real HSI datasets reveal that LS3T-Net can produce more accurate detection results than seven popular benchmark methods.

Deep learning methods, especially convolutional neural network (CNN)-based methods, have shown promising performance for hyperspectral image (HSI) change detection (CD). It is acknowledged widely that different spectral channels and spatial locations in input image patches may contribute differently to CD. However, they are treated equally in existing CNN-based approaches. To increase the accuracy of HSI CD, we propose an end-to-end Siamese CNN (SiamNet) with a spectral–spatial-wise attention (SSA-SiamNet) mechanism. The proposed SSA-SiamNet method can emphasize informative channels and locations and suppress less informative ones to refine the spectral–spatial features adaptively. Moreover, in the network training phase, the weighted contrastive loss function is used for more reliable separation of changed and unchanged pixels and to accelerate the convergence of the network. SSA-SiamNet was validated using four groups of bitemporal HSIs. The accuracy of CD using the SSA-SiamNet was found to be consistently greater than for ten benchmark methods.

Deep learning-based methods (e.g., convolutional neural network (CNN)-based methods), have shown increasing potential in hyperspectral image (HSI) change detection (CD). However, the recent advances in CNN-based methods in HSI CD tasks are mostly devoted to designing more complex architectures or adding additional hand-designed blocks. This increases the number of parameters making model training difficult. In this paper, we propose an end-to-end residual self-calibrated network (RSCNet) to increase the accuracy of HSI CD. To fully exploit the spatial information, the proposed RSCNet method adaptively builds inter-spatial and inter-spectral dependencies around each spatial location with fewer extra parameters and reduced complexity. Moreover, the introduced self-calibrated convolution (SCConv) helps to generate more discriminative representations by heterogeneously exploiting convolutional filters nested in the convolutional layer. The designed RSC module can explicitly incorporate richer information by introducing response calibration operation. The experiments on four bi-temporal HSI datasets demonstrated that the proposed RSCNet method is more accurate than ten widely used benchmark methods.

The very high spectral resolution in hyperspectral images (HSIs) offers an opportunity to detect subtle land-cover changes. However, the availability of HSIs acquired from different platforms requires the development of change detection (CD) methods capable of processing HSIs with different spatial resolutions. In this paper, we propose a general end-to-end subpixel convolution-based residual network (SPCNet) to accomplish the CD task between high spatial resolution (HR) and low spatial resolution (LR) HSIs. To effectively tackle the resolution matching issue, a super resolution (SR) block with an efficient subpixel convolution layer is introduced to upscale the LR feature maps into HR maps. The subpixel convolution layer can fully explore the subpixel context information by learning an array of upscaling filters. Moreover, the designed SPC module is embedded into the LR branch to generate more discriminative representations. More importantly, the SPC module as a plug-and-play unit has the potential to be embedded into other baseline networks to enhance the feature learning capability. Experimental results on four HSI datasets demonstrate the effectiveness of the proposed SPCNet.

Researches on the classification of hyperspectral images (HSIs) based on deep learning are in full swing, especially the spectral-spatial dependent global learning (SSDGL) framework, which is both efficient and robust. However, the global convolutional long short-term memory (GCL) module under this framework fails to take full consideration of the spectral characteristics contained in HSIs, and the hierarchically balanced (H-B) sampling strategy introduced in this framework prevents the training process from converging smoothly. In this article, we develop a novel regularized spectral-spatial global learning (RSSGL) framework. Compared with SSDGL, the proposed framework mainly makes three improvements. Above all, aiming at the problem that the GCL module used in SSDGL cannot fully tap the local spectral dependence, we apply 3D convolution to the gated units of long short-term memory (LSTM) as an alternative to the GCL module for adjacent and non-adjacent spectral dependencies learning. Furthermore, to extract the most discriminative features, an improved statistical loss regularization term is developed, in which we introduce a simple but effective diversity-promoting condition to make it more reasonable and suitable for deep metric learning in HSI classification. Finally, to effectively address the performance oscillation caused by the H-B sampling strategy, the proposed framework adopts an early stopping strategy to save and restore the optimal model parameters, making it more flexible and stable. Experiments conducted on three representative data sets show that the proposed RSSGL has superior classification performance compared with the existing relatively excellent research methods. The source code is released at https://github.com/swiftest/RSSGL.

Due to the limitations of hardware technology and budget constraints, there always exists a tradeoff between spatial and spectral resolutions in a hyperspectral image (HSI). Because of the limited spatial resolution, mixed pixels are a common issue in HSIs, and consequently, some targets appear as subpixels. The effectiveness of hyperspectral target detection is affected greatly by the subpixel targets, especially when the size of the targets is small. In this article, we proposed a double dictionary-based nonlinear representation model for hyperspectral subpixel target detection (DDNRTD). DDNRTD represents HSIs with a nonlinear model based on background and target dictionaries, which fully considers the spatial property of background and targets and can separate background and targets reliably, especially for small-sized subpixel targets. In addition, we designed an over-completed background dictionary construction strategy to represent the background part more effectively, which integrates spectral angle distance (SAD) with sparse representation. Experiments on two simulated and five real HSI datasets showed that the proposed DDNRTD method produced more accurate detection results than six state-of-the-art methods.

Due to the influences of imaging conditions, spectral imagery can be coarse and contain a large number of mixed pixels. These mixed pixels can lead to inaccuracies in the land-cover class (LC) mapping. Super-resolution mapping (SRM) can be used to analyze such mixed pixels and obtain the LC mapping information at the subpixel level. However, traditional SRM methods mostly rely on spatial correlation based on linear distance, which ignores the influences of nonlinear imaging conditions. In addition, spectral unmixing errors affect the accuracy of utilized spectral properties. In order to overcome the influence of linear and nonlinear imaging conditions and utilize more accurate spectral properties, the SRM based on spatial-spectral correlation (SSC) is proposed in this work. Spatial correlation is obtained using the mixed spatial attraction model (MSAM) based on the linear Euclidean distance. Besides, a spectral correlation that utilizes spectral properties based on the nonlinear Kullback-Leibler distance (KLD) is proposed. Spatial and spectral correlations are combined to reduce the influences of linear and nonlinear imaging conditions, which results in an improved mapping result. The utilized spectral properties are extracted directly by spectral imagery, thus avoiding the spectral unmixing errors. Experimental results on the three spectral images show that the proposed SSC yields better mapping results than state-of-the-art methods.

Subpixel mapping based on the Hopfield neural network (HNN) is a technique to handle mixed pixels for obtaining the spatial distribution information of land cover. However, the original low-resolution remote sensing image may contain some uncertainties, such as the diversity of the land cover classes and the limitation of the resolution of the satellite sensor, the existing HNN is unable to fully utilize the prior information of the original image. In order to resolve this problem, an improved HNN (I-HNN) is proposed in this letter. In the proposed I-HNN, additional prior information of the original image is supplied by adding a new processing path to the existing HNN. To validate the effectiveness of the proposed method, two experiments are conducted on real hyperspectral images. The obtained results demonstrate that the proposed I-HNN outperforms the existing HNN. Moreover, the I-HNN does not require any auxiliary data.

An improved version of supervised locally linear embedding is proposed. In this algorithm, the weight factors of the supervised method are adaptively achieved. This method can simplify the supervised feature extraction algorithm by reducing parameters. To improve classification accuracy, a clustering‐based fuzzy support vector machine (FSVM) is proposed. Different from traditional FSVMs, the proposed method constructs the fuzzy weights by inner‐class clusters. In the proposed method, loose density is defined to express the compactness of the inner‐class clusters. The proposed algorithm can restrain the noise and outliers by exploiting the method of endowing with smaller weight for big loose density and bigger weight for the small loose density of samples in the clusters. To inspect the performance of the proposed methods, we conduct experiments on two hyper‐spectral images. Results show that the two methods are competitive among the competitors.

In this paper, a framework based on fine resolution thematic map, namely, interpolation then classification (ITC) is proposed. Firstly interpolation algorithm is applied in the original coarse hyperspectral imagery to derive a high-resolution imagery with generous prior information. Then fine resolution thematic map is derived from the high-resolution imagery by the available classification methods. Experiments on two real hyperspectral imagery showed that the proposed method produced higher accuracy result than interpolation-based soft-then-hard super-resolution mapping (I-STHSRM).

In this paper, a sub-pixel mapping (SPM) method based on super-resolution then spectral unmixing (SRTSUSPM) is proposed. In the proposed framework, firstly projection onto convex set (POCS) model with the endmembers of interest is applied to original imagery to obtain a high-resolution imagery; then the fraction images are derived from the high-resolution imagery by linear spectral mixture analysis (LSMA); finally hard attribute values on a per sub-pixel basis is implemented to achieve SPM. Experiments show that the higher mapping accuracy can be derived from the proposed SPM method.

Several spectral unmixing techniques using multiple endmembers for each class have been developed. Although they can address within-class spectral variability, their unmixing results may have low unmixing resolution when the within-class variation is large due to the associated high uncertainty. Therefore, it is critical to represent data in an effective feature space so that the endmember classes are compact with small variation. In this letter, a minimum-class-variance support vector machine (MCVSVM) is further developed to extend its functions for both classification and spectral unmixing. Moreover, analytical expressions for spectral unmixing resolution (SUR) are provided to measure the spectral unmixing uncertainty in the new feature space. The extended MCVSVM (e_MCVSVM) can improve SUR and reduce the spectral unmixing uncertainty as it can effectively maximize the between-class scatter while minimizing the within-class scatter. Experimental results show that the e_MCVSVM algorithm performs better in terms of the unmixing accuracy and the computation speed compared with the other algorithms (e.g., fully constrained least squares and endmember bundles) in both linearly separable and nonseparable cases. This newly proposed approach advances the linear spectral mixture analysis with greater speed and higher accuracy based on the SVM after the SUR is effectively characterized.

In this letter, a novel subpixel sharpening for soft-then-hard subpixel mapping (SPM) is proposed. First, the fractional images for each class are, respectively, derived by spectral unmixing followed by spatial interpolation and by spectral interpolation followed by spectral unmixing. Bilinear and bicubic interpolation is used as the spatial and spectral interpolation methods. The fractional images for each class are then integrated together using the appropriate weighting parameter. Finally, the integrated finer fractional images are used to allocate hard class labels to subpixels. The proposed method is fast and does not need any prior spatial structure information. Experiments on two actual hyperspectral images show that the proposed method produces higher accuracy results than the existing algorithms. Moreover, both the spatial and spectral information is fully utilized to improve the accuracy of the SPM results.

An interactive color visualization method is proposed for hyperspectral imagery (HSI). The method visualizes complex information through different fusion results of multiple images in a color space which is under the interactive control of the observers. In order to solve the main problem of traditional visualization methods, i.e., they can at most display information from three bands in one image, this paper proposes an easy, vivid and effective method for color visualization. In the proposed approach, observers interactively control a cursor position to change the output fusion images and their fusion coefficients. In the approach, the dynamic display will include more than three bands of HSIs. The proposed method is also applicable for visualization of other multi-images, e.g., multispectral images, output images of direction filters, multi-focus images, and multi-temporal images, etc.

The kernel function plays an important role in machine learning methods such as the support vector machine. In this letter, a new kernel framework is developed for hyperspectral image classification. In contrast to existing composite kernels constructed via a linearly weighted combination, the multiple-mapping kernel proposed in this letter is obtained through repeated nonlinear mappings. Experiments indicate that the proposed multiple-mapping kernel framework (MMKF) is effective for hyperspectral image classification. Compared to the single kernel methods, the MMKF tends to be more advantageous in terms of classification accuracy, particularly for the situation with a small-size training set.

Due to the limited spatial resolution of multispectral/hyperspectral data, mixed pixels widely exist and various spectral unmixing techniques have been developed for information extraction at the subpixel level in recent years. One of the challenging problems in spectral mixture analysis is how to model the data of a primary class. Given that the within-class spectral variability (WSV) is inevitable, it is more realistic to associate a group of representative spectra with a pure class. The unmixing method using the extended support vector machines (eSVMs) has handled this problem effectively. However, it has simplified WSV in the mixed cases. In this paper, a further development of eSVMs is presented to address two problems in multiple-endmember spectral mixture analysis: 1) one mixed pixel may be unmixed into different fractions (model overlap); and 2) one fraction may correspond to a group of mixed pixels (fraction overlap). Then, spectral unmixing resolution (SUR) is introduced to characterize how finely the mixture in a mixed pixel can be quantified. The quantitative relationship between SUR and WSV of endmembers is derived via a geometry analysis in support vector machine feature space. Thus, the possible SUR can be estimated when multiple endmembers for each class are given. Moreover, if the requirement of SUR is fixed, the acceptance level of WSV is then limited, which can be used as a guide to remove outliers and purify endmembers for each primary class. Experiments are presented to illustrate model and fraction overlap problems and the application of SUR in uncertainty analysis of spectral unmixing.

There is a type of algorithm for subpixel mapping (SPM), namely, the soft-then-hard SPM (STHSPM) algorithm that first estimates soft attribute values for land cover classes at the subpixel scale level and then allocates classes (i.e., hard attribute values) for subpixels according to the soft attribute values. This paper presents a novel class allocation approach for STHSPM algorithms, which allocates classes in units of class (UOC). First, a visiting order for all classes is predetermined, and the number of subpixels belonging to each class is calculated using coarse fraction data. Then, according to the visiting order, the subpixels belonging to the being visited class are determined by comparing the soft attribute values of this class, and the remaining subpixels are used for the allocation of the next class. The process is terminated when each subpixel is allocated to a class. UOC was tested on three remote sensing images with five STHSPM algorithms: back-propagation neural network, Hopfield neural network, subpixel/pixel spatial attraction model, kriging, and indicator cokriging. UOC was also compared with three existing allocation methods, i.e., linear optimization technique (LOT), sequential assignment in units of subpixel (UOS), and a method that assigns subpixels with highest soft attribute values first (HAVF). Results show that for all STHSPM algorithms, UOC is able to produce higher SPM accuracy than UOS and HAVF; compared with LOT, UOC is able to achieve at least comparable accuracy but needs much less computing time. Hence, UOC provides an effective and real-time class allocation method for STHSPM algorithms.

Iterative Error Analysis (IEA) widely known as a good endmember extraction (EE) algorithm. It is robust, automatic and free of data transformation. However, IEA is faced with risks in some cases due to the sole use of unmxing distance, and its speed is lowed down by the iteration-based linear spectral mixture analysis (LSMA). To make IEA algorithm faster and more robust, its modified version is proposed based on two substitutions. One is substituting integrated distance for unmxing distance, which makes the algorithm more robust. The other is substituting SVM-based multiple endmember spectral mixture analysis (MESMA) for iteration-based LSMA, which speeds up the algorithm greatly. Experiments show that the modified IEA algorithm outperforms original one in terms of both robustness and running speed.

Spectral unmixing has been an important technique for hyperspectral imagery processing. In traditional spectral unmixing methods that are based on the linear spectral mixture model (LSMM), unmixing accuracy is limited by the inherent deficiency of the model. It was shown that the support vector machine (SVM) can be extended for spectral unmixing, based on the advantage that the SVM model can accommodate the variations within a relative pure class by using multiple pure samples instead of a single endmember for one class. In the SVM model, class label errors are considered in constraints. However, the errors concerned in spectral unmixing are the unmixing residue instead of the class label ones. This letter presents a method of imposing unmixing residue constraints on the least squares SVM unmixing model. The related problems, including deducing the closed-form solution and substituting the single endmember for multiple ones, were studied together. Experiments showed that the new SVM model was superior to the original SVM as well as the traditional LSMM in terms of unmixing residue, fractional abundance, and confused matrix criterions.

Subpixel mapping (SPM) is a promising technique to increase the spatial resolution of land cover maps. Markov random field (MRF)-based SPM has the advantages of considering spatial and spectral constraints simultaneously. In the conventional MRF, only the spectral information of one observed coarse spatial resolution image is utilized, which limits the SPM accuracy. In this letter, supplementary information from subpixel shifted remote sensing images (SSRSI) is used with MRF to produce more accurate SPM results. That is, spectral information from SSRSI is incorporated into the likelihood energy function of MRF to provide multiple spectral constraints. Simulated and real images were tested with the subpixel/pixel spatial attraction model, Hopfield neural networks (HNNs), HNN with SSRSI, image interpolation then hard classification, conventional MRF, and proposed MRF with SSRSI based SPM methods. Results showed that the proposed method can generate the most accurate SPM results among these methods.

Spectral unmixing is one of the important techniques for hyperspectral data processing. The analysis of spectral mixing is often based on a linear, fully constrained (FC) (i.e., nonnegative and sum-to-one mixture proportions), and least squares criterion. However, the traditional iterative processing of FC least squares (FCLS) linear spectral mixture analysis (LSMA) (FCLS-LSMA) is of heavy computational burden. Recently developed geometric LSMA methods decreased the complexity to some degree, but how to further reduce the computational burden and completely meet the FCLS criterion of minimizing the unmixing residual needs to be explored. In this paper, a simple distance measure is proposed, and then, a new geometric FCLS-LSMA method is constructed based on the distance measure. The method is in line with the FCLS criterion, free of iteration and dimension reduction, and with very low complexity. Experimental results show that the proposed method can obtain the same optimal FCLS solution as the traditional iteration-based FCLS-LSMA, and it is much faster than the existing spectral unmixing methods, particularly the traditional iteration-based method.

Extended support vector machines (ESVM) was introduced recently for spectral unmixing. It models a class using a group of representative spectra to accommodate within class spectral variation. This paper presents a further geometry analysis of this method, and an improved ESVM is developed, which takes into account both within-class spectral variability and within each mixed case. The experiments illustrate that the new proposed algorithm can obtain more realistic unmixing results.

In this paper, a new sub-pixel mapping algorithm is proposed based on sub-pixel/sub-pixel spatial attraction model (SSSAM). Different from the original sub-pixel/pixel spatial attraction model (SPSAM), the SSSAM considers the spatial distribution of each sub-pixel within neighbor pixels, when calculating the spatial attractions for sub-pixels within the centre pixel. Then the attractions are used to determine the class values of these sub-pixels. Two experiments on three artificial images and one real remote sensing image are processed. Both of the results show that compared with traditional SPSAM, the proposed method can produce sub-pixel mapping results with higher accuracy.

In this letter, the supervised classification algorithm support vector machines is extended to map both pure pixels and mixed pixels using hyperspectral data. The margins between the hyperplanes formed by the pixels on the class boundaries are recognized as mixed region, and the space beyond this region is related to pure pixels. In this way, each endmember is modeled by a set of training samples instead of a single (representative) spectrum to accommodate the variations within the relative pure pixels due to system noise. Unmixing outputs generate an integrated soft- and hard-classification map. The better performance comparing with conventional spectral unmixing method was demonstrated using hyperspectral data sets.

In this letter, a geometry-based feature-selection method is proposed for efficient analysis of hyperspectral imagery. It searches the vertices that form the largest simplex iteratively in pixel space. These vertices are representative subsets of spectral bands. A distance measure is introduced in the simplex volume comparison for fast implementation of the proposed method. Fast principal component analysis and spectral band indexing are suggested for data preprocessing. This method can be implemented in supervised or unsupervised manner. It is automatic, fast, and distribution-free. Experimental results show the superiority of the proposed method in terms of quality and speed