Pith. sign in

REVIEW 4 major objections 6 minor 36 references

DCT-Mamba3D: Spectral Decorrelation and Spatial-Spectral Feature Extraction for Hyperspectral Image Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A 3D discrete cosine transform front end lets a bidirectional state-space model classify hyperspectral images with 95-99.5% accuracy on three benchmarks.

desk verdict Reasonable architecture, but the baseline comparisons are so under-specified and implausible that the SOTA claim cannot be taken at face value. read the letter →

arxiv 2502.01986 v1 pith:U4MDLLAP submitted 2025-02-04 cs.CV eess.IV

Weijia Cao , Xiaofei Yang , Yicong Zhou , Zheng Zhang This is my paper · ORCID
classification cs.CVeess.IV
keywords hyperspectralimageclassification3DdiscretecosinetransformspectraldecorrelationMambastate-spacemodelspatial-spectralfeaturesIndianPinesKennedySpaceCenter
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

DCT-Mamba3D is a hyperspectral image classification framework that aims to turn spectral redundancy from an obstacle into an advantage. Its central claim is that applying a 3D discrete cosine transform over both spatial and spectral dimensions before feature extraction, then feeding the decorrelated frequency components through a bidirectional state-space model, yields higher classification accuracy than eight leading CNN, Transformer, and Mamba baselines. The paper reports overall accuracies of 95.23% on Indian Pines, 99.50% on Kennedy Space Center, and 98.15% on Houston2013 under a 10% training split, with particular gains on 'same object, different spectra' and 'different objects, same spectra' cases. A reader would care because these hard cases are exactly what limits real remote-sensing maps, and the proposed pipeline is a concrete recipe for attacking them.

What carries the argument

The load-bearing object is the 3D Spatial-Spectral Decorrelation Module (3D-SSDM), which applies the 3D discrete cosine transform basis functions $\psi_{i,j,k}$ of Eq. (2) to an HSI cube, converting correlated spatial-spectral pixels into independent frequency components. This is followed by the 3D-Mamba module, a bidirectional state-space model that selectively scans the spatial and spectral branches to capture long-range dependencies, and by the Global Residual Enhancement (GRE) module, which adds a learnable-weighted copy of the original frequency features to stabilize training. Together the modules turn frequency-domain decorrelation into a feature-extraction hierarchy: first concentrate energy, then model dependencies, then stabilize with a residual.

What would settle it

Retrain all eight baselines under the same 10% train/test splits with a hyperparameter search per model, and compare overall accuracy and per-class F1. If a properly tuned MiM, MorphF, or ViT matches or exceeds 95.23% on Indian Pines, 99.50% on KSC, or 98.15% on Houston2013, the central superiority claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a 3D discrete cosine transform can serve as a spatial-spectral decorrelation front end for Mamba-style state-space models, and that this combination separates classes whose spectral signatures overlap. The 3D-SSDM applies the 3D DCT basis functions $\psi_{i,j,k}(x,y,z)$ to the input cube $X\in\mathbb{R}^{H\times W\times C}$, producing frequency components in which redundant spectral and spatial content is compressed into fewer coefficients. The 3D-Mamba module then scans the decorrelated components bidirectionally along spatial and spectral axes, and the GRE module adds a learnable residual from the original frequency representation. The paper concludes from its experiments that this full pipeline outperforms 2D-CNN, 3D-CNN, HybridSN, ViT, HiT, MorphF, SSFTT, and MiM on all three benchmark datasets, and that the decorrelation module alone accounts for a large share of the gain in the ablation study.

Load-bearing premise

The comparison claim depends on the eight baseline models having been trained under the same 10% splits, the same hyperparameter tuning effort, and the same number of runs; if the baselines were undertuned, the reported accuracy margins could shrink.

Editorial extensions

If this is right

  • If the reported numbers hold, decorrelating spectra before state-space modeling is a broadly useful front end for hyperspectral classification, not a dataset-specific trick.
  • The 3D-SSDM's compression of redundant bands means DCT-Mamba3D maintains high accuracy with 1-3% training samples, which matters for labeling-scarce remote-sensing missions.
  • The Spearman heatmaps and training-loss curves imply the architecture's gains come from reduced inter-band correlation and faster, more stable convergence, so similar pipelines can be expected to train quickly.
  • For classes with near-identical spectral signatures (e.g., Buildings-Grass-Trees-Drives vs. Stone-Steel-Towers), the framework should yield materially better per-class F1 than non-decorrelating baselines.

Reading between the lines

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

  • A natural extension the authors only gesture at is adaptive frequency selection: choosing which DCT coefficients survive based on class-discriminative power could push the same architecture further on datasets with different band structures.
  • Because the decorrelation front end is architecture-agnostic, one could test whether applying the same 3D-DCT preprocessing to a plain CNN or Transformer reproduces part of the gain, isolating the contribution of Mamba.
  • The reported per-class zeroes for some baselines suggest a controlled re-benchmark under equal tuning could narrow the margins; that is a testable prediction rather than a settled result.
  • If the Spearman-heatmap logic is right, the same decorrelation should improve other redundancy-dominated tasks, such as multispectral or synthetic-aperture-radar classification.
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 / 6 minor

Summary. The paper proposes DCT-Mamba3D, a hyperspectral image classification framework that combines a 3D discrete cosine transform based spectral-spatial decorrelation module (3D-SSDM), a bidirectional state-space model (3D-Mamba), and a global residual enhancement (GRE) module. The authors claim state-of-the-art results on Indian Pines, Kennedy Space Center, and Houston2013 under a 10% training split, with particular gains on classes exhibiting spectral variability or spectral similarity. The paper includes an ablation study, t-SNE visualizations, training loss comparisons, and a complexity analysis to support the architecture's effectiveness. The central claim is empirical: DCT-Mamba3D outperforms eight baselines across three benchmarks.

Significance. If the reported results are reproducible under a fair protocol, DCT-Mamba3D would be a useful contribution to hyperspectral image classification, offering a frequency-domain decorrelation approach that is relatively underexplored in Mamba-based HSI methods. The architecture combines three established ideas (DCT, SSM, residual connections) in a new way, and the ablations suggest each component contributes positively. However, the significance of the claimed state-of-the-art margins is currently undermined by the lack of credible baseline comparisons: several baselines exhibit degenerate per-class accuracies and implausibly inconsistent performance across datasets, and no training details or error bars are provided. The paper's contribution is therefore best viewed as a promising architecture awaiting rigorous empirical validation.

major comments (4)
  1. [Section 4.3, Tables 1-3] The baseline comparison issue is load-bearing because the abstract and conclusion claim state-of-the-art performance; a fair comparison is required to support that claim.
  2. [Section 3.1, Eqs. (1)-(2)] This ambiguity is load-bearing because the decorrelation module is the paper's first contribution and all downstream modules depend on its output tensor shape.
  3. [Section 3.2, Eq. (4)] The 3D-Mamba module is a named contribution; the missing state update is a load-bearing gap in the method description.
  4. [Section 3.3, Eq. (7)] This is a secondary but still method-critical omission because the loss function is part of the proposed framework.
minor comments (6)
  1. [Section 3.2, Eq. (3)] The subscript in 'x spectral' has a space; it should be 'x_spectral' for clarity.
  2. [Section 4.5, Table 5] Table 5 reports only OA and Kappa for varying sample sizes, but the abstract claims improvements in 'same object, different spectra' and 'different objects, same spectra' scenarios; please report AA and per-class results for the classes in question at these sample sizes to substantiate that claim.
  3. [Section 4.4, Figure 6] The training loss comparison plots 2D-CNN and HiT only, not the other baselines; the caption should state the conditions under which each curve was recorded (same epochs, same optimizer, same hardware) to make the convergence claim meaningful.
  4. [Section 4.3.1, Figure 5] The t-SNE plots do not include class labels or a legend, making it difficult to assess whether clusters correspond to the ground-truth classes; please add a legend or describe the color mapping.
  5. [Table 6] The complexity table reports 'FLOPS' but does not specify the input size used for the measurement; please state the tensor dimensions and the hardware/software used for the FLOPs count.
  6. [Section 3.1 and 4.5] The module is called '3D-SSDM' in the methodology and '3D-SDM' in Table 4's first row; please standardize the naming to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported accuracy gains are empirical benchmark results, not derived by construction from the model's equations or from self-citation.

full rationale

DCT-Mamba3D's central claim is an empirical accuracy comparison (Tables 1-3). The architecture is specified by fixed 3D DCT basis functions (Eq. 2), a bidirectional SSM (Eq. 4), a residual combination (Eq. 6), and cross-entropy training (Eq. 7); none of these equations has a parameter that is fitted to the test set and then renamed as a prediction. The decorrelation benefit is a standard mathematical property of DCT, and the ablation in Table 4 is an ordinary component-removal test, not a prediction forced by construction. The only self-citations are to the authors' own earlier baselines (2D-CNN [28], 3D-CNN [29], HiT [30]), used as comparison methods in Tables 1-3 and Figure 6; these citations do not justify the claimed advantage, which rests on the reported benchmark numbers. The reviewer's concern that some baselines appear undertrained (e.g., MorphF at 80.36% on KSC and several 0.00 per-class scores) is a legitimate experimental-validity threat, but it is not circular reasoning: the proposed method's numbers are not manufactured from its own definitions. No equation or citation chain reduces the central claim to its inputs, so no circular step can be quoted.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method relies on well-known DCT and Mamba components, but key implementation parameters are left unspecified. No new theoretical entities are introduced.

free parameters (4)
  • patch_size (3x3x3) = 3x3x3 (implied)
    The text refers to a 3x3x3 DCT setup yielding 27 basis functions, but the patch size and overlapping strategy are not defined; this is a hand-chosen architectural parameter.
  • loss weight lambda = not specified
    Eq. (7) includes a regularization term lambda L_reg, but lambda and L_reg are never defined or reported.
  • Mamba hidden dimensions and number of layers = not specified
    The state-space model size, number of blocks, and scan order are not given, making the architecture under-specified.
  • GRE mixing weight alpha = learned
    Alpha in Eq. (6) is a learnable scalar; its value or initialization is not reported.
assumptions (3)
  • domain assumption The 3D DCT decorrelates spectral and spatial dimensions, reducing redundancy in HSI data (Eq. 1-2).
    This is a known property of the DCT for signals with high inter-band correlation, but the paper does not prove it for HSI patches.
  • ad hoc to paper The bidirectional state-space model in Eq. (4) can be trained with the selective scan algorithm of Mamba, though the state update is omitted.
    The paper invokes Mamba's machinery without providing the required equations, so the reader must assume a standard Mamba implementation fills the gap.
  • domain assumption The datasets and train/test splits are treated as ground truth.
    Standard evaluation protocol in HSI classification, but the exact split (10% training, 90% test) is asserted without a reference to a standard protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DCT-Mamba3D: Spectral Decorrelation and Spatial-Spectral Feature Extraction for Hyperspectral Image Classification." pith.science (2026). https://pith.science/paper/U4MDLLAP

@misc{pith2026250201986,
  author       = {Pith},
  title        = {Pith review of: DCT-Mamba3D: Spectral Decorrelation and Spatial-Spectral Feature Extraction for Hyperspectral Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U4MDLLAP}},
  note         = {Machine review of arXiv:2502.01986}
}
read the original abstract

Hyperspectral image classification presents challenges due to spectral redundancy and complex spatial-spectral dependencies. This paper proposes a novel framework, DCT-Mamba3D, for hyperspectral image classification. DCT-Mamba3D incorporates: (1) a 3D spectral-spatial decorrelation module that applies 3D discrete cosine transform basis functions to reduce both spectral and spatial redundancy, enhancing feature clarity across dimensions; (2) a 3D-Mamba module that leverages a bidirectional state-space model to capture intricate spatial-spectral dependencies; and (3) a global residual enhancement module that stabilizes feature representation, improving robustness and convergence. Extensive experiments on benchmark datasets show that our DCT-Mamba3D outperforms the state-of-the-art methods in challenging scenarios such as the same object in different spectra and different objects in the same spectra.

Figures

Figures reproduced from arXiv: 2502.01986 by the authors.

Figure 1
Figure 1. Spectral response functions illustrating HSI classifica [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DCT-Mamba3D framework. mospheric conditions, or intrinsic material differences) and similarity between materials exacerbate classification chal￾lenges [9, 21, 31]. High inter-band correlation leads to re￾dundant information, complicating differentiation, particu￾larly in mixed pixels where each pixel may represent multi￾ple materials [3]. Frequency-Domain Transformations for Enhanced Feature Extraction: Frequency-do… view at source ↗
Figure 3
Figure 3. 3D Spatial-Spectral Decorrelation Module (3D-SSDM), [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Spearman correlation heatmaps on the Indian Pines [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: t-SNE visualization of feature embeddings across models on the KSC dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Training loss comparison between 2D-CNN, HiT, and [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

  1. [1]

    WaveMamba: Spatial-Spectral Wavelet Mamba for Hyperspectral Image Classification

    Muhammad Ahmad, Muhammad Usama, and Manual Mazzara. Wavemamba: Spatial-spectral wavelet mamba for hyperspectral image classification. arXiv preprint arXiv:2408.01231, 2024. 2, 3

    work page Pith review arXiv 2024

  2. [2]

    Discrete cosine transform

    Nasir Ahmed, T Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE Transactions on Computers, 100(1): 90–93, 1974. 3

    work page 1974

  3. [3]

    Hyperspectral image resolution enhancement using high- resolution multispectral image based on spectral unmixing

    Mohamed Amine Bendoumi, Mingyi He, and Shaohui Mei. Hyperspectral image resolution enhancement using high- resolution multispectral image based on spectral unmixing. IEEE Transactions on Geoscience and Remote Sensing , 52 (10):6574–6583, 2014. 2

    work page 2014

  4. [4]

    Deep feature extraction and classification of hyperspectral images based on convolutional neural net- works

    Yushi Chen, Hanlu Jiang, Chunyang Li, Xiuping Jia, and Pe- dram Ghamisi. Deep feature extraction and classification of hyperspectral images based on convolutional neural net- works. IEEE Transactions on Geoscience and Remote Sens- ing, 54(10):6232–6251, 2016. 2

    work page 2016

  5. [5]

    Psrt: Pyramid shuffle-and- reshuffle transformer for multispectral and hyperspectral im- age fusion

    Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Dan- feng Hong, and Gemine Vivone. Psrt: Pyramid shuffle-and- reshuffle transformer for multispectral and hyperspectral im- age fusion. IEEE Transactions on Geoscience and Remote Sensing, 61:1–15, 2023. 1

    work page 2023

  6. [6]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 4

    arXiv 2010

  7. [7]

    Cat: Center attention transformer with strati- fied spatial-spectral token for hyperspectral image classifica- tion

    Jiaqi Feng, Qixiong Wang, Guangyun Zhang, Xiuping Jia, and Jihao Yin. Cat: Center attention transformer with strati- fied spatial-spectral token for hyperspectral image classifica- tion. IEEE Transactions on Geoscience and Remote Sensing,

    work page

  8. [8]

    Recent ad- vances on spectral–spatial hyperspectral image classifica- tion: An overview and new guidelines

    Lin He, Jun Li, Chenying Liu, and Shutao Li. Recent ad- vances on spectral–spatial hyperspectral image classifica- tion: An overview and new guidelines. IEEE Transactions on Geoscience and Remote Sensing, 56(3):1579–1597, 2017. 1

    work page 2017

Show all 36 references
  1. [9]

    Spectralgpt: Spectral remote sensing foun- dation model

    Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xi- uping Jia, et al. Spectralgpt: Spectral remote sensing foun- dation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2

    2024

  2. [10]

    3D Gabor convolu- tional neural network for hyperspectral image classification

    Sen Jia, Jianhui Liao, Meng Xu, Yan Li, Jiasong Zhu, Wei- wei Sun, Xiuping Jia, and Qingquan Li. 3D Gabor convolu- tional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing , 60: 1–16, 2022. 2

    2022

  3. [11]

    Spec- tral–spatial hyperspectral image segmentation using sub- space multinomial logistic regression and markov random fields

    Jun Li, Jos ´e M Bioucas-Dias, and Antonio Plaza. Spec- tral–spatial hyperspectral image segmentation using sub- space multinomial logistic regression and markov random fields. IEEE Transactions on Geoscience and Remote Sens- ing, 50(3):809–823, 2012. 1

    2012

  4. [12]

    Latent diffusion enhanced rect- angle transformer for hyperspectral image restoration

    Miaoyu Li, Ying Fu, Tao Zhang, Ji Liu, Dejing Dou, Cheng- gang Yan, and Yulun Zhang. Latent diffusion enhanced rect- angle transformer for hyperspectral image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence,

  5. [13]

    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. 3

    2024

  6. [14]

    Deep learning classifiers for hyperspectral imaging: A review

    Matteo E Paoletti, Jean M Haut, Javier Plaza, and Antonio Plaza. Deep learning classifiers for hyperspectral imaging: A review. ISPRS Journal of Photogrammetry and Remote Sensing, 158:279–317, 2019. 1

    2019

  7. [15]

    A dual frequency transformer network for hyperspectral image classification

    Xin Qiao and Weimin Huang. A dual frequency transformer network for hyperspectral image classification. IEEE Jour- nal of Selected Topics in Applied Earth Observations and Remote Sensing, pages 1–12, 2023. 2, 3

    2023

  8. [16]

    HybridSN: Exploring 3D–2D CNN Feature Hierarchy for Hyperspectral Image Classification

    Swapnil Karun Roy, Gopal Krishna, and Shiv Ram Dubey. HybridSN: Exploring 3D–2D CNN Feature Hierarchy for Hyperspectral Image Classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277–281, 2020. 2, 4

    2020

  9. [17]

    Spectral–spatial morpho- logical attention transformer for hyperspectral image clas- sification

    Soumya Kanti Roy, Akash Deria, Chirag Shah, Jean-Marc Haut, Qian Du, and Antonio Plaza. Spectral–spatial morpho- logical attention transformer for hyperspectral image clas- sification. IEEE Transactions on Geoscience and Remote Sensing, 61:1–15, 2023. 2, 4

    2023

  10. [18]

    Masked vision transformers for hyperspectral image classi- fication

    Linus Scheibenreif, Michael Mommert, and Damian Borth. Masked vision transformers for hyperspectral image classi- fication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2166–2176,

  11. [19]

    Dct-mask: Discrete cosine transform mask rep- resentation for instance segmentation

    Xing Shen, Jirui Yang, Chunbo Wei, Bing Deng, Jian- qiang Huang, Xian-Sheng Hua, Xiaoliang Cheng, and Kewei Liang. Dct-mask: Discrete cosine transform mask rep- resentation for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern reco...

    2021

  12. [20]

    Spectral–spatial feature tokenization transformer for hyper- spectral image classification

    Liang Sun, Guangming Zhao, Yisong Zheng, and Zuxin Wu. Spectral–spatial feature tokenization transformer for hyper- spectral image classification. IEEE Transactions on Geo- science and Remote Sensing, 60:1–14, 2022. 4

    2022

  13. [21]

    Spectral variability of remotely sensed target materials: Causes, models, and strategies for mitigation and robust exploitation

    James Theiler, Amanda Ziemann, Stefania Matteoli, and Marco Diani. Spectral variability of remotely sensed target materials: Causes, models, and strategies for mitigation and robust exploitation. IEEE Geoscience and Remote Sensing Magazine, 7(2):8–30, 2019. 1, 2

    2019

  14. [22]

    Har- monic networks for image classification

    Matej Ulicny, Vladimir A Krylov, and Rozenn Dahyot. Har- monic networks for image classification. 2019. 2, 3

    2019

  15. [23]

    Harmonic convolutional networks based on discrete cosine transform

    Matej Ulicny, Vladimir A Krylov, and Rozenn Dahyot. Harmonic convolutional networks based on discrete cosine transform. Pattern Recognition, 129:108707, 2022. 2, 3

    2022

  16. [24]

    Frequency domain- based features for hyperspectral image classification

    Ke Wang, Bin Yong, and Zhaohui Xue. Frequency domain- based features for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters , 16(9):1417–1421,

  17. [25]

    An efficient hy- perspectral image classification method using deep fusion of 3D discrete wavelet transform and CNN

    Jingran Xu, Jiankang Zhao, and Yuchao Fu. An efficient hy- perspectral image classification method using deep fusion of 3D discrete wavelet transform and CNN. IEEE Geoscience and Remote Sensing Letters, 20:1–5, 2023. 2

    2023

  18. [26]

    Learning in the frequency domain

    Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1740–1749, 2020. 2, 3

    2020

  19. [27]

    Ex- ploiting frequency correlation for hyperspectral image recon- struction

    Muge Yan, Lizhi Wang, Lin Zhu, and Hua Huang. Ex- ploiting frequency correlation for hyperspectral image recon- struction. arXiv preprint arXiv:2406.00683, 2024. 2

    2024 arXiv

  20. [28]

    Hyperspectral image classification with deep learning models

    Xiaofei Yang, Yunming Ye, Xutao Li, Raymond YK Lau, Xiaofeng Zhang, and Xiaohui Huang. Hyperspectral image classification with deep learning models. IEEE Transactions on Geoscience and Remote Sensing, 56(9):5408–5423, 2018. 2, 4

    2018

  21. [29]

    Synergistic 2D/3D con- volutional neural network for hyperspectral image classifica- tion

    Xiaofei Yang, Xin Zhang, Yun Ye, Raymond Y Lau, Sheng Lu, Xin Li, and Xiaodong Huang. Synergistic 2D/3D con- volutional neural network for hyperspectral image classifica- tion. Remote Sensing, 12(12):2033, 2020. 2, 4

    2020

  22. [30]

    Hy- perspectral image transformer classification networks

    Xiaofei Yang, Weijia Cao, Yanan Lu, and Yicong Zhou. Hy- perspectral image transformer classification networks. IEEE Transactions on Geoscience and Remote Sensing , 60:1–15,

  23. [31]

    Spectralmamba: Efficient mamba for hyperspectral im- age classification

    Jing Yao, Danfeng Hong, Chenyu Li, and Jocelyn Chanus- sot. Spectralmamba: Efficient mamba for hyperspectral im- age classification. arXiv preprint arXiv:2404.08489 , 2024. 1, 2, 3

    2024 arXiv

  24. [32]

    Wave-vit: Unifying wavelet and transformers for visual representation learning

    Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-vit: Unifying wavelet and transformers for visual representation learning. In European Conference on Com- puter Vision, pages 328–345. Springer, 2022. 3

    2022

  25. [33]

    Three-dimensional frequency domain transform network for cross-scene hyperspectral image classification

    Jun Zhang, Cheng Zhang, Shuai Liu, Zhenwei Shi, and Bin Pan. Three-dimensional frequency domain transform network for cross-scene hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing ,

  26. [34]

    Deep learning for remote sensing data: A technical tutorial on the state of the art

    Liangpei Zhang, Lefei Zhang, and Bo Du. Deep learning for remote sensing data: A technical tutorial on the state of the art. IEEE Geoscience and Remote Sensing Magazine , 4(2): 22–40, 2016. 1

    2016

  27. [35]

    Mamba-in-Mamba: Cen- tralized Mamba-Cross-Scan in Tokenized Mamba model for hyperspectral image classification

    Weilian Zhou, Sei-Ichiro Kamata, Haipeng Wang, Man Sing Wong, and Huiying Cynthia Hou. Mamba-in-Mamba: Cen- tralized Mamba-Cross-Scan in Tokenized Mamba model for hyperspectral image classification. Neurocomputing, page 128751, 2024. 2, 3, 4

    2024

  28. [36]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2

    2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.