Pith. sign in

REVIEW 4 major objections 4 minor 36 references

MVNet: Hyperspectral Remote Sensing Image Classification Based on Hybrid Mamba-Transformer Vision Backbone Architecture

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MVNet, a hybrid 3D-CNN/Transformer/Mamba architecture, claims near-perfect hyperspectral accuracy on Indian Pines, Pavia University, and KSC at lower cost than mainstream methods.

desk verdict Plausible hybrid Mamba-Transformer architecture undermined by inconsistent, likely leaked experimental results. read the letter →

arxiv 2507.04409 v1 pith:HU334B4H submitted 2025-07-06 cs.CV

Guandong Li , Mengxia Ye This is my paper
classification cs.CV
keywords hyperspectralimageclassificationMambastatespacemodelTransformer3D-CNNspatial-spectralfeatureextractiondecoupledattentionremotesensing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes MVNet, a hybrid architecture that combines 3D-CNN local feature extraction, Transformer global modeling, and Mamba linear-complexity sequence modeling for hyperspectral image classification. The core claim is that MVNet outperforms mainstream methods on Indian Pines, Pavia University, and Kennedy Space Center datasets in both accuracy and efficiency. The architecture's two key components are a dual-branch Mamba module that pairs a state space model branch with a non-SSM branch (1D convolution + SiLU), and an HSI-MambaVision Mixer that replaces causal convolution with regular convolution and uses decoupled attention to capture bidirectional spatial-spectral dependencies in one pass. If true, this gives a strong and efficient baseline for pixel-level HSI classification.

What carries the argument

The central machinery is the dual-branch Mamba module and the HSI-MambaVision Mixer. The dual-branch module pairs an SSM branch (selective scanning) with a non-SSM branch (1D convolution + SiLU), each projected to half the embedding dimension and then concatenated; the mixer replaces Mamba's causal convolution with regular convolution and applies decoupled attention to weight spatial and spectral features separately. Together they let the network capture bidirectional spatial-spectral dependencies in a single forward pass at linear complexity.

What would settle it

Retrain MVNet using a strict spatial split that places a buffer of at least the patch radius (6–8 pixels) between any training and test pixel, then re-measure overall accuracy on Indian Pines. If the OA falls substantially below the reported 99.74, or the per-class accuracies stop being exactly 1.0, the independence assumption behind the reported numbers is false.

Watch

Extended reading notes

Core claim

MVNet establishes that a redesigned Mamba block, adapted for non-sequential visual data, can serve as the core of a hyperspectral classifier when paired with a 3D-CNN backbone and Transformer-style attention. The paper argues that the autoregressive, causal-convolution design of the original Mamba block is ill-suited to HSI data, where spatial and spectral features need bidirectional modeling, and that replacing causal convolution with standard convolution while adding a symmetric non-SSM branch preserves linear complexity and recovers global context. The HSI-MambaVision Mixer then uses decoupled attention to weight spatial and spectral features separately, reducing parameter redundancy. On the three datasets, MVNet reports overall accuracies of 99.74 on Indian Pines, 99.98 on Pavia University, and 99.76 on KSC against the compared baselines.

Load-bearing premise

The results assume that test pixels are independent of training pixels even though the random split draws overlapping spatial patches from the same image; if neighboring training patches leak label information into test patches, the reported near-perfect accuracies do not reflect true generalization.

Editorial extensions

If this is right

  • If MVNet's reported results hold, HSI classification can exceed 99% overall accuracy on standard benchmarks with a single backbone, reducing the need for dataset-specific tuning.
  • The dual-branch Mamba design offers a template for adapting selective state space models to non-sequential, grid-structured data beyond hyperspectral images.
  • Replacing causal convolution with regular convolution plus decoupled attention could lower the computational cost of other Mamba-based vision architectures while capturing bidirectional context.
  • The claimed linear-complexity sequence modeling suggests MVNet could scale to larger scenes or higher spectral resolutions than quadratic-attention transformers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The decoupled attention idea could transfer to other high-dimensional, spectrally redundant datasets, such as multispectral imaging or medical hypercube analysis, where per-channel weighting matters.
  • The dual-branch Mamba module could serve as a reusable component for other non-sequential vision tasks, potentially improving Mamba-based backbones without quadratic attention cost.
  • The efficiency claim could be quantified further by reporting parameter counts and FLOPs against the cited baselines on the same three datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes MVNet, a hybrid 3D-CNN/Mamba/Transformer backbone for hyperspectral image classification. It introduces a dual-branch Mamba module (an SSM branch and a non-SSM 1D-convolution/SiLU branch) and an HSI-MambaVision Mixer that replaces causal convolution with non-causal convolution and uses decoupled attention to capture bidirectional spatial-spectral dependencies. The central claim, stated in the abstract, is that MVNet outperforms mainstream HSI classification methods on Indian Pines, Pavia University, and KSC in both accuracy and computational efficiency. The experimental section reports near-perfect metrics (e.g., OA/AA/Kappa of 99.74/99.67/99.70 on Indian Pines, 99.98/99.98/99.98 on Pavia, and 1.00/1.00/1.00 at block size 17 for both Pavia and KSC) and compares against SSRN, 3D-CNN, 3D-SE-DenseNet, DGCNet, Hit, and Spectralformer.

Significance. If the reported results were valid, MVNet would be a strong and computationally efficient baseline for HSI classification, combining local spectral-spatial features, global modeling, and linear-complexity sequence modeling in one architecture. The architectural ideas are motivated by task properties rather than fitted to benchmark numbers, which is a genuine strength: replacing causal convolution with non-causal convolution and adding a parallel non-SSM branch are sensible attempts to adapt Mamba to visual and hyperspectral data. However, the paper provides no code, no machine-checked proofs, and no reproducible protocol, and the empirical evidence is internally inconsistent and vulnerable to spatial leakage. The claimed advantage over existing methods is therefore not established, and the contribution remains an architectural proposal awaiting credible validation.

major comments (4)
  1. [§3.2.1-3.2.2, Tables 4-6] The evaluation protocol uses random pixel-level splits on overlapping spatial patches of size 13x13 or 17x17 drawn from the same image; no statement indicates that training, validation, and test pixels are spatially disjoint. Because patches overlap, every test patch shares pixels with many training patches, so the network can memorize local spatial context rather than learning transferable spectral-spatial features. This makes the near-perfect accuracies in Tables 5 and 6 (OA=AA=Kappa=1.00 for Pavia and KSC at block size 17) uninformative about generalization and invalidates the comparison with baselines in Tables 7 and 8. The paper must report results with spatially disjoint train/test regions or patch-level splits; without that, the abstract's claim of outperforming mainstream methods is unsupported. The internal contradiction that Section 3.2.2 says accuracy decreases at block size 17 while Tables 5 and 6 show perfect scores reinforces this concern.
  2. [Table 7] The reported MVNet AA of 99.67 for Indian Pines is not the average of the per-class accuracies listed in the same table: the mean of the 16 per-class values is approximately 99.04 (classes 1-9 and 13-15 listed as 1, i.e., 100%; classes 10-12 between 99.34 and 99.46; class 16 at 86.43). Since AA is by definition the mean per-class accuracy, either the per-class values or the aggregate metrics are incorrect. This internal inconsistency casts doubt on the reliability of the entire reported result set and must be resolved.
  3. [§3.2.1, §3.2.2, §3.3] The paper contradicts itself about the input configuration. Section 3.2.1 states that Pavia and KSC use 11x11 neighboring pixel blocks; Section 3.2.2 states that block size 13 is selected for Indian Pines and 17 for Pavia and KSC; Section 3.3 states that the input sizes are 13x13x200 for Indian Pines, 13x13x103 for Pavia, and 17x17x176 for KSC. The actual configuration used to produce Tables 7 and 8 is therefore unclear, making the experiments unreproducible as written.
  4. [§3.2.1, Tables 1-6] The training/validation/test ratio and the block size are selected by comparing OA/AA/Kappa values computed on the test set (e.g., the 6:1:3 and 5:1:4 ratios and block sizes 13/17 are chosen because they give the highest reported accuracies). This is test-set tuning; it inflates MVNet's scores and undermines the claim that MVNet outperforms the baselines in Tables 7 and 8. Moreover, no error bars or significance tests are reported for MVNet, whereas several baselines include standard deviations; fair comparison requires the same evaluation protocol for all methods.
minor comments (4)
  1. [Table 7 and text near it] The table and text refer to 'Hit [ ?]' with an unresolved citation; the reference entry is missing from the bibliography.
  2. [§2.3.1] The text mentions a channel attention module and a sparse convolution strategy, but gives no equations, implementation details, or ablation study, so these components cannot be reproduced or evaluated independently.
  3. [§2.2.1] The sentence about global convolution says the output of 'Eq. 4' is computed, but the relevant discrete output equation is Eq. (7); the cross-reference is wrong.
  4. [Tables 4-6] Tables 4-6 report '-' for Indian Pines block sizes 15 and 17, although Section 3.2.2 discusses the behavior at block size 17; either complete the table with the missing values or state explicitly why they are omitted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the architecture is not derived from benchmark outputs, and self-citations are contextual rather than load-bearing.

full rationale

The paper contains no derivation chain in which an output is shown to be equivalent to its inputs by construction. The proposed MVNet architecture is presented as a novel combination of 3D-CNN, Transformer, and Mamba components; the equations in Section 2.2.2 define the forward pass of the proposed mixers, and the experimental claims in Section 3 are empirical benchmark comparisons against external baselines. There is no fitted parameter that is later renamed as a prediction: the block-size selection in Section 3.2.2 is ordinary hyperparameter choice, and the reported OA/AA/Kappa values are evaluation results, not quantities derived from the training fit. The self-citations in the related-work discussion (references [12] through [19], all by the same authors) are contextual descriptions of prior architecture designs; none is invoked as a uniqueness theorem, a forbidden-alternative argument, or the sole justification for the central claim. The claim that MVNet outperforms mainstream methods is falsifiable by independent reproduction and does not reduce to any cited result. Internal inconsistencies in the experimental tables (e.g., the text in Section 3.2.2 says accuracy decreases at block size 17 while Tables 5 and 6 show perfect scores) and the potential spatial-patch leakage concern are correctness or soundness issues, not circularity per the specified definitions. No circular step can be quoted and exhibited, so the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The central claim relies on per-dataset hyperparameters (patch size, data split) that were chosen based on test performance, plus unstated assumptions about the evaluation protocol. No code was released, and the reported numbers are internally inconsistent, so the effective contribution beyond prior work is reduced.

free parameters (3)
  • Patch size (spatial block size M=N) per dataset = 13 (Indian Pines), 17 (Pavia University, KSC)
    Selected by comparing test OA across block sizes 7-17 in Tables 4-6; affects the input dimensions and the reported accuracy.
  • Train/validation/test ratio per dataset = 6:1:3 (Indian Pines, Pavia University), 5:1:4 (KSC)
    Selected by comparing test accuracy across ratios in Tables 1-3; the best ratio is reported as the method's setting.
  • Stage depths, window sizes, heads, mlp ratio, drop rate, conv3d dimension = 1,3,8,16; 4,4,7,7; 2,4,8,16; 4; 0.2; 80
    Hand-chosen architecture hyperparameters (Section 3.2); no ablation or sensitivity analysis is given.
assumptions (3)
  • standard math State space model equations (1)-(9) from Mamba/S4 are valid and can be discretized as stated.
    Background taken from Gu & Dao and Gu et al. (refs [5,6]); no proof given in this paper.
  • domain assumption Hyperspectral data exhibit high spectral redundancy and sparse ground objects, motivating the architecture.
    Section 1 and 2.2 state this as the problem motivation; it is not verified within the paper.
  • domain assumption Random pixel-level splitting with overlapping patches yields test samples independent of training samples.
    Section 3.2 assumes the standard HSI evaluation protocol; with large 17x17 patches this assumption is questionable.
invented entities (2)
  • HSI-MambaVision Mixer module
    purpose: Captures bidirectional spatial-spectral dependencies in a single forward pass and focuses on high-value features via decoupled attention.
    Only evaluated within this paper; the experiments are inconsistent and no external validation exists.
  • Dual-branch Mamba module (SSM + non-SSM branches)
    purpose: Models short- and long-range dependencies while reducing latency compared to bidirectional scanning.
    Introduced here; claimed efficiency is not measured, and the design closely follows MambaVision without citation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MVNet: Hyperspectral Remote Sensing Image Classification Based on Hybrid Mamba-Transformer Vision Backbone Architecture." pith.science (2026). https://pith.science/paper/HU334B4H

@misc{pith2026250704409,
  author       = {Pith},
  title        = {Pith review of: MVNet: Hyperspectral Remote Sensing Image Classification Based on Hybrid Mamba-Transformer Vision Backbone Architecture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HU334B4H}},
  note         = {Machine review of arXiv:2507.04409}
}
read the original abstract

Hyperspectral image (HSI) classification faces challenges such as high-dimensional data, limited training samples, and spectral redundancy, which often lead to overfitting and insufficient generalization capability. This paper proposes a novel MVNet network architecture that integrates 3D-CNN's local feature extraction, Transformer's global modeling, and Mamba's linear complexity sequence modeling capabilities, achieving efficient spatial-spectral feature extraction and fusion. MVNet features a redesigned dual-branch Mamba module, including a State Space Model (SSM) branch and a non-SSM branch employing 1D convolution with SiLU activation, enhancing modeling of both short-range and long-range dependencies while reducing computational latency in traditional Mamba. The optimized HSI-MambaVision Mixer module overcomes the unidirectional limitation of causal convolution, capturing bidirectional spatial-spectral dependencies in a single forward pass through decoupled attention that focuses on high-value features, alleviating parameter redundancy and the curse of dimensionality. On IN, UP, and KSC datasets, MVNet outperforms mainstream hyperspectral image classification methods in both classification accuracy and computational efficiency, demonstrating robust capability in processing complex HSI data.

Figures

Figures reproduced from arXiv: 2507.04409 by the authors.

Figure 1
Figure 1. Architecture of the MambaVision module [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Detailed design of HSI-MambaVision layer and overall hierarchical structure [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Overall network architecture diagram 3. Experiments and Analysis To evaluate the performance of MVNet, we conducted experiments on three repre￾sentative hyperspectral datasets: Indian Pines, Pavia University, and Kennedy Space Center (KSC). The classification metrics include Overall Accuracy (OA), Average Ac￾curacy (AA), and Kappa coefficient. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: : False color composite and ground truth labels of Indian Pines dataset [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: : False color composite and ground truth labels of Pavia University dataset [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: : Classification results of the best models for the Indian Pines dataset: (a) [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: : Training and validation curves of MVNet on Indian Pines showing loss and [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: : Training and validation curves of MVNet on Pavia University showing loss [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 33 canonical work pages

  1. [1]

    Intelligent remote sensing satellite system

    Zhang Bing. Intelligent remote sensing satellite system. Journal of Remote Sensing , 15(3):415–431, 2011

    work page 2011

  2. [2]

    Hyperspectral imaging: techniques for spectral detection and classification, volume 1

    Chein-I Chang. Hyperspectral imaging: techniques for spectral detection and classification, volume 1. Springer Science & Business Media, 2003

    work page 2003

  3. [3]

    Deep fea- ture extraction and classification of hyperspectral images based on convolutional neural networks

    Yushi Chen, Hanlu Jiang, Chunyang Li, Xiuping Jia, and Pedram Ghamisi. Deep fea- ture extraction and classification of hyperspectral images based on convolutional neural networks. IEEE transactions on geoscience and remote sensing , 54(10):6232–6251, 2016

    work page 2016

  4. [4]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks , 107:3–11, 2018

    work page 2018

  5. [5]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 , 2023

    arXiv 2023

  6. [6]

    Combining recurrent, convolutional, and continuous-time models with linear state space layers

    Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher R´ e. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems , 34:572–585, 2021

    work page 2021

  7. [7]

    CMTNet: Convolutional Meets Transformer Network for Hyperspectral Images Classification

    Faxu Guo, Quan Feng, Sen Yang, and Wanxia Yang. Cmtnet: Convolutional meets trans- former network for hyperspectral images classification. arXiv preprint arXiv:2406.14080 , 2024

    work page Pith review arXiv 2024

  8. [8]

    Spatial-spectral transformer for hyperspectral image classification

    Xin He, Yushi Chen, and Zhouhan Lin. Spatial-spectral transformer for hyperspectral image classification. Remote Sensing, 13(3):498, 2021

    work page 2021

Show all 36 references
  1. [9]

    3dss-mamba: 3d-spectral-spatial mamba for hyperspectral image classification

    Yan He, Bing Tu, Bo Liu, Jun Li, and Antonio Plaza. 3dss-mamba: 3d-spectral-spatial mamba for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  2. [10]

    Spectralformer: Rethinking hyperspectral image classification with transform- ers

    Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza, and Jocelyn Chanussot. Spectralformer: Rethinking hyperspectral image classification with transform- ers. IEEE Transactions on Geoscience and Remote Sensing , 60:1–15, 2021

    2021

  3. [11]

    Going deeper with contextual cnn for hyperspectral image classification

    Hyungtae Lee and Heesung Kwon. Going deeper with contextual cnn for hyperspectral image classification. IEEE Transactions on Image Processing , 26(10):4843–4855, 2017

    2017

  4. [12]

    3d wavelet convolutions with extended receptive fields for hyperspectral image classification

    Guandong Li and Mengxia Ye. 3d wavelet convolutions with extended receptive fields for hyperspectral image classification. arXiv preprint arXiv:2504.10795 , 2025

    2025 arXiv

  5. [13]

    Dgcnet: An efficient 3d-densenet based on dynamic group convolution for hyperspectral remote sensing image classification

    Guandong Li and Mengxia Ye. Dgcnet: An efficient 3d-densenet based on dynamic group convolution for hyperspectral remote sensing image classification. Spectroscopy Letters, pages 1–14, 2025

    2025

  6. [14]

    Dynamic 3d kan convolution with adaptive grid optimiza- tion for hyperspectral image classification

    Guandong Li and Mengxia Ye. Dynamic 3d kan convolution with adaptive grid optimiza- tion for hyperspectral image classification. arXiv preprint arXiv:2504.15155 , 2025

    2025 arXiv

  7. [15]

    Efficient dynamic attention 3d convolution for hyperspec- tral image classification

    Guandong Li and Mengxia Ye. Efficient dynamic attention 3d convolution for hyperspec- tral image classification. arXiv preprint arXiv:2503.23472 , 2025

    2025 arXiv

  8. [16]

    Expert kernel generation network driven by contextual mapping for hyperspectral image classification

    Guandong Li and Mengxia Ye. Expert kernel generation network driven by contextual mapping for hyperspectral image classification. arXiv preprint arXiv:2504.13045 , 2025

    2025 arXiv

  9. [17]

    Hyperspectral image classification via transformer- based spectral-spatial attention decoupling and adaptive gating

    Guandong Li and Mengxia Ye. Hyperspectral image classification via transformer- based spectral-spatial attention decoupling and adaptive gating. arXiv preprint arXiv:2506.08324, 2025

    2025 arXiv

  10. [18]

    Spatial-geometry enhanced 3d dynamic snake convolu- tional neural network for hyperspectral image classification, 2025

    Guandong Li and Mengxia Ye. Spatial-geometry enhanced 3d dynamic snake convolu- tional neural network for hyperspectral image classification, 2025

    2025

  11. [19]

    Spatial-spectral hyperspectral classification based on learnable 3d group convolution

    Guandong Li and Mengxia Ye. Spatial-spectral hyperspectral classification based on learnable 3d group convolution. Spectroscopy Letters, pages 1–13, 2025

    2025

  12. [20]

    Faster hyperspectral image classification based on selective kernel mechanism using deep convolutional networks

    Guandong Li and Chunju Zhang. Faster hyperspectral image classification based on selective kernel mechanism using deep convolutional networks. arXiv preprint arXiv:2202.06458, 2022

    2022 arXiv

  13. [21]

    Doubleconvpool-structured 3d-cnn for hyper- spectral remote sensing image classification

    GD Li, CJ Zhang, F Gao, and XY Zhang. Doubleconvpool-structured 3d-cnn for hyper- spectral remote sensing image classification. Journal of Image and Graphics , 24(4):639– 654, 2019

    2019

  14. [22]

    Hyperspectral remote sensing image classification using three-dimensional-squeeze-and- excitation-densenet (3d-se-densenet)

    Guandong Li, Chunju Zhang, Runmin Lei, Xueying Zhang, Zhourun Ye, and Xiaoli Li. Hyperspectral remote sensing image classification using three-dimensional-squeeze-and- excitation-densenet (3d-se-densenet). Remote Sensing Letters , 11(2):195–203, 2020

    2020

  15. [23]

    Mambahsi: Spatial- spectral mamba for hyperspectral image classification

    Yapeng Li, Yong Luo, Lefei Zhang, Zengmao Wang, and Bo Du. Mambahsi: Spatial- spectral mamba for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  16. [24]

    Swin transformer v2: Scaling up capacity and resolution

    Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, 20 Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 12009–12019, 2022

    2022

  17. [25]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012– 10022, 2021

    2021

  18. [26]

    Deep supervised learning for hyperspectral data classification through con- volutional neural networks

    Konstantinos Makantasis, Konstantinos Karantzalos, Anastasios Doulamis, and Nikolaos Doulamis. Deep supervised learning for hyperspectral data classification through con- volutional neural networks. In 2015 IEEE international geoscience and remote sensing symposium (IGARSS), p...

    2015

  19. [27]

    Spectral–spatial feature tokeniza- tion transformer for hyperspectral image classification

    Le Sun, Guangrui Zhao, Yuhui Zheng, and Zebin Wu. Spectral–spatial feature tokeniza- tion transformer for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing , 60:1–14, 2022

    2022

  20. [28]

    A fast dense spectral– spatial convolution network framework for hyperspectral images classification

    Wenju Wang, Shuguang Dou, Zhongmin Jiang, and Liujie Sun. A fast dense spectral– spatial convolution network framework for hyperspectral images classification. Remote sensing, 10(7):1068, 2018

    2018

  21. [29]

    Hyperspectral image transformer classification networks

    Xiaofei Yang, Weijia Cao, Yao Lu, and Yicong Zhou. Hyperspectral image transformer classification networks. IEEE Transactions on Geoscience and Remote Sensing , 60:1–15, 2022

    2022

  22. [30]

    Spectralmamba: Efficient mamba for hyperspectral image classification

    Jing Yao, Danfeng Hong, Chenyu Li, and Jocelyn Chanussot. Spectralmamba: Efficient mamba for hyperspectral image classification. arXiv preprint arXiv:2404.08489 , 2024

    2024 arXiv

  23. [31]

    Multi-scale dense networks for hyperspec- tral remote sensing image classification

    Chunju Zhang, Guandong Li, and Shihong Du. Multi-scale dense networks for hyperspec- tral remote sensing image classification. IEEE Transactions on Geoscience and Remote Sensing, 57(11):9201–9222, 2019

    2019

  24. [32]

    Three-dimensional densely connected convolutional network for hyperspectral remote sensing image classifi- cation

    Chunju Zhang, Guandong Li, Shihong Du, Wuzhou Tan, and Fei Gao. Three-dimensional densely connected convolutional network for hyperspectral remote sensing image classifi- cation. Journal of Applied Remote Sensing , 13(1):016519–016519, 2019

    2019

  25. [33]

    Deep feature aggregation network for hyperspectral remote sensing image classification

    Chunju Zhang, Guandong Li, Runmin Lei, Shihong Du, Xueying Zhang, Hui Zheng, and Zhaofu Wu. Deep feature aggregation network for hyperspectral remote sensing image classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13:5314–5325, 2020

    2020

  26. [34]

    A survey on visual mamba

    Hanwei Zhang, Ying Zhu, Dan Wang, Lijun Zhang, Tianxiang Chen, Ziyang Wang, and Zi Ye. A survey on visual mamba. Applied Sciences, 14(13):5683, 2024

    2024

  27. [35]

    Spectral–spatial feature extraction for hyperspectral image classification: A dimension reduction and deep learning approach

    Wenzhi Zhao and Shihong Du. Spectral–spatial feature extraction for hyperspectral image classification: A dimension reduction and deep learning approach. IEEE Transactions on Geoscience and Remote Sensing , 54(8):4544–4554, 2016

    2016

  28. [36]

    Spectral–spatial residual network for hyperspectral image classification: A 3-d deep learning framework

    Zilong Zhong, Jonathan Li, Zhiming Luo, and Michael Chapman. Spectral–spatial residual network for hyperspectral image classification: A 3-d deep learning framework. IEEE Transactions on Geoscience and Remote Sensing , 56(2):847–858, 2017. 21

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.