Search

Nonlinear characteristics, spectral variability, and high dimensionality pose significant challenges to the classification of hyperspectral images. Therefore, classifiers need more training samples to reach satisfactory results, particularly when only a small number of labeled samples are available. Existing active learning (AL) methods often neglect dimension reduction (DR), relying on the full spectrum, which limits their effectiveness. To address these issues, we use an end-to-end active learning framework that combines advanced preprocessing techniques, a probabilistic spatial-aware collaborative representation (PSACR) classifier, and a novel clustering-based active learning sampling strategy (MMU+CLSW) in a unified framework. The preprocessing phase, consisting of DR and domain transform filter (DtF), plays a crucial role in addressing spectral variability and high dimensionality. DR reduces the high dimensionality, and DtF helps in extracting important features, which together contribute to reducing intra-class variance and increasing inter-class variance, improving class separability. The probabilistic spatial aware collaborative representation (PSACR) classifier enables high accuracies to be achieved using these features under a limited training set. Moreover, sample selection is a crucial aspect of active learning. A novel clustering-based diversity and uncertainty sampling method (MMU+CLSW) is proposed for this purpose, leveraging collaborative weights obtained from PSACR. MMU+CLSW selects diverse and uncertain examples by clustering collaborative weights, benefiting from the intrinsic information of the PSACR classifier. We evaluate our method on three hyperspectral datasets: Indian Pines, Pavia University, and Salinas. Experimental results demonstrate the superiority of our approach over existing supervised and active learning techniques by achieving high classification accuracy even with an extremely limited number of labeled examples. Codes are available at https://github.com/alicankrc2/An-End-To-End-Active-Learning-Framework [ABSTRACT FROM AUTHOR]

Hyperspectral images are high-dimensional data that capture detailed spectral information across a wide range of wavelengths, enabling the precise identification and analysis of different materials or objects. However, the high dimensionality of the data also introduces information redundancy and increases the computational overhead, making it necessary to perform band selection to retain the most discriminative and informative bands for the target task. Traditional band selection methods, such as ranking-based, searching-based, and clustering-based approaches, often rely on handcrafted features and heuristic rules, which fail to fully exploit the latent information and complex spatial–spectral relationships in hyperspectral images. To address this issue, this paper proposes a two-stage unsupervised band selection method based on deep reinforcement learning. First, we performed noise estimation preprocessing to filter out bands with high noise levels to reduce the interference in the agent's learning process. Then, the band selection problem was formulated as a Markov Decision Process (MDP), where the agent learned an optimal band selection strategy through interactions with the environment. In the design of the reward function, the Optimal Index Factor (OIF) was introduced as the evaluation metric to encourage the agent to select bands with high information content and low redundancy, and thereby improve the efficiency and quality of the selection process. Experimental results on three hyperspectral datasets demonstrated that the proposed method could effectively improve the performance of the hyperspectral image band selection. [ABSTRACT FROM AUTHOR]

In contrast to conventional remote sensing images, hyperspectral remote sensing images are characterized by a greater number of spectral bands and exceptionally high resolution. The richness of both spectral and spatial information facilitates the precise classification of various objects within the images, establishing hyperspectral imaging as indispensable for remote sensing applications. However, the labor-intensive and time-consuming process of labeling hyperspectral images results in limited labeled samples, while challenges like spectral similarity between different objects and spectral variation within the same object further complicate the development of classification algorithms. Therefore, efficiently exploiting the spatial and spectral information in hyperspectral images is crucial for accomplishing the classification task. To address these challenges, this paper presents a multi-scale feature fusion convolutional neural network (MSFF). The network introduces a dual branch spectral and spatial feature extraction module utilizing 3D depthwise separable convolution for joint spectral and spatial feature extraction, further refined by an attention-based-on-central-pixels (ACP) mechanism. Additionally, a spectral–spatial joint attention module (SSJA) is designed to interactively explore latent dependency between spectral and spatial information through the use of multilayer perceptron and global pooling operations. Finally, a feature fusion module (FF) and an adaptive multi-scale feature extraction module (AMSFE) are incorporated to enable adaptive feature fusion and comprehensive mining of feature information. Experimental results demonstrate that the proposed method performs exceptionally well on the IP, PU, and YRE datasets, delivering superior classification results compared to other methods and underscoring the potential and advantages of MSFF in hyperspectral remote sensing classification. [ABSTRACT FROM AUTHOR]

The CNN-Transformer joint model stands as the leading architecture for contemporary hyperspectral image (HSI) classification, integrating global and local features through either successive or dual-branch CNN and Transformer networks. However, these methods often fall short in effectively incorporating spatial-spectral information with local-global attributes, resulting in incomplete feature representation. To address these challenges, we propose a spectral-spatial global-local interaction network that transmits global and local features into the spatial and spectral branches, facilitated by cross-interaction operators to ensure adequate feature flow. Initially, CNNs are employed to separately extract shallow features for the spectral and spatial branches. We then introduce a Spectral-Spatial Global-Local Interaction block designed for deep feature extraction, enhancing the flow of spectral and spatial features with local and global attributes using parallel transformers and dynamic convolutions. Transformers model the long-range dependencies of global spectral and spatial features, while dynamic convolutions enhance the context sensitivity of local spectral and spatial representations. Quadruple cross-interaction blocks are proposed to traverse both the spectral-spatial branches and local-global attribute dimensions, facilitating information exchange for complementary HSI representation. Extensive experiments and ablation studies on four public HSI datasets demonstrate the superiority of our proposed method. convolutional neural networks (CNNs); spectral and spatial; global and local; hyperspectral image classification; cross-interaction. [ABSTRACT FROM AUTHOR]

Hyperspectral image classification grapples with the twin challenges of high dimensionality and limited labelled data. These limitations hinder the development of generalizable classification models that can perform well across diverse datasets. To overcome these limitations, this paper proposes a novel semi-supervised framework that synergizes autoencoders, generative adversarial networks and zero-shot learning. This semi-supervised approach significantly improves feature extraction and data augmentation by harnessing the power of generative adversarial networks built upon autoencoders, ultimately enhancing classification accuracy. It further pushes the boundaries beyond traditional methods by enabling zero-shot learning, allowing the model to classify unseen data from classes not present in the training set. Additionally, the proposed model incorporates text embeddings to enrich feature representation, resulting in improved performance. This multimodal classification approach empowers the way for robust training and testing on cross-sensor datasets, even handling data with diverse spectra. Experimentally, it demonstrates remarkable accuracy across various domains, achieving a peak performance of 92.35% for cross-domain data and 91.83% for same-domain data, marking a significant leap forward in the generalizability of semi-supervised classification models. [ABSTRACT FROM AUTHOR]

The hyperspectral remote sensing images of agricultural crops contain rich spectral information, which can provide important details about crop growth status, diseases, and pests. However, existing crop classification methods face several key limitations when processing hyperspectral remote sensing images, primarily in the following aspects. First, the complex background in the images. Various elements in the background may have similar spectral characteristics to the crops, and this spectral similarity makes the classification model susceptible to background interference, thus reducing classification accuracy. Second, the differences in crop scales increase the difficulty of feature extraction. In different image regions, the scale of crops can vary significantly, and traditional classification methods often struggle to effectively capture this information. Additionally, due to the limitations of spectral information, especially under multi-scale variation backgrounds, the extraction of crop information becomes even more challenging, leading to instability in the classification results. To address these issues, a semantic-guided transformer network (SGTN) is proposed, which aims to effectively overcome the limitations of these deep learning methods and improve crop classification accuracy and robustness. First, a multi-scale spatial–spectral information extraction (MSIE) module is designed that effectively handle the variations of crops at different scales in the image, thereby extracting richer and more accurate features, and reducing the impact of scale changes. Second, a semantic-guided attention (SGA) module is proposed, which enhances the model's sensitivity to crop semantic information, further reducing background interference and improving the accuracy of crop area recognition. By combining the MSIE and SGA modules, the SGTN can focus on the semantic features of crops at multiple scales, thus generating more accurate classification results. Finally, a two-stage feature extraction structure is employed to further optimize the extraction of crop semantic features and enhance classification accuracy. The results show that on the Indian Pines, Pavia University, and Salinas benchmark datasets, the overall accuracies of the proposed model are 98.24%, 98.34%, and 97.89%, respectively. Compared with other methods, the model achieves better classification accuracy and generalization performance. In the future, the SGTN is expected to be applied to more agricultural remote sensing tasks, such as crop disease detection and yield prediction, providing more reliable technical support for precision agriculture and agricultural monitoring. [ABSTRACT FROM AUTHOR]

In recent years, convolutional neural networks (CNNs) have gained attention in the field of hyperspectral image classification. While CNNs demonstrate proficiency in capturing spatial nuances, their ability to model the continuous nature of spectral information remains challenging. In this paper, we propose a model that integrates CNNs for spatial feature extraction with Gated Recurrent Units (GRUs) to model the spectral dependencies. Recognizing the importance of specific bands and regions, dual attention mechanisms are introduced separately in both spectral and spatial modules. These attention modules emphasize crucial spectral bands and significant spatial regions, enhancing the discriminative power of the model. Chosen for its computational efficiency and ability to retain long-term dependencies, the GRU provides a balance between the performance and the computational cost of the proposed model. Experiments on benchmark hyperspectral datasets unveil that our model outperforms state-of-the-art methods, achieving remarkable improvements in classification accuracy. This fusion of architectures not only elevates the model's performance but also provides insights into the delicate balance between spatial and spectral nuances inherent in hyperspectral data. Our results set a new benchmark in hyperspectral image classification by highlighting the transformative potential of seamlessly integrating diverse neural network architectures while optimizing computational demands. [ABSTRACT FROM AUTHOR]

Recently, advancements in convolutional neural networks (CNNs) have significantly contributed to the advancement of hyperspectral image (HSI) classification. However, the problem of limited training samples is the primary obstacle to obtaining further improvements in HSI classification. The traditional methods relying solely on 2D-CNN for feature extraction underutilize the inter-band correlations of HSI, while the methods based on 3D-CNN alone for feature extraction lead to an increase in training parameters. To solve the above problems, we propose an HSI classification network based on hybrid depth-wise separable convolution and dual-branch feature fusion (HDCDF). The dual-branch structure is designed in HDCDF to extract simultaneously integrated spectral–spatial features and obtain complementary features via feature fusion. The proposed modules of 2D depth-wise separable convolution attention (2D-DCAttention) block and hybrid residual blocks are applied to the dual branch, respectively, further extracting more representative and comprehensive features. Instead of full 3D convolutions, HDCDF uses hybrid 2D–3D depth-wise separable convolutions, offering computational efficiency. Experiments are conducted on three benchmark HSI datasets: Indian Pines, University of Pavia, and Salinas Valley. The experimental results show that the proposed method showcases superior performance when the training samples are extremely limited, outpacing the state-of-the-art method by an average of 2.03% in the overall accuracy of three datasets, which shows that HDCDF has a certain potential in HSI classification. [ABSTRACT FROM AUTHOR]

Hyperspectral images are rich in spectral and spatial information, providing a detailed and comprehensive description of objects, which makes hyperspectral image analysis technology essential in intelligent agriculture. With various corn seed varieties exhibiting significant internal structural differences, accurate classification is crucial for planting, monitoring, and consumption. However, due to the large volume and complex features of hyperspectral corn image data, existing methods often fall short in feature extraction and utilization, leading to low classification accuracy. To address these issues, this paper proposes a spectral-spatial attention transformer network (SSATNet) for hyperspectral corn image classification. Specifically, SSATNet utilizes 3D and 2D convolutions to effectively extract local spatial, spectral, and textural features from the data while incorporating spectral and spatial morphological structures to understand the internal structure of the data better. Additionally, a transformer encoder with cross-attention extracts and refines feature information from a global perspective. Finally, a classifier generates the prediction results. Compared to existing state-of-the-art classification methods, our model performs better on the hyperspectral corn image dataset, demonstrating its effectiveness. [ABSTRACT FROM AUTHOR]

Remote sensing hyperspectral images (HSIs) are rich sources of information about land cover captured across hundreds of narrow, contiguous spectral wavelength bands. However, using the entire original HSI for practical applications can lead to suboptimal classification accuracy. To address this, band reduction techniques, categorized as feature extraction and feature selection methods, are employed to enhance classification results. One commonly used feature extraction approach for HSIs is Principal Component Analysis (PCA). However, PCA may fall short of capturing the local and specific characteristics present in the HSI data. In this paper, we introduce two novel feature extraction methods: Segmented Truncated Singular Value Decomposition (STSVD) and Spectrally Segmented Truncated Singular Value Decomposition (SSTSVD) to improve classification performance. Segmentation is carried out based on highly correlated bands' segments and spectral bands' segments within the HSI data. Our study evaluates and compares these newly proposed methods against classical feature extraction methods, including PCA, Incremental PCA, Sparse-PCA, Kernel PCA, Segmented-PCA (SPCA), and Truncated Singular Value Decomposition (TSVD). We perform this analysis on three distinct HSI datasets, namely the Indian Pines HSI, the Pavia University HSI, and the Kennedy Space Center HSI, using per-pixel Support Vector Machine (SVM) and Random Forest (RF) classification. The experimental results demonstrate the superiority of our proposed methods for all three datasets. The best-performing feature extraction methods when classification is performed using an SVM classifier are STSVD3 (89.03%), SSTSVD2 (95.55%), and STSVD3 (97.74%) for the Indian Pines, Pavia University, and Kennedy Space Center datasets, respectively. Similarly, for the RF classifier, the best-performing feature extraction methods are SSTSVD4 (88.98%), SSTSVD3 (96.04%), and SSTSVD4 (96.09%) for Indian Pines, Pavia University, and Kennedy Space Center datasets, respectively. [ABSTRACT FROM AUTHOR]

Hyperspectral image classification in remote sensing often encounters challenges due to limited annotated data. Semi-supervised learning methods present a promising solution. However, their performance is heavily influenced by the quality of pseudo labels. This limitation is particularly pronounced during the early stages of training, when the model lacks adequate prior knowledge. In this paper, we propose an Iterative Pseudo Label Generation (IPG) framework based on the Segment Anything Model (SAM) to harness structural prior information for semi-supervised hyperspectral image classification. We begin by using a small number of annotated labels as SAM point prompts to generate initial segmentation masks. Next, we introduce a spectral voting strategy that aggregates segmentation masks from multiple spectral bands into a unified mask. To ensure the reliability of pseudo labels, we design a spatial-information-consistency-driven loss function that optimizes IPG to adaptively select the most dependable pseudo labels from the unified mask. These selected pseudo labels serve as iterative point prompts for SAM. Following a suitable number of iterations, the resultant pseudo labels can be employed to enrich the training data for the classification model. Experiments conducted on the Indian Pines and Pavia University datasets demonstrate that even a simple 2D CNN based classification model trained with our generated pseudo labels significantly outperforms eight state-of-the-art hyperspectral image classification methods. [ABSTRACT FROM AUTHOR]

Hyperspectral image consists of multiple contiguous spectral bands, which are crucial for precise land classification. In earlier studies, convolutional neural network has emerged as effective methods for hyperspectral image classification due to their powerful feature extraction capabilities. Recently, vision transformers have been applied in the field of hyperspectral image classification. However, most existing transformer methods mainly focus on global relationships, lacking the ability to capture multiscale features, leading to subpar classification performance. To address this issue, this paper proposes a hyperspectral image classification with embedded linear vision transformer (ELViT). Firstly, the ELViT employs a token generator to embed multiscale semantic tokens, providing the model with richer image features by leveraging the local representation capability of convolutional neural network. Then, We propose a transformer with linear complexity designed to capture global correlations between different tokens, allowing the model to prioritize distinctive feature information. Additionally, a gated linear unit activation function is utilized to supplement the establishment of long-range relationships in the transformer. Experimental results demonstrate that ELViT outperforms both convolutional neural network based and transformer based methods, achieving excellent classification performance with overall accuracies of 98.43%, 99.67%, and 99.27% on the Pavia University, Salinas Valley, and WHU-Hi-LongKou datasets, respectively. [ABSTRACT FROM AUTHOR]

The dense and nearly continuous spectral bands in hyperspectral images result in strong inter-band correlations, which can diminish performance of the model in classification tasks. Moreover, most convolutional neural network-based methods for hyperspectral image classification typically depend on a fixed scale to extract spectral–spatial features, which ignore the detail features of some objects. To address the above issues, a novelty Spectral Spatial Adaptive Weighted Fusion and Residual Dense Network (S2AWF-RDN) is proposed for Hyperspectral image classification. Specifically, the proposed S2AWF-RDN consists of spectral–spatial adaptive weighted fusion module, multi-channel feature concatenation residual dense module, and spatial feature fusion module. Firstly, the spectral information optimization branch is developed to adjust the weights assigned to various spectral channels. Similarly, the spatial information optimization branch is developed to adjust the weights for different spatial regions. Secondly, to obtain rich spectral spatial information from different levels, multi-channel feature concatenation residual dense module has been proposed. In addition, a multi-channel feature concatenation block is designed guiding the model to extract spectral spatial information at different scales. Finally, spatial feature fusion module is introduced to retain more spatial information. The experimental outcomes illustrate that the proposed network model exhibits superior classification performance on three renowned hyperspectral image datasets. Furthermore, the efficacy of the proposed network model is further corroborated through comparative and ablation studies.

Radiofrequency ablation (RFA) is used to treat atrial fibrillation (AF). Viability gaps between ablated regions can lead to AF recurrence; thus, correctly detecting RFA lesions is important for successful treatment. Hyperspectral imaging (HSI) has been previously shown to aid in visualizing RFA-affected areas. However, automated classification of ablated/unabled tissue in HSI remains open. We comparatively analyze state-of-the-art deep learning models for HSI classification from the field of remote sensing. We group them based on model architecture and utilization of spectral and spatial contexts. We deploy a pre-clinical HSI dataset obtained from excised left atrial tissue of four porcine and five bovine samples. The 45 classification algorithms are benchmarked through two types of experiments: objective and subjective. For objective evaluation, per HSI cube, we limit the training and testing to rectangular areas of known classes (ablated/unablated), as reliable classification masks for whole HSI cubes are not available. For both porcine and bovine data, models incorporating graph neural networks demonstrate superior performance. The best eight algorithms are further compared in a subjective experiment. We ask seventeen participants to rate RGB visualizations of model outputs. In this experiment, the performance of an algorithm is assessed on the whole HSI cube, and is subject to participants’ interpretation of the extent of RFA-affected tissue. Our comparative analysis allows us to observe the suitability of different deep learning architectures for HSI classification of affected tissue. These efforts contribute towards the potential future utilization of HSI for surgical guidance in RFA or similar procedures.

Hyperspectral data acquired through remote sensing are invaluable for environmental and resource studies. While rich in spectral information, various complexities, such as environmental conditions, material properties, and sensor characteristics can cause significant variability even among pixels belonging to the same material class. This variability poses nuisance for accurate land-cover classification and analysis. Focusing on the spectral domain, we utilize an autoencoder architecture called the symmetric autoencoder (SymAE), which leverages permutation invariant representation and stochastic regularization in tandem to disentangle class-invariant “coherent” features from variability-causing “nuisance” features on a pixel-by-pixel basis. This disentanglement is achieved through a purely data-driven process, without the need for hand-crafted modeling, noise distribution priors, or reference “clean signals.” In addition, SymAE can generate virtual spectra through manipulations in latent space. Using AVIRIS instrument data, we demonstrate these virtual spectra offer insights on the disentanglement. Extensive experiments across six benchmark hyperspectral datasets show that coherent features extracted by SymAE can be used to achieve state-of-the-art pixel-based classification. Furthermore, we leverage these coherent features to enhance the performance of some leading spectral–spatial hyperspectral image (HSI) classification methods. Our approach especially shows improvement in scenarios where training and test regions are disjoint, a common challenge in real-world applications where existing methods often struggle to maintain relatively high performance.

Recently, the transformer-based model has shown superior performance in hyperspectral image classification (HSIC) due to its excellent ability to model long-term dependencies on sequence data. An important component of the transformer is the tokenizer, which can transform the features into semantic token sequences (STS). Nonetheless, transformer's semantic tokenization strategy is hardly representative of local relatively important high-level semantics because of its global receptive field. Recently, the Mamba-based methods have shown even stronger spatial context modeling ability than Transformer for HSIC. However, these Mamba-based methods mainly focus on spectral and spatial dimensions. They tend to extract semantic information in very long feature sequences or represent semantic information in several typical tokens, which may ignore some important semantics of the HSIs. In order to represent the semantic information of HSIs more holistically in Mamba, this article proposes a semantic tokenization-based Mamba (STMamba) model. In STMamba, a spectral-spatial feature extraction module is used to extract the spectral–spatial joint features. Then, a generated semantic token sequences module is designed to transform the features into STS. Subsequently, the STS are fed into the semantic token state spatial model to capture relationships between different semantic tokens. Finally, the fused semantic token is passed into a classifier for classification. Experimental results on three HSI datasets demonstrate that the proposed STMamba outperforms existing state-of-the-art deep learning and transformer-based methods.

In recent years, significant advances in unmanned aerial vehicle (UAV) technology and hyperspectral remote sensing have spurred rapid and innovative developments in UAV-based hyperspectral image (HSI) classification across a range of fields, including environmental monitoring, precision agriculture, forest health assessment, and disaster management. Compared to spaceborne platforms, the spectra of ground objects observed by UAV platforms exhibit notable variations, presenting more pronounced challenges for accurate classification. This article provides an in-depth and systematic review of UAV HSI classification techniques, systematically examining the evolution from traditional machine learning approaches, such as sparse coding, compressed sensing, and kernel methods, to cutting-edge deep learning frameworks, including convolutional neural networks, Transformer models, recurrent neural networks, graph convolutional networks, generative adversarial networks, and hybrid models. Although traditional methods demonstrate effectiveness in certain scenarios, their limitations become increasingly apparent when dealing with high-dimensional, nonlinear spectral data. In contrast, deep learning-based models excel at capturing intricate relationships between spectral and spatial features, significantly boosting classification accuracy and emerging as the dominant paradigm in the field. The WHU-Hi hyperspectral remote sensing dataset is utilized as a case study to elucidate the advantages and limitations of various deep learning methods through rigorous qualitative and quantitative comparisons. The potential of UAV hyperspectral image classification techniques in addressing high-dimensional data and complex scenarios is also thoroughly described. Furthermore, this article delves into cutting-edge research trends, such as lightweight model development, hyperspectral large models, multisource data fusion, and model interpretability, while also highlighting future trends for UAV hyperspectral remote sensing classification technology, particularly in real-time monitoring and intelligent applications.

Extracting discriminative spectral-spatial features from hyperspectral images (HSIs) remains a crucial topic within the remote sensing community. However, most feature extraction methods suffer from coarse textures, leading to poor performance in classifying HSIs. Moreover, limited training samples can significantly hinder the generalization ability of classifiers, hindering performance improvements in HSI classification. To address these issues, a proposed dual-branch structural feature extraction method fully leverages the global and local information, preserving key structures while removing unimportant textures. The key steps of our method are detailed below: First, an image fusion technique is performed on HSIs to obtain dimension-reduced HSIs. Second, a global multiscale feature extraction method is proposed by performing different parameter settings of pyramid texture filtering (PTF) on HSIs and global probability can be obtained. Third, a PTF-based local multiscale feature extraction method is proposed using multiscale superpixel segmentation on the finest layer of PTFs and local probability can be obtained. Fourth, a decision fusion rule is applied to integrate class probabilities from the two branches. Experiments performed on three different datasets demonstrate that the proposed method can outperform other state-of-the-art classification methods.

Advancements in sensor technologies have brought about significant improvements in the resolution and quality of imagery by enhancing spatial, temporal, spectral, and radiometric aspects. These remarkable progressions have sparked enhancements in hyperspectral image classification (HSIC) applications, including land cover mapping, vegetation classification, urban monitoring, and resource understanding, which are crucial for optimal earth resource management. Effective HSIC demands advanced algorithms that exhibit high accuracy, low computational complexity, and robustness in extracting intricate spectral-spatial features. The advent of deep convolutional neural networks (DCNNs) has revolutionized image classification, introducing robust architectures that continue to evolve. However, a notable challenge remains in supervised HSIC due to the scarcity of training samples, a bottleneck that has yet to be comprehensively addressed in the literature. To catalyze further exploration, this study reviews existing methods designed to mitigate the limitations posed by limited labeled data. It also examines current techniques for feature learning in HSIC using DCNNs. Additionally, the study presents results obtained from various methods applied to the most widely recognized public HSIC datasets, accompanied by insightful observations that lay the groundwork for future research endeavors within the hyperspectral community. [ABSTRACT FROM AUTHOR]

Abstract Hyperspectral image (HSI) data has a wide range of valuable spectral information for numerous tasks. HSI data encounters challenges such as small training samples, scarcity, and redundant information. Researchers have introduced various research works to address these challenges. Convolution Neural Network (CNN) has gained significant success in the field of HSI classification. CNN’s primary focus is to extract low-level features from HSI data, and it has a limited ability to detect long-range dependencies due to the confined filter size. In contrast, vision transformers exhibit great success in the HSI classification field due to the use of attention mechanisms to learn the long-range dependencies. As mentioned earlier, the primary issue with these models is that they require sufficient labeled training data. To address this challenge, we proposed a spectral-spatial feature extractor group attention transformer that consists of a multiscale feature extractor to extract low-level or shallow features. For high-level semantic feature extraction, we proposed a group attention mechanism. Our proposed model is evaluated using four publicly available HSI datasets, which are Indian Pines, Pavia University, Salinas, and the KSC dataset. Our proposed approach achieved the best classification results in terms of overall accuracy (OA), average accuracy (AA), and Kappa coefficient. As mentioned earlier, the proposed approach utilized only 5%, 1%, 1%, and 10% of the training samples from the publicly available four datasets.

Hyperspectral image classification aims to classify pixels in hyperspectral image into different land-cover classes. Considering multi-view data can improve the classification performance as it contains more abundant information than single view, a multi-view based method is proposed for hyperspectral image classification in this paper. In the proposed method, spatial neighboring pixels are adaptively determined by the mean and deviation of samples in the fixed-size spatial window. Then, the sparse representation algorithm is realized to obtain the sparse coefficients for the multi-view matrix set of the spatial neighboring pixels. Moreover, the renyi-entropy is calculated to determine the weight of each view in the classification task since the performance of different views is different. Finally, the obtained sparse coefficients and weights are jointly used to determine the class label of each test pixel. Experimental results on the Indian Pines dataset and University of Pavia dataset demonstrate the superior performance of the proposed method compared to several state-of-art methods. [ABSTRACT FROM AUTHOR]

Recently, there has been a surge in the adoption of deep learning (DL) techniques, especially convolutional neural networks (CNNs), to perform hyperspectral image (HSI) classification. Although deep learners have been shown to achieve impressive performance in HSI classification, they are known to be extremely sensitive to even slight perturbations to their inputs and models. When applied in safety-critical applications, it is crucial to know how robust they really are against perturbations. However, there is still limited tool support for DL testing in terms of their robustness, nor are the existing RGB testing approaches able to address the HSI-specific challenges. In this paper, we propose a mutation analysis framework specialized for DL models trained to classify HSIs, which facilitates a critical evaluation of the robustness of DL-based HSI classifiers. First, we introduce a set of mutation operators to inject faults into the inputs and models to simulate distortions of remote sensing HSI classifiers. By utilizing the mutation testing technique, we implement a novel framework which supports the multidimensional evaluation of individual DL-based classifiers. Finally, a comparative study of the robustness of seven popular CNN-based HSI classifiers (i.e., 3D-CNN, FDSSC, HybridSN, MCNN, FC3DCNN, DWTDENSE, and Tri-CNN) on six HSI datasets is provided. Results show that FDSSC and Tri-CNN achieve higher robustness in the presence of distortions, and FDSSC maintains a relatively stable level of robustness even with few training samples. These empirical findings can be partly explained by the characteristics of the classifiers' architectures. The results substantiate the efficacy of our evaluation framework in assessing the robustness of HSI classifiers and thus confirm its contribution to the field of remote sensing image classification. [ABSTRACT FROM AUTHOR]