Pith. sign in

REVIEW 4 major objections 7 minor 37 references

MSCM-net classifies hyperspectral images by pairing multi-scale CNN local features with Mamba long-sequence modeling, beating strong baselines at lower parameter cost.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 12:27 UTC pith:OJSHQLRO

load-bearing objection Coherent multi-scale-CNN + Mamba HSI classifier with real efficiency gains, but SOTA margins rest on single-run point estimates tuned on the test sets. the 4 major comments →

arxiv 2607.28277 v1 pith:OJSHQLRO submitted 2026-07-30 cs.CV

MSCM-net: A hyperspectral image classiffcation method based on multi-scale convolution and Mamba

Jianjun Chen , Linlin Wang , Lifang Chang , Limin Huo , Shujiang Song , Yanjia Zhao , Mingwei Shao This is my paper
classification cs.CV
keywords hyperspectral image classificationMambaselective state space modelmulti-scale convolutionSENetdeep learningspectral-spatial features
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Hyperspectral image classification needs both fine local spatial-spectral detail and long-range context, but CNNs are local and Transformers are expensive. This paper argues that a hybrid called MSCM-net solves that tradeoff: a multi-scale convolution plus SENet block (MCSE) extracts multi-receptive-field features, stacked Mamba blocks model long pixel sequences efficiently, residual links keep original information, and a dual-branch head fuses the center pixel with global average context. On Indian Pines, WHU-Hi-HongHu, and Salinas with 30 labeled samples per class, the model reports the best overall accuracy, average accuracy, and Kappa among the compared methods while using only about 0.2 million parameters. A sympathetic reader cares because remote-sensing maps must be accurate under scarce labels and limited compute; if the hybrid really delivers both, it is a practical path past pure CNN or Transformer designs.

Core claim

The authors claim that combining multi-scale CNN feature extraction (MCSE with 1×1, 3×3, and 5×5 branches plus SENet) with stacked selective state-space (Mamba) blocks, residual fusion of original features, and a center-plus-global dual-branch aggregator yields state-of-the-art hyperspectral classification accuracy on three standard benchmarks while cutting parameter count and keeping competitive FLOPs and inference time relative to CNN, Transformer, and prior Mamba baselines.

What carries the argument

MSCM-net: PCA-patched HSI cubes go through spectral compression, an MCSE multi-scale convolution+SENet module, residual-fused serialization into stacked Mamba encoder blocks, then a dual-branch center/global feature aggregator into an MLP classifier.

Load-bearing premise

That single-run averages under a fixed 30-samples-per-class split and hyperparameters tuned on the same three test sets are enough to prove lasting superiority over strong baselines.

What would settle it

Retrain MSCM-net and the same baselines on multiple independent 30-per-class draws (or published standard splits) with several random seeds and check whether the reported OA margins (roughly 0.4–2 points) and the parameter/accuracy ranking still hold with reported means and standard deviations.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Multi-scale CNN local extractors plus selective state-space sequence models can replace heavier self-attention stacks for patch-based HSI classification.
  • Center-pixel emphasis fused with global average pooling is a lightweight way to keep category-critical spectral cues without full attention.
  • With ~0.2M parameters and sub-second-class test times on Indian Pines-scale data, the design is usable where label budgets and onboard compute are tight.
  • Ablations imply four Mamba blocks, SENet channel reweighting, and the residual path before Mamba are load-bearing for the reported gains.

Where Pith is reading between the lines

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

  • If the efficiency claim generalizes, similar multi-scale+Mamba stacks may transfer to other dense spectral cubes (e.g., medical or industrial spectroscopy) where Transformers are currently too costly.
  • The dual-branch center/global head is close in spirit to query-center designs in recent HSI Transformers; testing it as a drop-in head on pure Mamba or pure CNN backbones would isolate how much of the gain is architectural versus backbone choice.
  • Absence of cross-dataset hyperparameter transfer tests leaves open whether the chosen patch size 15 and four-block depth are dataset-specific or a stable default.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes MSCM-net for hyperspectral image classification, combining a multi-scale convolutional front-end (MCSE: parallel 1×1, 3×3 and 5×5 depthwise-separable branches followed by SENet channel reweighting) with stacked selective state-space (Mamba) blocks and a dual-branch center/global feature aggregation head. After PCA and a 3×3 spectral compression layer, patches are serialized, residual-fused with projected original features plus positional encoding, processed by L=4 Mamba blocks, then classified via element-wise sum of an enhanced center token and global average pooling. Experiments on Indian Pines, WHU-Hi-HongHu and Salinas under a fixed 30-samples-per-class protocol report best OA/AA/Kappa versus SVM, 3DCNN, SSRN, SSFTT, MASSFormer, SQSFormer and SSMamba, with the smallest parameter count (0.201M) and competitive FLOPs/test time on Indian Pines (Table 5). Ablations cover Mamba depth, SENet, residual connection and center vs global branches; a hyperparameter study covers learning rate, batch size and patch size.

Significance. If the accuracy–efficiency claims hold under proper multi-run evaluation, the work is a useful incremental contribution to HSI classification: it shows that multi-scale depthwise CNN + SENet as a local front-end, residual injection into Mamba, and a simple center/global head can match or slightly exceed recent Transformer and Mamba baselines at markedly lower parameter count. The architecture is coherent, equations for MCSE, SENet, Mamba discretization and dual-branch fusion are given, and ablations plus complexity numbers support the design choices. The contribution is primarily empirical/engineering rather than theoretical; significance rests on whether the reported ~0.4–2% OA margins are stable.

major comments (4)
  1. [§3.3, Tables 2–5] Tables 2–4 and Table 5 report single point estimates (OA/AA/Kappa, params, FLOPs, test time) with no standard deviations, no multi-seed means, and no repeated stratified train/test splits. OA margins over the strongest baselines are small (+1.16% vs SQSFormer on Indian Pines, +0.38% vs SSMamba on WHU-Hi-HongHu, +0.43% vs MASSFormer on Salinas). In 30-shot HSI, split/seed variance commonly reaches or exceeds these margins. Without multi-run statistics the central claim of “advanced classification performance while reducing computational complexity” is not established. Please report mean±std over ≥5–10 random seeds/splits for all methods (or at least MSCM-net and the top 2–3 baselines) on all three datasets.
  2. [§3.4, Figures 6–9; §4, Figure 10] §3.4 and Figures 6–9 select learning rate (1e-3), batch size (64), patch size (15) and, via Fig. 10 / §4, Mamba depth L=4 by sweeping on the same three evaluation sets used for the leaderboard. This couples model selection to the reported test metrics and inflates the risk that the ranking is overfit to these three scenes. At minimum, state the selection protocol clearly, freeze hyperparameters chosen on one dataset when evaluating the others, or use a nested/validation split; ideally add results under a fixed protocol shared with baselines.
  3. [§3.5, Table 5] Complexity (parameters, FLOPs, test time) is reported only for Indian Pines (Table 5). The efficiency half of the title claim should be supported on WHU-Hi-HongHu and Salinas as well (same patch/protocol), and FLOPs measurement settings (input shape, whether PCA/conv front-end is included) should be specified so comparisons to SSMamba (2.844G) and SSFTT (0.078G) are interpretable.
  4. [§3.3, Tables 2–4] Text in §3.3 refers to “Table 4” for Indian Pines and “Table 5”/“Table 6” for HongHu/Salinas, but the actual tables are numbered 2–4 (results) and 5 (complexity). Several per-class “best” claims and the narrative comparison to “suboptimal” methods are hard to audit under this mismatch. Correct cross-references and ensure every bold/best entry in Tables 2–4 is consistent with the stated OA/AA/Kappa ranking.
minor comments (7)
  1. [Title, Abstract, passim] Title and elsewhere: “classiffcation” → “classification”; abstract “methodshaveadvanced, theystill facelocalityconstraints”; numerous missing spaces after punctuation and concatenated words (e.g., “themulti-scalecon-volution”, “dependencyrelationship”). A full copy-edit pass is needed.
  2. [§2.1–2.3, Figure 1] Figure 1 and §2.1: patch size is stated as P=15, but MCSE/Mamba text repeatedly uses 13×13 / L=169. Clarify whether spatial size after the initial 3×3 conv (no explicit stride/padding stated) is 13 or 15, and make figure and equations consistent.
  3. [§2.3] Eqs. (15)–(21): D_t is described as “state dimension” while appearing as a skip term in y_t = C_t h_t + D_t x_t; standard SSM notation uses D as a feedthrough matrix. Align notation with Mamba literature and define A’s shape and whether it is input-dependent.
  4. [§3.1, Table 1; §3.3] Table 1 training counts: several classes use fewer than 30 samples (e.g., Grass-pasture-mowed 14, Oats 10) while the text says “training samples are set to 30.” State the exact sampling rule (min(30, n_c) or similar).
  5. [Figures 3–5] Classification maps (Figures 3–5) lack a ground-truth panel and color legend matching Figure 2, which weakens visual comparison. Add GT and a shared legend.
  6. [§1] Related work cites several 2025 Mamba-HSI methods (DBMGNet, SEDGM, MambaHSI, DBMLLA) but does not discuss how MCSE+center/global differs from their spatial-spectral dual-branch designs beyond a high-level claim. A short differentiation paragraph would help novelty assessment.
  7. [Highlights] Highlights still contain placeholders (“Research highlight 1” …). Replace with the actual bullet claims before publication.

Circularity Check

0 steps flagged

No derivation circularity: empirical CNN–Mamba architecture paper evaluated on external HSI benchmarks.

full rationale

MSCM-net is an engineering/architecture paper, not a first-principles derivation. The load-bearing claim is empirical superiority (OA/AA/Kappa and parameter/FLOP counts) versus named external baselines (SVM, 3DCNN, SSRN, SSFTT, MASSFormer, SQSFormer, SSMamba) on three public datasets under a fixed 30-shot protocol. Architecture pieces (multi-scale depthwise conv, SENet, stacked selective SSM/Mamba blocks, residual fusion, center/global dual-branch head) are composed from standard components and described with ordinary layer equations; none of those equations defines the reported accuracy in terms of a fitted constant that is then re-presented as a prediction. Citations are to external methods and the original Mamba/SSM literature, not to an author-owned uniqueness theorem that forces the design. Hyperparameter sweeps (lr, batch, patch, Mamba depth) on the same three sets are ordinary model selection and do not collapse the leaderboard metrics by construction. No self-definitional loop, fitted-input-as-prediction, or renaming of a known closed-form result is present. Score 0 with empty steps is the appropriate finding.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 3 invented entities

Load-bearing content is engineering choice plus benchmark protocol, not physical axioms. Claims rest on standard DL building blocks, the assumption that 15×15 patches with 30 labels/class represent the task, and hand-chosen depth/width hyperparameters validated on the same sets used for leaderboard comparison.

free parameters (8)
  • learning_rate = 1e-3
    Selected by sweep on the three evaluation datasets; peak claimed at 0.001 (Figure 6).
  • batch_size = 64
    Chosen by sweep; 64 best on all three sets (Figure 7).
  • patch_size_P = 15
    Spatial context window swept and fixed at 15 for all reported SOTA tables (Figure 8).
  • Mamba_block_depth_L = 4
    Ablated 0–5; L=4 chosen as best accuracy/complexity tradeoff (Figure 10).
  • post_PCA_conv_channels = 64
    Spectral compression width set to 64 channels before MCSE (§2.1).
  • Mamba_feature_dim = 128
    Sequence model width set to dim=128 after rearrange (§2.3).
  • MCSE_branch_output_channels_N = 64 per branch (192 concat)
    Each of three conv branches outputs N=64, concat to 192 (§2.2).
  • train_samples_per_class = 30 (mostly)
    Few-shot protocol fixed at 30 labeled pixels per class (14/10 for tiny IP classes) for all main tables.
axioms (6)
  • domain assumption Selective state-space (Mamba) layers provide efficient long-range sequence modeling with hardware-efficient scan, suitable after flattening HSI patches.
    Invoked throughout §1–2.3 and Eqs. 15–21; taken from Gu & Dao Mamba line and prior HSI-Mamba papers.
  • domain assumption PCA plus a learned 3×3 spectral compression to 64 bands preserves discriminative spatial-spectral information for classification.
    Preprocessing stage in §2.1; standard HSI practice but not proven optimal here.
  • domain assumption The center pixel of a patch carries the most complete class label information, justifying explicit center-feature extraction.
    Stated in §2.1 and §2.4 as prior knowledge for the dual-branch head.
  • ad hoc to paper Depthwise-separable 3×3/5×5 multi-scale conv + SENet channel reweighting is an adequate local spatial-spectral front-end.
    MCSE design choice in §2.2; supported by ablation Table 6 but not derived.
  • ad hoc to paper Element-wise sum of enhanced center features and global average pool is sufficient fusion (no learned gate).
    Eq. 26 dual-branch fusion; ablation Table 8 supports it empirically only.
  • domain assumption Standard supervised cross-entropy training on randomly (or fixed) sampled 30-shot splits yields comparable rankings across methods.
    Experimental protocol §3.2–3.3; split randomness and seeds not fully specified.
invented entities (3)
  • MSCM-net (full pipeline) no independent evidence
    purpose: Name the combined HSI classifier: MCSE + residual + stacked Mamba + dual-branch aggregation + MLP head.
    Composite architecture; no new physical entity. Independent evidence is only the paper’s own benchmark tables.
  • MCSE module no independent evidence
    purpose: Multi-scale 1×1/3×3/5×5 conv branches concatenated and SENet-reweighted for local spatial-spectral features.
    Named module assembling known conv and SENet parts; falsifiable only via ablations in this paper.
  • Dual-branch center-global feature aggregation module no independent evidence
    purpose: Extract center token and sequence mean from Mamba output and sum them for classification.
    Task-specific head design; supported only by internal ablation Table 8.

pith-pipeline@v1.2.0-daily-grok45 · 22247 in / 4102 out tokens · 75258 ms · 2026-07-31T12:27:17.972526+00:00 · methodology

0 comments
read the original abstract

Hyperspectral imaging is widely used in remote sensing and engineering. Therefore, research on its classification methods is crucial. While CNN and Transformer-based methods have advanced, they still face locality constraints and high computational complexity. To address these issues, we propose an innovative hyperspectral image classification model, MSCM-net. Specifically, first of all, a model architecture combining multi-scale CNN and Mamba is proposed. It consists of a multi-scale feature extraction module (MCSE) and multiple stacked Mamba blocks, which integrates the local feature extraction capability of multi-scale CNN and the long sequence modeling advantage of Mamba. Secondly, the proposed MCSE module consists of multi-scale convolution and SENet. Convolution kernels of different scales extract local information with different receptive fields, enhancing the fusion of spatial and spectral information. Meanwhile, the SENet enables the model to automatically learn the importance of each channel in the multi-scale features. Furthermore, we also propose a dual-branch feature aggregation module, which further effectively extracts and integrates the spectral information contained in the central pixel and the spatial information in the surrounding pixels. Our model has undergone numerous experiments on three widely used benchmark datasets. The experimental results show that MSCM-net can achieve advanced classification performance while reducing computational complexity.

Figures

Figures reproduced from arXiv: 2607.28277 by Jianjun Chen, Lifang Chang, Limin Huo, Linlin Wang, Mingwei Shao, Shujiang Song, Yanjia Zhao.

Figure 1
Figure 1. Figure 1: Overall architecture of the MSCM-net model [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Images and category information of the three datasets. (a) Indian Pines. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Classification results of different methods on the Indian Pines dataset (a)SVM [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Classification results using different methods on the WHU-Hi-HongHu dataset [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Classification results of different methods on the Salinas dataset (a)SVM [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Line graph showing the impact of learning rate on accuracy on three datasets [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Line graph showing the impact of batch size on accuracy across three datasets [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Line graph showing the influence of patch size on accuracy across the three [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Heatmap showing the impact of hyperparameters on the classification perfor [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Line graph showing the ablation experiment results of Mamba block quantity [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗

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

37 extracted references · 18 canonical work pages

  1. [1]

    J. Gao, G. Chen, X. Ji, C. Shen, Reinforced graph aggrega- tion cross-domain few-shot learning for hyperspectral remote sens- ing image classification, SIGNAL PROCESSING 238 (2026). doi:10.1016/j.sigpro.2025.110101

    arXiv 2026

  2. [2]

    J. Zhao, K. Wu, L. Zhang, W. Huang, C. Ruan, L. Huang, Patch- based hierarchical residual spectral-spatial convolutional network for hyperspectral image classification, SIGNAL PROCESSING 230 (2025). doi:10.1016/j.sigpro.2024.109850

    arXiv 2025

  3. [3]

    X. Wang, Q. Hu, Y. Cheng, J. Ma, Hyperspectral image super- resolution meets deep learning: A survey and perspective, IEEE- CAA JOURNAL OF AUTOMATICA SINICA 10 (2023) 1668–1691. doi:10.1109/JAS.2023.123681. 26

    arXiv 2023

  4. [4]

    X. Yang, K. Ma, D. Zhang, S. Song, X. An, Classification of soybean seeds based on rgb reconstruction of hyperspectral images, PLOS ONE 19 (2024). doi:10.1371/journal.pone.0307329

    work page doi:10.1371/journal.pone.0307329 2024

  5. [5]

    F. Ali, A. Razzaq, W. Tariq, A. Hameed, A. Rehman, K. Razzaq, S. Sarfraz, N. A. Rajput, H. E. M. Zaki, M. S. Shahid, G. On- drasek, Spectral intelligence: Ai-driven hyperspectral imaging for agri- cultural and ecosystem applications, AGRONOMY-BASEL 14 (2024). doi:10.3390/agronomy14102260

    work page doi:10.3390/agronomy14102260 2024

  6. [6]

    J. Wang, G. Chen, J. Ju, T. Lin, R. Wang, Z. Wang, Characteriza- tion and classification of urban weed species in northeast china using terrestrial hyperspectral images, WEED SCIENCE 71 (2023) 353–368. doi:10.1017/wsc.2023.36

    work page doi:10.1017/wsc.2023.36 2023

  7. [7]

    Ruo-xi, F

    S. Ruo-xi, F. Yi-ning, C. Wei, W. Xiang-hai, Advance in hyperspec- tral images change detection, SPECTROSCOPY AND SPECTRAL ANALYSIS 43 (2023) 2354–2362. doi:10.3964/j.issn.1000-0593(2023)08- 2354-09

    work page doi:10.3964/j.issn.1000-0593(2023)08- 2023

  8. [8]

    P. Du, K. Tan, X. Xing, A novel binary tree support vector machine for hyperspectral remote sensing image classification, OPTICS COMMU- NICATIONS 285 (2012) 3054–3060. doi:10.1016/j.optcom.2012.02.092

    work page doi:10.1016/j.optcom.2012.02.092 2012

  9. [9]

    Zhang, G

    Y. Zhang, G. Cao, X. Li, B. Wang, Cascaded random forest for hyper- spectral image classification, IEEE JOURNAL OF SELECTED TOP- ICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENS- ING 11 (2018) 1082–1094. doi:10.1109/JSTARS.2018.2809781

    arXiv 2018

  10. [10]

    G. Liu, L. Wang, D. Liu, L. Fei, J. Yang, Hyperspectral image classifica- tionbasedonnon-parallelsupportvectormachine, REMOTESENSING 14 (2022). doi:10.3390/rs14102447

    work page doi:10.3390/rs14102447 2022

  11. [11]

    Zhou, Deep distributed convolutional neural networks: Uni- versality, ANALYSIS AND APPLICATIONS 16 (2018) 895–919

    D.-X. Zhou, Deep distributed convolutional neural networks: Uni- versality, ANALYSIS AND APPLICATIONS 16 (2018) 895–919. doi:10.1142/S0219530518500124

    work page doi:10.1142/s0219530518500124 2018

  12. [12]

    Yasin, M

    M. Yasin, M. Sarigul, M. Avci, Logarithmic learning differential convolutional neural network, NEURAL NETWORKS 172 (2024). doi:10.1016/j.neunet.2024.106114. 27

    arXiv 2024

  13. [13]

    D. J. Atha, M. R. Jahanshahi, Evaluation of deep learning approaches based on convolutional neural networks for corrosion detection, STRUC- TURAL HEALTH MONITORING-AN INTERNATIONAL JOURNAL 17 (2018) 1110–1128. doi:10.1177/1475921717737051

    work page doi:10.1177/1475921717737051 2018

  14. [14]

    X. Yang, Y. Ye, X. Li, R. Y. K. Lau, X. Zhang, X. Huang, Hyperspectral image classification with deep learning models, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 56 (2018) 5408–5423. doi:10.1109/TGRS.2018.2815613

    arXiv 2018

  15. [15]

    A. Paul, S. Bhoumik, N. Chaki, Ssnet: an improved deep hybrid net- work for hyperspectral image classification, NEURAL COMPUTING & APPLICATIONS 33 (2021) 1575–1585. doi:10.1007/s00521-020-05069- 1

    work page doi:10.1007/s00521-020-05069- 2021

  16. [16]

    doi:10.3390/s19235276

    F.Feng, S.Wang, C.Wang, J.Zhang, Learningdeephierarchicalspatial- spectral features for hyperspectral image classification based on residual 3d-2d cnn, SENSORS 19 (2019). doi:10.3390/s19235276

    work page doi:10.3390/s19235276 2019

  17. [17]

    X. Lei, F. Xie, C. Jin, Gram angle field-based siamese graph convolutional neural network for hyperspectral images classification, IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 22 (2025). doi:10.1109/LGRS.2024.3521461

    arXiv 2025

  18. [18]

    X. Hu, T. Li, T. Zhou, Y. Liu, Y. Peng, Contrastive learning based on transformer for hyperspectral image classification, APPLIED SCIENCES-BASEL 11 (2021). doi:10.3390/app11188670

    work page doi:10.3390/app11188670 2021

  19. [19]

    Y. Tan, M. Li, L. Yuan, C. Shi, Y. Luo, G. Wen, Hyperspectral image classification with embedded linear vision transformer, EARTH SCI- ENCE INFORMATICS 18 (2024). doi:10.1007/s12145-024-01651-6

    work page doi:10.1007/s12145-024-01651-6 2024

  20. [20]

    C. Fu, T. Zhou, T. Guo, Q. Zhu, F. Luo, B. Du, Cnn-transformer and channel-spatial attention based network for hyperspectral image classification with few samples, NEURAL NETWORKS 186 (2025). doi:10.1016/j.neunet.2025.107283

    arXiv 2025

  21. [21]

    Fan, H.-k

    X.-y. Fan, H.-k. Li, K.-m. Liu, X.-l. Wang, Y. Yu, X.-y. Li, A multi-layer attentionconvolutionalneuralnetworkmodelforfineclassificationofhy- perspectral images in rare earth mining areas, SPECTROSCOPY AND 28 SPECTRAL ANALYSIS 45 (2025) 2666–2675. doi:10.3964/j.issn.1000- 0593(2025)09-2666-10

    work page doi:10.3964/j.issn.1000- 2025

  22. [22]

    Schuessler, Machine learning with nonlinear state space mod- els, AT-AUTOMATISIERUNGSTECHNIK 70 (2022) 1027–1028

    M. Schuessler, Machine learning with nonlinear state space mod- els, AT-AUTOMATISIERUNGSTECHNIK 70 (2022) 1027–1028. doi:10.1515/auto-2022-0089

    work page doi:10.1515/auto-2022-0089 2022

  23. [23]

    J. Sun, K. Chen, S. Wang, Y. Zhang, Z. Xu, X. Wu, C. Tang, Dgfe- mamba: Mamba-based 2d image segmentation network, JOURNAL OF BIONICENGINEERING22(2025)2135–2150.doi:10.1007/s42235-025- 00711-x

    work page doi:10.1007/s42235-025- 2025

  24. [24]

    doi:10.1038/s41598-025-20179-3

    C.Zhang, J.Zheng, M.Cao, Amusicsourceseparationmethodintegrat- ing time-frequency decoupling and mamba-based state space modeling, SCIENTIFIC REPORTS 15 (2025). doi:10.1038/s41598-025-20179-3

    work page doi:10.1038/s41598-025-20179-3 2025

  25. [25]

    H. Wang, P. Zhuang, X. Zhang, J. Li, Dbmgnet: A dual-branch mamba- gcn network for hyperspectral image classification, IEEE TRANS- ACTIONS ON GEOSCIENCE AND REMOTE SENSING 63 (2025). doi:10.1109/TGRS.2025.3564364

    arXiv 2025

  26. [26]

    Jiang, S

    Y. Jiang, S. Zhang, C. Wang, G. Zhang, M. Tan, B. Du, X. Shen, Sedgm: A structure-enhanced spatial-spectral dynamic gat- ing mamba for hyperspectral image classification, IEEE TRANS- ACTIONS ON GEOSCIENCE AND REMOTE SENSING 63 (2025). doi:10.1109/TGRS.2025.3626930

    arXiv 2025

  27. [27]

    Y. Li, Y. Luo, L. Zhang, Z. Wang, B. Du, Mambahsi: Spatial- spectral mamba for hyperspectral image classification, IEEE TRANS- ACTIONS ON GEOSCIENCE AND REMOTE SENSING 62 (2024). doi:10.1109/TGRS.2024.3430985

    arXiv 2024

  28. [28]

    Liang, P

    L. Liang, P. Xie, Y. Zhang, J. Li, Z. Zhang, J. Li, A. Plaza, Dbmlla: Double-branch mamba-like linear attention network for hyperspectral image classification, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 63 (2025). doi:10.1109/TGRS.2025.3605798

    arXiv 2025

  29. [29]

    F. Kong, M. Huang, J. Tang, G. Ren, End-to-end feature domain resid- ual coding network for multispectral image compression based on inter- spectral prediction, JOURNAL OF APPLIED REMOTE SENSING 18 (2024). doi:10.1117/1.JRS.18.036508. 29

    work page doi:10.1117/1.jrs.18.036508 2024

  30. [30]

    doi:10.1109/TNN.2008.2003271

    J.-M.Wu, Multilayerpottsperceptronswithlevenberg-marquardtlearn- ing, IEEE TRANSACTIONS ON NEURAL NETWORKS 19 (2008) 2032–2043. doi:10.1109/TNN.2008.2003271

    arXiv 2008

  31. [31]

    T. Kun, D. Pei-Jun, Hyperspectral remote sensing image clas- sification based on support vector machine, JOURNAL OF INFRARED AND MILLIMETER WAVES 27 (2008) 123–128. doi:10.3724/SP.J.1010.2008.00123

    Pith/arXiv arXiv 2008

  32. [32]

    Firat, D

    H. Firat, D. Hanbay, Comparison of 3d cnn based deep learning archi- tectures using hyperspectral images, JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY 38 (2023) 521–534. doi:10.17341/gazimmfd.977688

    work page doi:10.17341/gazimmfd.977688 2023

  33. [33]

    Zhong, J

    Z. Zhong, J. Li, Z. Luo, M. Chapman, Spectral-spatial residual network for hyperspectral image classification: A 3-d deep learning framework, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENS- ING 56 (2018) 847–858. doi:10.1109/TGRS.2017.2755542

    arXiv 2018

  34. [34]

    L. Sun, G. Zhao, Y. Zheng, Z. Wu, Spectralspatial feature tokeniza- tion transformer for hyperspectral image classification, IEEE TRANS- ACTIONS ON GEOSCIENCE AND REMOTE SENSING 60 (2022). doi:10.1109/TGRS.2022.3144158

    arXiv 2022

  35. [35]

    L. Sun, H. Zhang, Y. Zheng, Z. Wu, Z. Ye, H. Zhao, Massformer: Memory-augmented spectral-spatial transformer for hyperspectral im- age classification, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 62 (2024). doi:10.1109/TGRS.2024.3392264

    arXiv 2024

  36. [36]

    N. Chen, L. Fang, Y. Xia, S. Xia, H. Liu, J. Yue, Spectral query spa- tial: Revisiting the role of center pixel in transformer for hyperspectral image classification, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 62 (2024). doi:10.1109/TGRS.2024.3361652

    arXiv 2024

  37. [37]

    Huang, Y

    L. Huang, Y. Chen, X. He, Spectral-spatial mamba for hy- perspectral image classification, REMOTE SENSING 16 (2024). doi:10.3390/rs16132449. 30 Declaration of competing interest The authors declare that they have no known competing financial inter- ests or personal relationships that could have appeared to influence the work reported in this paper. Acknow...

    doi:10.3390/rs16132449 2024