pith. machine review for the scientific record. sign in

arxiv: 2604.23622 · v1 · submitted 2026-04-26 · 💻 cs.CV

Recognition: unknown

A Synergistic CNN-Transformer Network with Pooling Attention Fusion for Hyperspectral Image Classification

Feng Qian, Guangyao Shi, Jingwen Yan, Peng Chen, Wenxuan He

Authors on Pith no claims yet

Pith reviewed 2026-05-08 06:31 UTC · model grok-4.3

classification 💻 cs.CV
keywords hyperspectral image classificationCNN-Transformer hybridspatial-spectral fusionpooling attentionfeature extraction moduleremote sensingland cover mapping
0
0 comments X

The pith

A new network uses parallel CNN and transformer branches to extract and fuse spatial-spectral features from hyperspectral images, yielding higher classification accuracy than prior methods.

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

The paper presents a synergistic CNN-Transformer architecture that processes spatial and spectral information through separate branches while adding dedicated modules to combine them and limit information loss across layers. It claims this design overcomes two recurring difficulties in hyperspectral classification: ineffective joint use of local and non-local features, and progressive degradation of detail during network propagation. Experiments on standard benchmark datasets show the resulting model outperforms existing state-of-the-art approaches. A sympathetic reader would care because hyperspectral images supply rich material signatures for applications such as land-cover mapping, yet current networks still struggle to exploit both the spatial layout and the many spectral bands without losing critical signals.

Core claim

The central claim is that a Twin-Branch Feature Extraction module running 3D and 2D convolutions in parallel, a hybrid pooling attention module for spatial aggregation, a cascade transformer encoder for global spectral context, and a cross-layer feature fusion module together allow CNNs and vision transformers to collaborate on spatial-spectral data, producing superior pixel-level classification results on representative hyperspectral datasets.

What carries the argument

The Twin-Branch Feature Extraction (TBFE) module that applies 3D and 2D convolutions in parallel to capture spectral and spatial features separately, supported by hybrid pooling attention (HPA) for spatial weighting and cross-layer feature fusion (CFF) to retain information from earlier layers.

If this is right

  • Pixel classification into land-cover categories improves on multiple public HSI benchmarks.
  • Spatial and spectral features can be handled separately before fusion without excessive loss of detail.
  • Global spectral dependencies captured by the cascade transformer contribute to the observed accuracy gains.
  • Cross-layer fusion preserves information that would otherwise degrade in deeper networks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same branch-separation plus fusion pattern could be tested on other multi-band remote-sensing modalities such as multispectral or SAR data.
  • Computational cost comparisons with pure CNN or pure transformer baselines would clarify whether the added modules remain practical for large-scale mapping.
  • Ablation results on the individual modules could be examined across datasets to identify which component drives most of the gain.

Load-bearing premise

The newly added TBFE, HPA, and CFF modules together solve spatial-spectral fusion and layer-wise information loss without adding dataset-specific biases or requiring hyperparameter choices that were not disclosed in the experiments.

What would settle it

Apply the full model and each of its ablated variants to a fresh hyperspectral dataset never used in training or tuning; if accuracy gains disappear or if removing any single module leaves performance unchanged, the claimed benefit of the synergistic design is refuted.

Figures

Figures reproduced from arXiv: 2604.23622 by Feng Qian, Guangyao Shi, Jingwen Yan, Peng Chen, Wenxuan He.

Figure 1
Figure 1. Figure 1: An overview illustration of the proposed synergistic CNN-Transformer network for the HSI classification task. The proposed network consists of several view at source ↗
Figure 2
Figure 2. Figure 2: An illustration of the hybrid pooling attention module (HPA). The “g” view at source ↗
Figure 3
Figure 3. Figure 3: An illustration of the cross-layer feature fusion module (CFF). view at source ↗
Figure 4
Figure 4. Figure 4: The OA impact curves of different parameters. (a) Remained bands. (b) Patch size. (c) Learning rate. (d) Attention Head view at source ↗
Figure 5
Figure 5. Figure 5: Classification maps of the Salinas dataset. (a) Ground-truth map, (b) SVM, (c) 2-D-CNN, (d) 3-D-CNN, (e) Hybrid, (f)SpectralFormer, (g)SSFTT, (h)MASSFormer, (i)SS-Mamba, (j)Ours. 12 view at source ↗
Figure 6
Figure 6. Figure 6: Classification maps of the Pavia Universit dataset. (a) Ground-truth map, (b) SVM, (c) 2-D-CNN, (d) 3-D-CNN, (e) Hybrid, (f)SpectralFormer, (g)SSFTT, (h)MASSFormer, (i)SS-Mamba, (j)Ours. (a) (j) (g) (f) (b) (d) (e) (c) (h) (i) view at source ↗
Figure 7
Figure 7. Figure 7: Classification maps of the Houston2013 dataset. (a) Ground-truth map, (b) SVM, (c) 2-D-CNN, (d) 3-D-CNN, (e) Hybrid, (f)SpectralFormer, (g)SSFTT, (h)MASSFormer, (i)SS-Mamba, (j)Ours. 13 view at source ↗
Figure 8
Figure 8. Figure 8: Classification maps of the WHU-Hi-HanChuan dataset. (a) Ground-truth map, (b) SVM, (c) 2-D-CNN, (d) 3-D-CNN, (e) Hybrid, (f)SpectralFormer, (g)SSFTT, (h)MASSFormer, (i)SS-Mamba, (j)Ours view at source ↗
Figure 9
Figure 9. Figure 9: Classification maps of the Houston2018 dataset. (a) Ground-truth map, (b) SVM, (c) 2-D-CNN, (d) 3-D-CNN, (e) Hybrid, (f)SpectralFormer, (g)SSFTT, (h)MASSFormer, (i)SS-Mamba, (j)Ours. In the future, leveraging multimodal knowledge may im￾prove the proposed architecture beyond solely vision modality. Future research will explore enhancing the model’s ability to encode multimodal features to achieve a richer … view at source ↗
read the original abstract

In the hyperspectral image (HSI) classification task, each pixel is categorized into a specific land-cover category or material. Convolutional neural networks (CNNs) and transformers have been widely used to extract local and non-local features in HSI classification. Recent works have utilized a multi-scale vision transformer (ViT) to enhance spectral feature capture and yield promising results. However, most existing methods still face challenges in the effective joint use of spatial-spectral information and in preserving information across layers during the propagation process. To address these issues, we propose a synergistic CNN-Transformer network with pooling attention fusion for HSI classification, which collaboratively utilizes CNNs and ViT to process spatial and spectral features separately. Specifically, we propose a Twin-Branch Feature Extraction (TBFE) module, which employs 3D and 2D convolution in parallel to comprehensively extract spectral and spatial features from HSI. A hybrid pooling attention (HPA) module is designed to aggregate spatial attention. Moreover, a cascade transformer encoder is employed for global spectral feature extraction, and a simple yet efficient cross-layer feature fusion (CFF) module is designed to reduce the loss of crucial information in the previous network layers. Extensive experiments are conducted on several representative datasets to demonstrate the superior performance of our proposed method compared to the state-of-the-art works. Code is available at https://github.com/chenpeng052/SCT-Net.git.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a synergistic CNN-Transformer network for hyperspectral image classification. It introduces the Twin-Branch Feature Extraction (TBFE) module with parallel 3D and 2D convolutions to extract spectral and spatial features, the Hybrid Pooling Attention (HPA) module to aggregate spatial attention, a cascade transformer encoder for global spectral feature extraction, and the Cross-Layer Feature Fusion (CFF) module to reduce information loss across layers. The central claim is that this architecture collaboratively utilizes CNNs and ViT to achieve superior performance over state-of-the-art methods on representative HSI datasets.

Significance. If the results hold under controlled conditions, this work could advance hybrid CNN-Transformer models in hyperspectral imaging by addressing spatial-spectral fusion and information preservation. The open-source code link is a strength for reproducibility. However, the significance is limited because the empirical gains are not yet shown to be attributable to the proposed modules rather than training variations.

major comments (2)
  1. [Experimental results] Experimental results section: The manuscript reports superior accuracy on standard HSI benchmarks but provides no ablation studies isolating the contributions of TBFE, HPA, and CFF. This is load-bearing for the central claim, as the strongest assertion attributes the margins to the synergistic design and these modules; without component ablations or controlled re-runs of baselines using identical optimizer, patch size, and augmentation, attribution cannot be verified.
  2. [Method] Method section on CFF: The claim that the cross-layer feature fusion module reduces loss of crucial information across layers is supported only by end-to-end accuracy; no quantitative metrics (e.g., layer-wise feature similarity or information retention scores) are given to substantiate the reduction in layer-wise loss.
minor comments (1)
  1. [Abstract] The abstract states that CNNs and ViT process spatial and spectral features separately, yet the TBFE module uses parallel 3D/2D convolutions on the same input; a short clarification on the separation mechanism would aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that additional experiments are needed to strengthen the attribution of performance gains and to provide quantitative support for the CFF module. We will revise the paper accordingly.

read point-by-point responses

  1. Referee: [Experimental results] Experimental results section: The manuscript reports superior accuracy on standard HSI benchmarks but provides no ablation studies isolating the contributions of TBFE, HPA, and CFF. This is load-bearing for the central claim, as the strongest assertion attributes the margins to the synergistic design and these modules; without component ablations or controlled re-runs of baselines using identical optimizer, patch size, and augmentation, attribution cannot be verified.

    Authors: We agree that ablation studies are essential to isolate module contributions and ensure fair attribution. In the revised manuscript, we will add comprehensive ablation experiments removing or replacing TBFE, HPA, and CFF individually. We will also re-implement all baseline methods under identical conditions (same optimizer, patch size, augmentation, and training protocol) to enable direct comparison and verify that the reported margins stem from the proposed synergistic design rather than implementation differences. revision: yes

  2. Referee: [Method] Method section on CFF: The claim that the cross-layer feature fusion module reduces loss of crucial information across layers is supported only by end-to-end accuracy; no quantitative metrics (e.g., layer-wise feature similarity or information retention scores) are given to substantiate the reduction in layer-wise loss.

    Authors: We acknowledge that end-to-end accuracy alone is insufficient to directly demonstrate information preservation by CFF. In the revision, we will include quantitative analyses such as layer-wise cosine similarity between features before and after fusion, as well as information retention metrics (e.g., mutual information or reconstruction error across layers), to provide explicit evidence supporting the claim that CFF reduces crucial information loss. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical architecture proposal

full rationale

The manuscript proposes a new CNN-Transformer network (TBFE, HPA, cascade transformer encoder, CFF) for HSI classification and validates it via end-to-end accuracy on standard benchmarks. No mathematical derivations, equations, or first-principles results exist that could reduce to inputs by construction. Claims rest entirely on experimental comparisons rather than self-definitional fits, fitted inputs renamed as predictions, or load-bearing self-citations. The work is self-contained as an empirical ML architecture paper.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 3 invented entities

The central claim rests on standard deep-learning assumptions plus several newly introduced architectural modules whose effectiveness is shown only empirically. No new physical entities are postulated.

free parameters (2)
  • Number of transformer layers and attention heads
    Chosen during architecture design; typical hyperparameter that must be tuned for the reported performance.
  • Pooling sizes and fusion weights in HPA and CFF
    Design choices that directly affect feature aggregation and are not derived from first principles.
axioms (2)
  • domain assumption 3D and 2D convolutions can separately capture spectral and spatial features in HSI data
    Invoked when describing the TBFE module; standard premise in CNN-based HSI papers.
  • domain assumption Transformer encoders can extract global spectral dependencies
    Basis for the cascade transformer encoder component.
invented entities (3)
  • Twin-Branch Feature Extraction (TBFE) module no independent evidence
    purpose: Parallel extraction of spectral and spatial features via 3D and 2D convolutions
    Newly proposed module; no independent evidence outside the paper.
  • Hybrid Pooling Attention (HPA) module no independent evidence
    purpose: Aggregation of spatial attention via hybrid pooling
    Newly proposed module; no independent evidence outside the paper.
  • Cross-Layer Feature Fusion (CFF) module no independent evidence
    purpose: Reduction of information loss across network layers
    Newly proposed module; no independent evidence outside the paper.

pith-pipeline@v0.9.0 · 5568 in / 1686 out tokens · 40502 ms · 2026-05-08T06:31:24.333928+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages · 2 internal anchors

  1. [1]

    Y . Fu, T. Zhang, Y . Zheng, D. Zhang, H. Huang, Joint camera spectral response selection and hyperspec- tral image recovery, IEEE Transactions on Pattern Anal- ysis and Machine Intelligence 44 (1) (2022) 256–272. doi:10.1109/TPAMI.2020.3009999

    work page doi:10.1109/tpami.2020.3009999 2022

  2. [2]

    Liang, S

    Z. Liang, S. Wang, T. Zhang, Y . Fu, Blind super- resolution of single remotely sensed hyperspectral image, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–14. doi:10.1109/TGRS.2023.3302128

    work page doi:10.1109/tgrs.2023.3302128 2023

  3. [3]

    Mohamed, M

    S. Mohamed, M. Haghighat, T. Fernando, S. Sridharan, C. Fookes, P. Moghadam, Factoformer: Factorized hy- perspectral transformers with self-supervised pretraining, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–14. doi:10.1109/TGRS.2023.3343392

    work page doi:10.1109/tgrs.2023.3343392 2024

  4. [4]

    J. Deng, R. Wang, L. Yang, X. Lv, Z. Yang, K. Zhang, C. Zhou, L. Pengju, Z. Wang, A. Abdullah, M. Zhan- hong, Quantitative estimation of wheat stripe rust disease index using unmanned aerial vehicle hyperspectral im- agery and innovative vegetation indices, IEEE Transac- tions on Geoscience and Remote Sensing 61 (2023) 1–11. doi:10.1109/TGRS.2023.3292130

    work page doi:10.1109/tgrs.2023.3292130 2023

  5. [5]

    J. Wang, S. Guo, R. Huang, L. Li, X. Zhang, L. Jiao, Dual-channel capsule generation adversarial network for hyperspectral image classification, IEEE Transactions on Geoscience and Remote Sensing 60 (2022) 1–16. doi:10.1109/TGRS.2020.3044312

    work page doi:10.1109/tgrs.2020.3044312 2022

  6. [6]

    P. Chen, W. He, F. Qian, G. Shi, J. Yan, A synergistic cnn- transformer network with pooling attention fusion for hy- perspectral image classification, Digital Signal Processing 160 (2025) 105070

    2025

  7. [7]

    S. Li, W. Song, L. Fang, Y . Chen, P. Ghamisi, J. A. Benediktsson, Deep learning for hyperspectral image classification: An overview, IEEE Transactions on Geo- science and Remote Sensing 57 (9) (2019) 6690–6709. doi:10.1109/TGRS.2019.2907932

    work page doi:10.1109/tgrs.2019.2907932 2019

  8. [8]

    Melgani, L

    F. Melgani, L. Bruzzone, Classification of hyper- spectral remote sensing images with support vec- tor machines, IEEE Transactions on Geoscience and Remote Sensing 42 (8) (2004) 1778–1790. doi:10.1109/TGRS.2004.831865. 15

    work page doi:10.1109/tgrs.2004.831865 2004

  9. [9]

    L. Ma, M. M. Crawford, J. Tian, Local mani- fold learning-basedk-nearest-neighbor for hyperspec- tral image classification, IEEE Transactions on Geo- science and Remote Sensing 48 (11) (2010) 4099–4109. doi:10.1109/TGRS.2010.2055876

    work page doi:10.1109/tgrs.2010.2055876 2010

  10. [10]

    J. Ham, Y . Chen, M. Crawford, J. Ghosh, Investi- gation of the random forest framework for classifica- tion of hyperspectral data, IEEE Transactions on Geo- science and Remote Sensing 43 (3) (2005) 492–501. doi:10.1109/TGRS.2004.842481

    work page doi:10.1109/tgrs.2004.842481 2005

  11. [11]

    Fauvel, J

    M. Fauvel, J. A. Benediktsson, J. Chanussot, J. R. Sveins- son, Spectral and spatial classification of hyperspectral data using svms and morphological profiles, IEEE Trans- actions on Geoscience and Remote Sensing 46 (11) (2008) 3804–3814. doi:10.1109/TGRS.2008.922034

    work page doi:10.1109/tgrs.2008.922034 2008

  12. [12]

    Benediktsson, J

    J. Benediktsson, J. Palmason, J. Sveinsson, Classifica- tion of hyperspectral data from urban areas based on extended morphological profiles, IEEE Transactions on Geoscience and Remote Sensing 43 (3) (2005) 480–491. doi:10.1109/TGRS.2004.842478

    work page doi:10.1109/tgrs.2004.842478 2005

  13. [13]

    Dalla Mura, A

    M. Dalla Mura, A. Villa, J. A. Benediktsson, J. Chanus- sot, L. Bruzzone, Classification of hyperspectral im- ages by using extended morphological attribute pro- files and independent component analysis, IEEE Geo- science and Remote Sensing Letters 8 (3) (2011) 542–

    2011

  14. [14]

    doi:10.1109/LGRS.2010.2091253

    work page doi:10.1109/lgrs.2010.2091253 2010

  15. [15]

    M. Wang, Y . Sun, J. Xiang, Y . Zhong, Citnet: Con- volution interaction transformer network for hyperspec- tral and lidar image classification, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–18. doi:10.1109/TGRS.2024.3477965

    work page doi:10.1109/tgrs.2024.3477965 2024

  16. [16]

    L. Cao, K. Chua, W. Chong, H. Lee, Q. Gu, A comparison of pca, kpca and ica for dimensional- ity reduction in support vector machine, Neurocomput- ing 55 (1) (2003) 321–336, support Vector Machines. doi:https://doi.org/10.1016/S0925-2312(03)00433-8

    work page doi:10.1016/s0925-2312(03)00433-8 2003

  17. [17]

    H. Yuan, Y . Lu, L. Yang, H. Luo, Y . Y . Tang, Spectral- spatial linear discriminant analysis for hyperspectral im- age classification, in: 2013 IEEE International Con- ference on Cybernetics (CYBCO), 2013, pp. 144–149. doi:10.1109/CYBConf.2013.6617430

    work page doi:10.1109/cybconf.2013.6617430 2013

  18. [18]

    Z. Li, Z. Xue, Q. Xu, L. Zhang, T. Zhu, M. Zhang, Spformer: Self-pooling transformer for few-shot hy- perspectral image classification, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–19. doi:10.1109/TGRS.2023.3345923

    work page doi:10.1109/tgrs.2023.3345923 2024

  19. [19]

    P. Chen, C. Huang, Wmoe-clip: Wavelet-enhanced mixture-of-experts prompt learning for zero-shot anomaly detection, in: ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2026, pp. 22257–22261

    2026

  20. [20]

    Farrell, R

    M. Farrell, R. Mersereau, On the impact of pca dimen- sion reduction for hyperspectral detection of difficult tar- gets, IEEE Geoscience and Remote Sensing Letters 2 (2) (2005) 192–195. doi:10.1109/LGRS.2005.846011

    work page doi:10.1109/lgrs.2005.846011 2005

  21. [21]

    C. Yu, Y . Zhu, M. Song, Y . Wang, Q. Zhang, Unseen fea- ture extraction: Spatial mapping expansion with spectral compression network for hyperspectral image classifica- tion, IEEE Transactions on Geoscience and Remote Sens- ing 62 (2024) 1–15. doi:10.1109/TGRS.2024.3420137

    work page doi:10.1109/tgrs.2024.3420137 2024

  22. [22]

    P. Chen, F. Huang, C. Huang, Dyc-clip: Dynamic context- aware multi-modal prompt learning for zero-shot anomaly detection, Pattern Recognition (2026) 113215

    2026

  23. [23]

    Y . Chen, X. Zhao, X. Jia, Spectral–spatial classifica- tion of hyperspectral data based on deep belief network, IEEE Journal of Selected Topics in Applied Earth Ob- servations and Remote Sensing 8 (6) (2015) 2381–2392. doi:10.1109/JSTARS.2015.2388577

    work page doi:10.1109/jstars.2015.2388577 2015

  24. [24]

    W. Zhao, S. Du, Spectral–spatial feature extraction for hy- perspectral image classification: A dimension reduction and deep learning approach, IEEE Transactions on Geo- science and Remote Sensing 54 (8) (2016) 4544–4554. doi:10.1109/TGRS.2016.2543748

    work page doi:10.1109/tgrs.2016.2543748 2016

  25. [25]

    J. Yue, W. Zhao, S. Mao, H. Liu, Spectral– spatial classification of hyperspectral images us- ing deep convolutional neural networks, Re- mote Sensing Letters 6 (6) (2015) 468–477. doi:https://doi.org/10.1080/2150704X.2015.1047045

    work page doi:10.1080/2150704x.2015.1047045 2015

  26. [26]

    Chakraborty, U

    T. Chakraborty, U. Trehan, Spectralnet: Explor- ing spatial-spectral waveletcnn for hyperspectral image classification, arXiv preprint arXiv:2104.00341 (2021). doi:https://doi.org/10.48550/arXiv.2104.00341

    work page doi:10.48550/arxiv.2104.00341 2021

  27. [27]

    C. Shi, S. Yue, L. Wang, A dual-branch multiscale trans- former network for hyperspectral image classification, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–20. doi:10.1109/TGRS.2024.3351486

    work page doi:10.1109/tgrs.2024.3351486 2024

  28. [28]

    Y . Chen, H. Jiang, C. Li, X. Jia, P. Ghamisi, Deep fea- ture extraction and classification of hyperspectral images based on convolutional neural networks, IEEE Transac- tions on Geoscience and Remote Sensing 54 (10) (2016) 6232–6251. doi:10.1109/TGRS.2016.2584107

    work page doi:10.1109/tgrs.2016.2584107 2016

  29. [29]

    S. K. Roy, G. Krishna, S. R. Dubey, B. B. Chaud- huri, Hybridsn: Exploring 3-d–2-d cnn feature hierar- chy for hyperspectral image classification, IEEE Geo- science and Remote Sensing Letters 17 (2) (2020) 277–

    2020

  30. [30]

    doi:10.1109/LGRS.2019.2918719

    work page doi:10.1109/lgrs.2019.2918719 2019

  31. [31]

    J. Zhou, S. Zeng, G. Gao, Y . Chen, Y . Tang, A novel spatial–spectral pyramid network for hyper- spectral image classification, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–14. doi:10.1109/TGRS.2023.3303338. 16

    work page doi:10.1109/tgrs.2023.3303338 2023

  32. [32]

    K. Yang, H. Sun, C. Zou, X. Lu, Cross-attention spec- tral–spatial network for hyperspectral image classifica- tion, IEEE Transactions on Geoscience and Remote Sens- ing 60 (2022) 1–14. doi:10.1109/TGRS.2021.3133582

    work page doi:10.1109/tgrs.2021.3133582 2022

  33. [33]

    J. Wang, W. Li, M. Zhang, J. Chanussot, Large ker- nel sparse convnet weighted by multi-frequency attention for remote sensing scene understanding, IEEE Transac- tions on Geoscience and Remote Sensing 61 (2023) 1–12. doi:10.1109/TGRS.2023.3333401

    work page doi:10.1109/tgrs.2023.3333401 2023

  34. [34]

    Ullah, I

    F. Ullah, I. Ullah, R. U. Khan, S. Khan, K. Khan, G. Pau, Conventional to deep ensemble methods for hy- perspectral image classification: A comprehensive sur- vey, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 17 (2024) 3878–3916. doi:10.1109/JSTARS.2024.3353551

    work page doi:10.1109/jstars.2024.3353551 2024

  35. [35]

    L. Mou, P. Ghamisi, X. X. Zhu, Deep recurrent neural net- works for hyperspectral image classification, IEEE Trans- actions on Geoscience and Remote Sensing 55 (7) (2017) 3639–3655. doi:10.1109/TGRS.2016.2636241

    work page doi:10.1109/tgrs.2016.2636241 2017

  36. [36]

    L. Zhu, Y . Chen, P. Ghamisi, J. A. Benedikts- son, Generative adversarial networks for hyperspec- tral image classification, IEEE Transactions on Geo- science and Remote Sensing 56 (9) (2018) 5046–5063. doi:10.1109/TGRS.2018.2805286

    work page doi:10.1109/tgrs.2018.2805286 2018

  37. [37]

    Y . Ding, Z. Zhang, X. Zhao, D. Hong, W. Cai, C. Yu, N. Yang, W. Cai, Multi-feature fusion: Graph neu- ral network and cnn combining for hyperspectral im- age classification, Neurocomputing 501 (2022) 246–257. doi:https://doi.org/10.1016/j.neucom.2022.06.031

    work page doi:10.1016/j.neucom.2022.06.031 2022

  38. [38]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, arXiv preprint arXiv:2312.00752 (2023)

    work page internal anchor Pith review arXiv 2023

  39. [39]

    Huang, Y

    L. Huang, Y . Chen, X. He, Spectral-spatial mamba for hy- perspectral image classification, Remote Sensing 16 (13) (2024). doi:10.3390/rs16132449

    work page doi:10.3390/rs16132449 2024

  40. [40]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing systems 30 (2017)

    2017

  41. [41]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    work page internal anchor Pith review arXiv 2010

  42. [42]

    D. Hong, Z. Han, J. Yao, L. Gao, B. Zhang, A. Plaza, J. Chanussot, Spectralformer: Rethinking hyperspectral image classification with transformers, IEEE Transac- tions on Geoscience and Remote Sensing 60 (2022) 1–15. doi:10.1109/TGRS.2021.3130716

    work page doi:10.1109/tgrs.2021.3130716 2022

  43. [43]

    L. Sun, G. Zhao, Y . Zheng, Z. Wu, Spectral–spatial feature tokenization transformer for hyperspec- tral image classification, IEEE Transactions on Geoscience and Remote Sensing 60 (2022) 1–14. doi:10.1109/TGRS.2022.3144158

    work page doi:10.1109/tgrs.2022.3144158 2022

  44. [44]

    S. Mei, C. Song, M. Ma, F. Xu, Hyperspectral image classification using group-aware hierarchical transformer, IEEE Transactions on Geoscience and Remote Sensing 60 (2022) 1–14. doi:10.1109/TGRS.2022.3207933

    work page doi:10.1109/tgrs.2022.3207933 2022

  45. [45]

    Z. Shu, Y . Wang, Z. Yu, Dual attention transformer net- work for hyperspectral image classification, Engineering Applications of Artificial Intelligence 127 (2024) 107351. doi:https://doi.org/10.1016/j.engappai.2023.107351

    work page doi:10.1016/j.engappai.2023.107351 2024

  46. [46]

    L. Sun, H. Zhang, Y . Zheng, Z. Wu, Z. Ye, H. Zhao, Mass- former: Memory-augmented spectral-spatial transformer for hyperspectral image classification, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–15. doi:10.1109/TGRS.2024.3392264. 17

    work page doi:10.1109/tgrs.2024.3392264 2024