Hyperspectral Image Classification Research Articles

Hybrid convolutional network with enhanced graph attention mechanism (HCN-EGAM) for hyperspectral image classification

ABSTRACT Due to their reliance on superficial spectral or spatial information, current hyperspectral image classification techniques frequently fail to generate similarity measures that are accurate enough to create graph structures. We have developed a new end-to-end model, the hybrid convolutional network with enhanced graph attention mechanism (HCN-EGAM), to address this issue. It integrates an enhanced graph attention mechanism with 3D–2D hybrid convolution. In order to improve classification accuracy, this model synergistically uses the graph attention network module with the hybrid convolutional feature extraction module. The model starts by creating a 3D–2D hybrid convolutional network. This lets it quickly pull out deep spatial-spectral features that make different ground objects stand out in hyperspectral images. The model then employs these deep spatial-spectral properties to construct a graph structure, thereby enhancing its feature representation capabilities. Lastly, we use the Enhanced Graph Attention Mechanism (EGAM) to learn long-range spatial associations, successfully distinguishing between intra-class differences and inter-class similarities between distinct samples. Comparing our suggested method to other state-of-the-art methods, experimental assessments on three datasets, including the University of Pavia, Kennedy Space Center (KSC) and Indian Pines (IP) datasets, show a considerable improvement in classification accuracy.

SSFAN: A Compact and Efficient Spectral-Spatial Feature Extraction and Attention-Based Neural Network for Hyperspectral Image Classification

Hyperspectral image (HSI) classification is a crucial technique that assigns each pixel in an image to a specific land cover category by leveraging both spectral and spatial information. In recent years, HSI classification methods based on convolutional neural networks (CNNs) and Transformers have significantly improved performance due to their strong feature extraction capabilities. However, these improvements often come with increased model complexity, leading to higher computational costs. To address this, we propose a compact and efficient spectral-spatial feature extraction and attention-based neural network (SSFAN) for HSI classification. The SSFAN model consists of three core modules: the Parallel Spectral-Spatial Feature Extraction Block (PSSB), the Scan Block, and the Squeeze-and-Excitation MLP Block (SEMB). After preprocessing the HSI data, it is fed into the PSSB module, which contains two parallel streams, each comprising a 3D convolutional layer and a 2D convolutional layer. The 3D convolutional layer extracts spectral and spatial features from the input hyperspectral data, while the 2D convolutional layer further enhances the spatial feature representation. Next, the Scan Block module employs a layered scanning strategy to extract spatial information at different scales from the central pixel outward, enabling the model to capture both local and global spatial relationships. The SEMB module combines the Spectral-Spatial Recurrent Block (SSRB) and the MLP Block. The SSRB, with its adaptive weight assignment mechanism in the SToken Module, flexibly handles time steps and feature dimensions, performing deep spectral and spatial feature extraction through multiple state updates. Finally, the MLP Block processes the input features through a series of linear transformations, GELU activation functions, and Dropout layers, capturing complex patterns and relationships within the data, and concludes with an argmax layer for classification. Experimental results show that the proposed SSFAN model delivers superior classification performance, outperforming the second-best method by 1.72%, 5.19%, and 1.94% in OA, AA, and Kappa coefficient, respectively, on the Indian Pines dataset. Additionally, it requires less training and testing time compared to other state-of-the-art deep learning methods.

Endoscopic In Vivo Hyperspectral Imaging for Head and Neck Tumor Surgeries Using a Medically Approved CE-Certified Camera with Rapid Visualization During Surgery

Background: In vivo visualization of malignant tumors remains the main challenge during head and neck cancer surgery. This can result in inadequate tumor margin assessment and incomplete tumor resection, adversely affecting patient outcomes. Hyperspectral imaging (HSI) is a promising approach to address this issue. However, its application in surgery has been limited by the lack of medically approved HSI devices compliant with MDR regulations, as well as challenges regarding the integration into the surgical workflow. Methods: In this feasibility study, we employed endoscopic HSI during surgery to visualize the tumor sites of 12 head and neck cancer patients. We optimized the HSI workflow to minimize time required during surgery and to reduce the adaptation period needed for surgeons to adjust to the new workflow. Additionally, we implemented data processing to enable real-time classification and visualization of HSI within the intraoperative setting. HSI evaluation was conducted using principal component analysis and k-means clustering, with this clustering validated through comparison with expert annotations. Results: Our complete HSI workflow requires two to three minutes, with each HSI measurement—including evaluation and visualization—taking less than 10 s, achieving an accuracy of 79%, sensitivity of 72%, and specificity of 84%. Medical personnel became proficient with the HSI system after two surgeries. Conclusions: This study presents an HSI workflow for in vivo tissue differentiation during head and neck cancer surgery, providing accurate and visually accessible results within minimal time. This approach enhances the in vivo evaluation of tumor margins, leading to more clear margins and, consequently, improved patient outcomes.

SSML: Semi-supervised metric learning with hard samples for hyperspectral image classification

Deep learning is widely used in hyperspectral image (HSI) classification due to its powerful learning capabilities. However, its excellent performance typically requires a large number of samples, which can be time-consuming and labor-intensive to produce. The limitation of available samples greatly constrains the model's generalization performance. To alleviate the sample pressure, we propose a semi-supervised metric learning method for HSI classification that focuses on hard samples and can obtain reliable pseudo-labels through multiscale prediction. Additionally, we employ a hard sample learning strategy for network training, which concentrates on enhancing network discrimination by adaptively optimizing intra-class and inter-class distances. Performance tests on three public datasets indicate that the proposed method surpasses other state-of-the-art methods across multiple validation metrics.

Tensor-Based Few-Shot Learning for Cross-Domain Hyperspectral Image Classification

Tensor-Based Few-Shot Learning for Cross-Domain Hyperspectral Image Classification

Residual channel attention based sample adaptation few-shot learning for hyperspectral image classification.

Few-shot learning (FSL) uses prior knowledge and supervised experience to effectively classify hyperspectral images (HSIs), thereby reducing the cost of large numbers of labeled samples. However, existing few-shot methods ignore the correlation between cross-domain feature channels, and the feature representation ability is insufficient. To address above issue, this paper proposes a novel Residual Channel Attention Based Sample Adaptation Few-Shot Learning for Hyperspectral Image Classification(RCASA-FSL) for hyperspectral image classification (HSIC), which can capture and enhance cross-domain dependencies through multi-layer residual connection and random-based feature recalibration. Specifically, a Deep Residual Feature Channel Attention Mechanism (DRFCAM) is designed to obtain cross-domain dependencies by residual concatenation, and further the residual structure is stacked for mining depth discrimination information. Furthermore, a new Random-based Feature Recalibration Module (RFRM) is proposed to reassign the feature weights via random matrix, which fully explore feature weight relationships to guide the sample adaptation process. Besides, we design a joint loss function with combining the FSL loss and domain adaptive loss for further optimization model. Experiments conducted on several standard hyperspectral datasets demonstrate that the proposed RCASA-FSL is superior to other FSL techniques in both quantitative and qualitative aspects.

Segmentation-based truncated-SVD for effective feature extraction in hyperspectral image classification

ABSTRACT 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.

Center-Highlighted Multiscale CNN for Classification of Hyperspectral Images

Hyperspectral images (HSIs) capture a wide range of spectral features across multiple bands of light, from visible to near-infrared. Hyperspectral image classification technology enables researchers to accurately identify and analyze the composition and distribution of surface materials. Current mainstream deep learning methods typically use block sampling to capture spatial features for the model. However, this approach can affect classification results due to the influence of neighboring features within the sample block. To improve the model’s focus on the center of the sampling block, this study proposes a center highlight with multiscale CNN for hyperspectral image classification (CHMSC). The network utilizes an automatic channel selector (Auto-CHS) to fully consider every channel feature and capture the correlation between the channels. Then, CHMSC enhances the model’s ability to concentrate on the central features of the sampling block utilizing structures such as the center highlight. Finally, before outputting the prediction results, an SENet is employed to further refine the features and learn associate interactions between different scales of spatial features and spectral features. Experimental results from three hyperspectral datasets validate the effectiveness of the proposed method. Specifically, when 15 samples from each class are selected for training, CHMSC achieves the highest overall accuracy (OA) of 90.05%, 92.78%, and 90.15% on the three datasets, outperforming other methods with increases of more than 3.11%, 1.8%, and 2.01% in OA, respectively.

How to Learn More? Exploring Kolmogorov–Arnold Networks for Hyperspectral Image Classification

Convolutional neural networks (CNNs) and vision transformers (ViTs) have shown excellent capability in complex hyperspectral image (HSI) classification. However, these models require a significant number of training data and are computational resources. On the other hand, modern Multi-Layer Perceptrons (MLPs) have demonstrated a great classification capability. These modern MLP-based models require significantly less training data compared with CNNs and ViTs, achieving state-of-the-art classification accuracy. Recently, Kolmogorov–Arnold networks (KANs) were proposed as viable alternatives for MLPs. Because of their internal similarity to splines and their external similarity to MLPs, KANs are able to optimize learned features with remarkable accuracy, in addition to being able to learn new features. Thus, in this study, we assessed the effectiveness of KANs for complex HSI data classification. Moreover, to enhance the HSI classification accuracy obtained by the KANs, we developed and proposed a hybrid architecture utilizing 1D, 2D, and 3D KANs. To demonstrate the effectiveness of the proposed KAN architecture, we conducted extensive experiments on three newly created HSI benchmark datasets: QUH-Pingan, QUH-Tangdaowan, and QUH-Qingyun. The results underscored the competitive or better capability of the developed hybrid KAN-based model across these benchmark datasets over several other CNN- and ViT-based algorithms, including 1D-CNN, 2DCNN, 3D CNN, VGG-16, ResNet-50, EfficientNet, RNN, and ViT.

Bidirectional Mamba with Dual-Branch Feature Extraction for Hyperspectral Image Classification.

The hyperspectral image (HSI) classification task is widely used in remote sensing image analysis. The HSI classification methods based on convolutional neural networks (CNNs) have greatly improved the classification performance. However, they cannot well utilize the sequential properties of spectral features and face the challenge of increasing computational cost with the increase in network depth. To address these shortcomings, this paper proposes a novel network with a CNN-Mamba architecture, called DBMamba, which uses a bidirectional Mamba to process spectral feature sequences at a linear computational cost. In the DBMamba, principal component analysis (PCA) is first used to extract the main features of the data. Then, a dual-branch CNN structure, with the fused features from spectral-spatial features by 3D-CNN and spatial features by 2D-CNN, is used to extract shallow spectral-spatial features. Finally, a bidirectional Mamba is used to effectively capture global contextual information in features and significantly enhance the extraction of spectral features. Experimental results on the Indian Pines, Salinas, and Pavia University datasets demonstrate that the classification performance surpasses that of many cutting-edge methods, improving by 1.04%, 0.15%, and 0.09%, respectively, over the competing SSFTT method. The research in this paper enhances the existing knowledge on HSI classification and provides valuable insights for future research in this field.

Capsule Attention Network for Hyperspectral Image Classification

While many neural networks have been proposed for hyperspectral image classification, current backbones cannot achieve accurate results due to the insufficient representation by scalar features and always cause a cumbersome calculation burden. To solve the problem, we propose the capsule attention network (CAN), which combines an activity vector with an attention mechanism to improve HSI classification. In particular, we consider two attention mechanisms to improve the effectiveness of the activity vectors. First, an attention-based feature extraction (AFE) module is proposed to preprocess the spectral-spatial features of HSI data, which effectively mines useful information before the generation of the activity vectors. Second, we propose a self-weighted mechanism (SWM) to distinguish the importance of different capsule convolutions, which enhances the representation of the primary activity vectors. Experiments on four well-known HSI datasets have shown our CAN surpasses state-of-the-art (SOTA) methods on three widely used metrics with a much lower computational burden.

Hyperspectral Image Classification Algorithm for Forest Analysis Based on a Group-Sensitive Selective Perceptual Transformer

Substantial advancements have been achieved in hyperspectral image (HSI) classification through contemporary deep learning techniques. Nevertheless, the incorporation of an excessive number of irrelevant tokens in large-scale remote sensing data results in inefficient long-range modeling. To overcome this hurdle, this study introduces the Group-Sensitive Selective Perception Transformer (GSAT) framework, which builds upon the Vision Transformer (ViT) to enhance HSI classification outcomes. The innovation of the GSAT architecture is primarily evident in several key aspects. Firstly, the GSAT incorporates a Group-Sensitive Pixel Group Mapping (PGM) module, which organizes pixels into distinct groups. This allows the global self-attention mechanism to function within these groupings, effectively capturing local interdependencies within spectral channels. This grouping tactic not only boosts the model’s spatial awareness but also lessens computational complexity, enhancing overall efficiency. Secondly, the GSAT addresses the detrimental effects of superfluous tokens on model efficacy by introducing the Sensitivity Selection Framework (SSF) module. This module selectively identifies the most pertinent tokens for classification purposes, thereby minimizing distractions from extraneous information and bolstering the model’s representational strength. Furthermore, the SSF refines local representation through multi-scale feature selection, enabling the model to more effectively encapsulate feature data across various scales. Additionally, the GSAT architecture adeptly represents both global and local features of HSI data by merging global self-attention with local feature extraction. This integration strategy not only elevates classification precision but also enhances the model’s versatility in navigating complex scenes, particularly in urban mapping scenarios where it significantly outclasses previous deep learning methods. The advent of the GSAT architecture not only rectifies the inefficiencies of traditional deep learning approaches in processing extensive remote sensing imagery but also markededly enhances the performance of HSI classification tasks through the deployment of group-sensitive and selective perception mechanisms. It presents a novel viewpoint within the domain of hyperspectral image classification and is poised to propel further advancements in the field. Empirical testing on six standard HSI datasets confirms the superior performance of the proposed GSAT method in HSI classification, especially within urban mapping contexts, where it exceeds the capabilities of prior deep learning techniques. In essence, the GSAT architecture markedly refines HSI classification by pioneering group-sensitive pixel group mapping and selective perception mechanisms, heralding a significant breakthrough in hyperspectral image processing.

DMCCT: Dual-Branch Multi-Granularity Convolutional Cross-Substitution Transformer for Hyperspectral Image Classification

In the field of hyperspectral image classification, deep learning technology, especially convolutional neural networks, has achieved remarkable progress. However, convolutional neural network models encounter challenges in hyperspectral image classification due to limitations in their receptive fields. Conversely, the global modeling capability of Transformers has garnered attention in hyperspectral image classification. Nevertheless, the high computational cost and inadequate local feature extraction hinder its widespread application. In this study, we propose a novel fusion model of convolutional neural networks and Transformers to enhance performance in hyperspectral image classification, namely the dual-branch multi-granularity convolutional cross-substitution Transformer (DMCCT). The proposed model adopts a dual-branch structure to separately extract spatial and spectral features, thereby mitigating mutual interference and information loss between spectral and spatial data during feature extraction. Moreover, a multi-granularity embedding module is introduced to facilitate multi-scale and multi-level local feature extraction for spatial and spectral information. In particular, the improved convolutional cross-substitution Transformer module effectively integrates convolution and Transformer, reducing the complexity of attention operations and enhancing the accuracy of hyperspectral image classification tasks. Subsequently, the proposed method is evaluated against existing approaches using three classical datasets, namely Pavia University, Kennedy Space Center, and Indian Pines. Experimental results demonstrate the efficacy of the proposed method, achieving significant classification results on these datasets with overall classification accuracies of 98.57%, 97.96%, and 96.59%, respectively. These results establish the superiority of the proposed method in the context of hyperspectral image classification under similar experimental conditions.

A Low-Measurement-Cost-Based Multi-Strategy Hyperspectral Image Classification Scheme.

The cost of hyperspectral image (HSI) classification primarily stems from the annotation of image pixels. In real-world classification scenarios, the measurement and annotation process is both time-consuming and labor-intensive. Therefore, reducing the number of labeled pixels while maintaining classification accuracy is a key research focus in HSI classification. This paper introduces a multi-strategy triple network classifier (MSTNC) to address the issue of limited labeled data in HSI classification by improving learning strategies. First, we use the contrast learning strategy to design a lightweight triple network classifier (TNC) with low sample dependence. Due to the construction of triple sample pairs, the number of labeled samples can be increased, which is beneficial for extracting intra-class and inter-class features of pixels. Second, an active learning strategy is used to label the most valuable pixels, improving the quality of the labeled data. To address the difficulty of sampling effectively under extremely limited labeling budgets, we propose a new feature-mixed active learning (FMAL) method to query valuable samples. Fine-tuning is then used to help the MSTNC learn a more comprehensive feature distribution, reducing the model's dependence on accuracy when querying samples. Therefore, the sample quality is improved. Finally, we propose an innovative dual-threshold pseudo-active learning (DSPAL) strategy, filtering out pseudo-label samples with both high confidence and uncertainty. Extending the training set without increasing the labeling cost further improves the classification accuracy of the model. Extensive experiments are conducted on three benchmark HSI datasets. Across various labeling ratios, the MSTNC outperforms several state-of-the-art methods. In particular, under extreme small-sample conditions (five samples per class), the overall accuracy reaches 82.97% (IP), 87.94% (PU), and 86.57% (WHU).

RS-Net: Hyperspectral Image Land Cover Classification Based on Spectral Imager Combined with Random Forest Algorithm

Recursive neural networks and transformers have recently become dominant in hyperspectral (HS) image classification due to their ability to capture long-range dependencies in spectral sequences. Despite the success of these sequential architectures, mainstream deep learning methods primarily handle two-dimensional structured data. However, challenges such as the curse of dimensionality, spectral variability, and confounding factors in hyperspectral remote sensing images limit their effectiveness, especially in remote sensing applications. To address this issue, this paper proposes a novel land cover classification algorithm that integrates random forests with a spectral transformer network structure (RS-Net). Firstly, this paper presents a combination of the Gramian Angular Field (GASF) and Gramian Angular Difference Field (GADF) algorithms, which effectively maps the multidimensional time series constructed for each pixel onto two-dimensional image features, enabling precise extraction and recognition in the backend network algorithms and improving the classification accuracy of land cover types. Secondly, to capture the relationships between features at different scales, this paper proposes a SpectralFormer network architecture using the Context and Structure Encoding (CASE) module to effectively learn dependencies between channels. This architecture enhances important features and suppresses unimportant ones, thereby addressing the semantic gap and improving the recognition capability of land cover features. Finally, the final prediction results are determined by a voting mechanism from the Random Forest algorithm, which synthesizes predictions from multiple decision trees to enhance classification stability and accuracy. To better compare the performance of RS-Net, this paper conducted extensive experiments on three benchmark HS datasets obtained from satellite and airborne imagers, comparing various classic neural network models. Surprisingly, the RS-Net algorithm achieves high performance and efficiency, offering a new and effective tool for land cover classification.

GroupFormer for hyperspectral image classification through group attention

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.

Fast Hyperspectral Image Classification with Strong Noise Robustness Based on Minimum Noise Fraction

A fast hyperspectral image classification algorithm with strong noise robustness is proposed in this paper, aiming at the hyperspectral image classification problems under noise interference. Based on the Fast 3D Convolutional Neural Network (Fast-3DCNN), this algorithm enables the classification model to have good tolerance for various types of noise by using a Minimum Noise Fraction (MNF) as dimensionality reduction module for hyperspectral image input data. In addition, by introducing lightweight hybrid attention modules with the spatial and the channel information, the deep features extracted by the Convolutional Neural Network are further refined, ensuring that the model has high classification accuracy. Public dataset experiments have shown that compared to traditional methods, the MNF in this algorithm reduces the dimensionality of input spectral data, preserves information with higher signal-to-noise ratio(SNR) in the spectral bands, and aggregates spectral features into class feature vectors, greatly improving the noise robustness of the model. At the same time, based on a lightweight spectral–spatial hybrid attention mechanism, combined with fewer spectral dimensions, the model effectively avoids overfitting. With less loss in model training speed, it achieved better classification accuracy in small-scale training sample experiments, fully demonstrating the good generalization ability of this algorithm.

MultiScale spectral–spatial convolutional transformer for hyperspectral image classification

AbstractDue to the powerful ability in capturing the global information, transformer has become an alternative architecture of CNNs for hyperspectral image classification. However, general transformer mainly considers the global spectral information while ignores the multiscale spatial information of the hyperspectral image. In this paper, we propose a multiscale spectral–spatial convolutional transformer (MultiFormer) for hyperspectral image classification. First, the developed method utilizes multiscale spatial patches as tokens to formulate the spatial transformer and generates multiscale spatial representation of each band in each pixel. Second, the spatial representation of all the bands in a given pixel are utilized as tokens to formulate the spectral transformer and generate the multiscale spectral–spatial representation of each pixel. Besides, a modified spectral–spatial CAF module is constructed in the MultiFormer to fuse cross‐layer spectral and spatial information. Therefore, the proposed MultiFormer can capture the multiscale spectral–spatial information and provide better performance than most of other architectures for hyperspectral image classification. Experiments are conducted over commonly used real‐world datasets and the comparison results show the superiority of the proposed method.

A graph-guided transformer based on dual-stream perception for hyperspectral image classification

ABSTRACT The excellent capabilities of Transformers and Graph Neural Networks (GNNs) in modelling long-range dependencies and handling irregular data have led to their widespread application in hyperspectral image (HSI) classification tasks. However, the Graph Transformer combining both advantages is rarely used in this field and has some limitations. Current Graph Transformers consider interactions between all nodes within the graph, adding complexity and introducing unnecessary information from noisy nodes. Moreover, the rich spectral information in HSIs is often ignored, and there is a lack of effective fusion of spatial information. In this paper, we propose a dual-stream graph-guided Transformer for HSI classification. In spatial dimension, superpixels are utilized to guide spatial graph generation, capturing global topological dependencies and local details effectively in HSIs. In terms of spectrum, we innovatively construct a spectral graph based on spectral channels and adopt a contribution score-based strategy to adaptively filter out irrelevant edges, achieving sparsity while preserving spectral context relationships. Experimental results demonstrate the significant competitive advantage of our method in HSI classification tasks on three public datasets. The code is available at https://github.com/youngboy03/GTDPNet.

Hyperspectral image classification with token fusion on GPU

Hyperspectral images capture material nuances with spectral data, vital for remote sensing. Transformer has become a mainstream approach for tackling the challenges posed by high-dimensional hyperspectral data with complex structures. However, a major challenge they face when processing hyperspectral images is the presence of a large number of redundant tokens, which leads to a significant increase in computational load, adding to the model’s computational burden and affecting inference speed. Therefore, we propose a token fusion algorithm tailored to the operational characteristics of the hyperspectral image and pure transformer network, aimed at enhancing the final accuracy and throughput of the model. The token fusion algorithm introduces a token merging step between the attention mechanism and the multi-layer perceptron module in each Transformer layer. Experiments on four hyperspectral image datasets demonstrate that our token fusion algorithm can significantly improve inference speed without any training, while only causing a slight decrease in the pure transformer network’s classification accuracy.