Pith. sign in

REVIEW 3 major objections 5 minor 26 references

PNEC-Mamba: Prototype-Guided Positive-Negative Evidence Calibration for Hyperspectral Image Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read PNEC-Mamba, a prototype-guided positive-negative evidence calibration framework, achieves the best overall accuracy and Kappa among compared methods on three hyperspectral classification benchmarks.

desk verdict A new architecture with credible benchmark gains, but the evidence decomposition is under-supervised and the tables disagree; worth a referee, not a desk reject. read the letter →

arxiv 2608.01910 v1 pith:MBSQKUL2 submitted 2026-08-03 cs.CV

Mingzhen Xu , Can Xu , Di Wang , Haonan Guo , Bo Du This is my paper
classification cs.CV
keywords hyperspectralimageclassificationevidencereliabilityclassprototypespositive-negativeuncertaintyestimationstate-spacemodelMambaselectivecalibration
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 claims that pixel representations in hyperspectral images mix reliable evidence with contradictory cues from spectrally similar classes, mixed pixels, and surrounding context. PNEC-Mamba tries to separate these by comparing each pixel with dynamic class prototypes, deriving positive evidence that supports the true class and negative evidence that encodes competing classes, then using multi-source uncertainty to calibrate only unreliable pixels. On three benchmark datasets, it reports the highest overall accuracy and Kappa among compared methods, with full-scene inference faster than the other full-image models. The authors argue that this shows limiting accuracy is no longer feature quantity but feature purity.

What carries the argument

The core mechanism is prototype-guided positive-negative evidence calibration: a momentum-updated class prototype bank provides class references; pixel-prototype competition (scaled cosine similarity followed by softmax) yields positive context from the target class prototype and negative context from the most competitive non-target prototypes; two lightweight evidence heads (g+ and g-) produce E+ and E- from pixel features, branch features, and the contexts. The calibrated feature is F_i + G_i(alpha E+ - beta E-), where G_i is a learned gate fed by three uncertainty estimates (predictive entropy, Jensen-Shannon divergence between branch probabilities, and prototype-margin). The gate floor r

What would settle it

Train two variants, one with only positive evidence and one with only negative evidence; if the combined model does not outperform both, or if the outputs of g+ and g- are highly correlated, the evidence separation claim would collapse. Alternatively, add an explicit loss enforcing discriminativeness or orthogonality of E+ and E- and observe whether accuracy changes significantly.

Watch

Extended reading notes

Core claim

The central claim is that explicitly modeling whether pixel-level evidence supports or conflicts with classification decisions improves hyperspectral classification beyond better feature extraction. PNEC-Mamba maintains momentum-updated class prototypes as semantic anchors, computes scaled cosine similarity between pixel features and prototypes, and constructs positive evidence from the ground-truth/predicted class prototype and negative evidence from the most competitive non-target prototypes. A gate derived from predictive entropy, branch disagreement, and prototype margin controls the strength of calibration so uncertain pixels are adjusted more. A parameter-free full-resolution window av

Load-bearing premise

The method's novelty rests on the untested assumption that the positive and negative evidence heads learn genuinely complementary class-supporting and class-confusing cues rather than correlated or degenerate residuals, because only the final classification loss supervises them.

Editorial extensions

If this is right

  • On three benchmark datasets with different sensors, resolutions, and scene types, replacing direct feature-to-label mapping with evidence reliability modeling yields the best overall accuracy and Kappa among compared CNN, Transformer, and state-space baselines.
  • Uncertainty estimates localize errors: the most uncertain 20% of pixels contain 73.3-83.0% of all mistakes, and error AUROC ranges from 0.889 to 0.916.
  • Reliability-aware calibration and positive-negative evidence construction are complementary; removing either degrades accuracy on all datasets, most clearly on UP and HanChuan.
  • Full-scene inference is the fastest among compared methods (e.g., 0.0149 s on UP), because the full-image state-space encoder avoids patch-based repeated evaluation.
  • The accuracy gap among the best CNN, Transformer, and Mamba baselines is small (<1% OA), supporting the claim that feature quantity is sufficient and feature purity is the remaining limiting factor.

Reading between the lines

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

  • The positive-negative evidence decomposition resembles a contrastive learning mechanism at the prototype level; connecting it to explicit contrastive losses could strengthen the separation and test whether the heads learn complementary representations.
  • The gate floor and recovery window are hand-set per dataset; a sensitivity study or an adaptive mechanism could reveal whether selective calibration gains are robust or largely driven by tuned hyperparameters.
  • The 'feature purity' finding suggests that hybrid methods that add evidence calibration on top of any strong backbone could transfer to other pixel-wise classification tasks, such as medical imaging or remote sensing beyond hyperspectral.
  • The uncertainty localization results imply the framework could double as an error-detection module, for example in active learning where the most uncertain pixels are queried for labels.
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

3 major / 5 minor

Summary. PNEC-Mamba proposes a hyperspectral image classification framework built around pixel-level evidence reliability modeling. A full-image Mamba-based encoder produces pixel representations; dynamic class prototypes are updated by momentum; positive and negative evidence are constructed from pixel–prototype similarities via two lightweight heads; a multi-source uncertainty map (entropy, branch disagreement, prototype margin) gates a calibration step; and a full-resolution local-average refinement is applied at inference. The paper reports state-of-the-art OA and κ on Pavia University, Houston 2013, and WHU-Hi-HanChuan, with ten-run means and standard deviations, and reports the fastest full-scene inference among the compared methods.

Significance. If the evidence-separation mechanism is genuinely responsible for the gains, the paper would make a useful conceptual contribution by shifting HSI classification from representation learning toward explicit modeling of reliable versus interfering evidence. The experimental protocol is a strength: three diverse datasets, fixed partitions, ten seeded runs, standard deviations, and full-scene inference timing. However, the two load-bearing issues below—internal numeric inconsistency and lack of supervision/verification for the positive–negative decomposition—mean the central claims are not yet established as stated. The work is within the scope of the journal and likely of interest to the HSI community, but the mechanism needs stronger evidence.

major comments (3)
  1. [Comparative Experiments vs. Ablation Study (Tables 2 and 5)] The complete PNEC-Mamba model is reported with different numbers in the two tables under claimed identical settings. UP OA is 96.77 in Table 2 but 97.39 in Table 5; Houston OA is 92.18 vs. 92.47; HanChuan OA is 94.11 vs. 94.79. Since both tables describe the full model and the ablation text says 'the same experimental settings,' this is internally inconsistent and load-bearing: the central 'best OA/κ' claim rests on Table 2, while Table 5 reports even larger gains. The authors must reconcile the numbers, specify any differences in configuration, or explain why the full model differs across experiments.
  2. [Prototype-Guided Positive-Negative Evidence Construction and Training Objective (Eqs. 6–9, 19–23)] The positive and negative evidence heads E+ and E− are constructed from pixel–prototype competition, but no training objective directly supervises either head to encode class-consistent vs. competing-class content. The only losses act on the final calibrated logits (L_ce), the branch logits (L_br), the prototype similarities (L_pro), and the consistency regularizer (L_con). Therefore the claimed semantic separation is imposed by construction, not verified; the model could learn a generic gated residual in which the prototype contexts are mere conditioning inputs. The 'w/o EC' ablation removes the entire evidence-calibration branch, so its performance drop conflates the decomposition mechanism with extra parameters and a residual connection. Add an explicit evidence-level loss (e.g., aligning E+ with the true-class prototype and E− with competing prototypes) or provide diagnostics showing
  3. [Experimental Setup / Appendix: per-dataset hyperparameters] The gate floor ρ, recovery window k, and recovery strength η are hand-set per dataset (ρ=0.5/0.3/0.5, k=9/11/13, η=1.0/0.85/1.0). No sensitivity analysis is provided. Since the 'selective calibration' narrative depends on these choices, the claim of robust reliability-guided calibration is not fully supported. Report sensitivity to these values, or justify a common setting, to rule out dataset-specific tuning as the source of the reported gains.
minor comments (5)
  1. [Finding 3 (Evidence Reliability Analysis)] The statement that 'best models in CNN, Transformer, and Mamba frameworks' reach accuracies '92.87%/92.26%/92.73%' does not match Table 3: the best Mamba baseline is S2Mamba at 92.26%, not 92.73%; 92.73% is the Base OA in Table 5. Correct the numbers or rephrase the finding.
  2. [Eq. (22)] The consistency loss uses U^fuse_i, but only eG_i is defined in Eq. (14). Clarify whether U^fuse_i = eG_i or how the fused uncertainty score is obtained.
  3. [Ablation Study (Table 5)] The variants 'w/o EC' and 'w/o RC' are not precisely specified. Does 'w/o EC' remove only the evidence heads or the entire calibration branch? Does 'w/o RC' remove the gate and keep an unweighted residual adjustment? State the exact architecture of each variant so the ablations are interpretable.
  4. [Table 2] On UP, PNEC-Mamba's AA (95.84) is lower than MambaMoE's AA (96.73). Since the paper claims the best OA and κ, this is not a contradiction, but the text should explicitly note that AA is not best on UP to avoid overclaiming 'superior classification performance.'
  5. [General] Minor language issues: 'Incorrectly pixels' in Finding 1; 'the model should identify' in the Introduction. Also, no statistical significance tests are reported; given the small standard deviations, a paired test would strengthen the comparison claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported accuracies are external benchmark measurements and no constructed quantity is predicted from its own fitted inputs.

full rationale

The paper's derivation chain is a supervised learning pipeline, not a chain of equations whose outputs are identical to their inputs. The calibrated feature (Eq. 15) is a gated residual update F_i + G_i(αE+_i − βE−_i); E+ and E− are learned functions of the pixel feature and prototype contexts (Eqs. 8–9), and the only supervision on them is indirect through the final classification loss (Eq. 19) and consistency regularizer (Eq. 22). No reported quantity (OA, AA, κ) is a fitted parameter renamed as a prediction: accuracies are computed on fixed external test splits (UP, Houston, HanChuan), and the method is compared against independent baselines. The self-citations present (Zhu et al. 2026 includes author Xu, M.; MambaMoE includes author Wang, D.; MambaHSI includes author Du, B.) appear as related work and baselines, not as load-bearing evidence for the claimed SOTA; no uniqueness theorem or ansatz is imported from prior work by the same authors. The skeptical concern that E+/E− separation is not directly supervised is a legitimate experimental-design criticism (whether the reported gains come from the evidence semantics or from a generic gated residual), but it is not circularity: the model is not defined in terms of the result it is claimed to predict, and the benchmark numbers are external measurements. Finding 1 (errors have higher estimated uncertainty) is an empirical correlation, not a tautology, because the uncertainty maps are not trained against the error labels. Therefore no specific circular step can be quoted, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that prototype competition yields a meaningful positive/negative decomposition (ad hoc to the paper), that fused uncertainty correlates with actual error (validated only on the three datasets), that momentum prototypes are stable at 15 to 30 samples per class, and that upsampling plus local averaging recovers spatial detail. The per-dataset gate floors, recovery windows, and recovery strengths are hand-set per benchmark. The paper introduces no physical entities; positive/negative evidence and dynamic prototypes are learned feature-space constructs.

free parameters (5)
  • evidence scaling coefficients alpha, beta = not reported
    Introduced in Eq 10 and 15 as 'non-negative scaling coefficients'; the paper never states whether they are learned or tuned, nor their values, yet they directly set the strength of positive/negative calibration.
  • similarity scale tau = learnable, value not reported
    Eq 4 defines scaled cosine similarity with learnable tau > 0; no value or bound is reported, though it controls the sharpness of the prototype competition.
  • prototype momentum m = not reported
    Eq 3 updates prototypes with momentum m; the value is never given, although it controls how fast semantic references track the evolving feature space.
  • per-dataset gate floor rho = 0.5 (UP), 0.3 (Houston), 0.5 (HanChuan)
    Appendix: gate floors are set separately per dataset, i.e., hand-tuned to each benchmark; rho directly sets the minimum calibration applied to every pixel.
  • per-dataset recovery window k and strength eta = k = 9/11/13; eta = 1.0/0.85/1.0
    Appendix: local averaging window and recovery strength are chosen per dataset for the full-resolution consistency refinement, directly affecting boundary smoothing in the final predictions.
assumptions (4)
  • ad hoc to paper Pixel-prototype competition decomposes representations into class-consistent positive evidence and competing-class negative evidence.
    Eqs 6 to 10 construct E+ and E- from prototype contexts, but no loss term supervises the semantic content of E+ or E-; the decomposition is imposed by construction and is the paper's core novelty.
  • domain assumption The fused uncertainty score (entropy, branch disagreement, prototype margin) indicates which pixels need calibration.
    Eq 22 uses the fused uncertainty to weight a consistency loss; Table 6 shows correlation with errors on the three tested datasets, but the transfer to any new scene is assumed.
  • domain assumption With only 15 to 30 labeled pixels per class, momentum-updated prototypes are stable semantic anchors.
    Eq 3 averages over labeled pixels each iteration; at 15 samples per class the prototype estimates are noisy, yet the method relies on them for evidence construction and uncertainty estimation.
  • domain assumption Full-image state-space encoding preserves enough spatial detail for boundary recovery after downsampling and upsampling.
    The encoder applies convolution and pooling to reduce sequence length, and Eqs 16 to 18 assume bilinear upsampling plus local averaging recovers fine spatial structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PNEC-Mamba: Prototype-Guided Positive-Negative Evidence Calibration for Hyperspectral Image Classification." pith.science (2026). https://pith.science/paper/MBSQKUL2

@misc{pith2026260801910,
  author       = {Pith},
  title        = {Pith review of: PNEC-Mamba: Prototype-Guided Positive-Negative Evidence Calibration for Hyperspectral Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBSQKUL2}},
  note         = {Machine review of arXiv:2608.01910}
}
read the original abstract

In real-world hyperspectral scenes, pixel representations are often ambiguous due to factors such as spectral similarity, mixed pixels, and local context interference, which may simultaneously encode discriminative evidence and interfering information. Existing methods mainly focus on learning more powerful representations or modeling broader contexts, but rarely investigate whether the learned representations provide reliable evidence or introduce interference into classification decisions. To address this issue, we view hyperspectral image classification from the perspective of pixel-level evidence reliability modeling and propose PNEC-Mamba, a prototype-guided positive-negative evidence calibration framework. The framework progressively establishes semantic references, separates class-related evidence from interference, estimates pixel-level reliability, and performs selective calibration. First, a full-image state-space encoder extracts pixel representations, while dynamic class prototypes provide semantic references that evolve jointly with the feature space. Subsequently, positive and negative evidence is derived from pixel-prototype competition, explicitly separating discriminative cues that support classification from confusing signals associated with competing classes. Based on these evidence relationships, a multi-source uncertainty estimation strategy is introduced to assess pixel-level reliability, enabling stronger evidence calibration for uncertain regions. Finally, a full-resolution consistency refinement step is applied to recover local spatial details and improve boundary coherence in the final predictions. Extensive experiments on three benchmark datasets demonstrate that PNEC-Mamba achieves superior classification performance compared with state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2608.01910 by the authors.

Figure 1
Figure 1. Motivation for evidence reliability modeling. Spec [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of PNEC-Mamba. The framework sequentially performs semantic-anchored encoding, evidence [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Classification maps on WHU-Hi-HanChuan: (a) ground truth, (b) PNEC-Mamba. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Spatial visualization of high-uncertainty errors and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Classification maps on Pavia University: (a) ground truth, (b) DBMA, (c) DBDA, (d) GAHT, (e) MASSFormer, (f) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Classification maps on Houston: (a) ground truth, (b) DBMA, (c) DBDA, (d) GAHT, (e) MASSFormer, (f) S2Mamba, [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Classification maps on WHU-Hi-HanChuan: (a) ground truth, (b) DBMA, (c) DBDA, (d) GAHT, (e) MASSFormer, [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 2 linked inside Pith

  1. [1]

    M.; Plaza, A.; Camps-Valls, G.; Scheunders, P.; Nasrabadi, N.; and Chanussot, J

    Bioucas-Dias, J. M.; Plaza, A.; Camps-Valls, G.; Scheunders, P.; Nasrabadi, N.; and Chanussot, J. 2013. Hyperspectral Remote Sensing Data Analysis and Future Challenges. IEEE Geoscience and Remote Sensing Magazine, 1(2): 6--36

    2013

  2. [2]

    Chu, Y.; Li, P.; Peng, J.; Ding, W.; and Li, P. 2026 a . Fuzzy Graph Mamba Fusion Network with KAN for Hyperspectral Image Classification. Information Fusion, 133: 104333

    2026

  3. [3]

    Chu, Y.; Yan, C.; Peng, J.; Ding, W.; and Li, P. 2026 b . Fuzzy Hyperbolic Hypergraph Convolutional Fusion Network for Hyperspectral Image Classification under Limited Samples. Information Fusion, 135: 104375

    2026

  4. [4]

    Debes, C.; Merentitis, A.; Heremans, R.; Hahn, J.; Frangiadakis, N.; van Kasteren, T.; Liao, W.; Bellens, R.; Pizurica, A.; Gautama, S.; Philips, W.; Prasad, S.; Du, Q.; and Pacifici, F. 2014. Hyperspectral and LiDAR Data Fusion: Outcome of the 2013 GRSS Data Fusion Contest. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,...

    2014

  5. [5]

    Gu, A.; and Dao, T. 2023. Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv:2312.00752

    arXiv 2023

  6. [6]

    Guo, Y.; Fan, B.; Zhang, Z.; Feng, Y.; Dai, Y.; and He, M. 2026. PCRNet : Progressive Correlation Refinement for Few-Shot Hyperspectral Image Classification. IEEE Transactions on Geoscience and Remote Sensing, 64: 1--16

    2026

  7. [7]

    Hong, D.; Han, Z.; Yao, J.; Gao, L.; Zhang, B.; Plaza, A.; and Chanussot, J. 2022. SpectralFormer : Rethinking Hyperspectral Image Classification with Transformers. IEEE Transactions on Geoscience and Remote Sensing, 60: 1--15

    2022

  8. [8]

    Hu, W.; Huang, Y.; Wei, L.; Zhang, F.; and Li, H. 2015. Deep Convolutional Neural Networks for Hyperspectral Image Classification. Journal of Sensors, 2015: 258619

    2015

Show all 26 references
  1. [9]

    Huang, L.; Chen, Y.; and He, X. 2024. Spectral-Spatial Mamba for Hyperspectral Image Classification. Remote Sensing, 16(13): 2449

    2024

  2. [10]

    Li, R.; Zheng, S.; Duan, C.; Yang, Y.; and Wang, X. 2020. Classification of Hyperspectral Image Based on Double-Branch Dual-Attention Mechanism Network. Remote Sensing, 12(3): 582

    2020

  3. [11]

    Li, S.; Song, W.; Fang, L.; Chen, Y.; Ghamisi, P.; and Benediktsson, J. A. 2019. Deep Learning for Hyperspectral Image Classification: An Overview. IEEE Transactions on Geoscience and Remote Sensing, 57(9): 6690--6709

    2019

  4. [12]

    Li, Y.; Luo, Y.; Zhang, L.; Wang, Z.; and Du, B. 2024. MambaHSI : Spatial-Spectral Mamba for Hyperspectral Image Classification. IEEE Transactions on Geoscience and Remote Sensing, 62: 1--16

    2024

  5. [13]

    Li, Z.; Li, Y.; and Wang, N. 2026. DiffMamba : Gated Differential Spectral-Spatial Mamba for Few-Shot Hyperspectral Image Classification. Computational Intelligence, 42(3): e70239

    2026

  6. [14]

    Ma, W.; Yang, Q.; Wu, Y.; Zhao, W.; and Zhang, X. 2019. Double-Branch Multi-Attention Mechanism Network for Hyperspectral Image Classification. Remote Sensing, 11(11): 1307

    2019

  7. [15]

    Mei, S.; Song, C.; Ma, M.; and Xu, F. 2022. Hyperspectral Image Classification Using Group-Aware Hierarchical Transformer. IEEE Transactions on Geoscience and Remote Sensing, 60: 1--14

    2022

  8. [16]

    Pan, Z.; Li, C.; Plaza, A.; Chanussot, J.; and Hong, D. 2025. Hyperspectral Image Classification with Mamba. IEEE Transactions on Geoscience and Remote Sensing, 63: 1--14

    2025

  9. [17]

    K.; Krishna, G.; Dubey, S

    Roy, S. K.; Krishna, G.; Dubey, S. R.; and Chaudhuri, B. B. 2020. HybridSN : Exploring 3-D--2-D CNN Feature Hierarchy for Hyperspectral Image Classification. IEEE Geoscience and Remote Sensing Letters, 17(2): 277--281

    2020

  10. [18]

    Snell, J.; Swersky, K.; and Zemel, R. S. 2017. Prototypical Networks for Few-Shot Learning. In Advances in Neural Information Processing Systems, volume 30

    2017

  11. [19]

    Song, Y.; Zhang, J.; Liu, Z.; Xu, Y.; Quan, S.; Sun, L.; Bi, J.; and Wang, X. 2025. Deep Learning for Hyperspectral Image Classification: A Comprehensive Review and Future Predictions. Information Fusion, 123: 103285

    2025

  12. [20]

    Sun, L.; Zhang, H.; Zheng, Y.; Wu, Z.; Ye, Z.; and Zhao, H. 2024. MASSFormer : Memory-Augmented Spectral-Spatial Transformer for Hyperspectral Image Classification. IEEE Transactions on Geoscience and Remote Sensing, 62: 1--15

    2024

  13. [21]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention Is All You Need. In Advances in Neural Information Processing Systems, volume 30

    2017

  14. [22]

    Wang, G.; Zhang, X.; Peng, Z.; Zhang, T.; and Jiao, L. 2025. S2Mamba : A Spatial-Spectral State Space Model for Hyperspectral Image Classification. IEEE Transactions on Geoscience and Remote Sensing, 63: 1--13

    2025

  15. [23]

    Xu, Y.; Wang, D.; Jiao, H.; Zhang, L.; and Zhang, L. 2026. MambaMoE : Mixture-of-Spectral-Spatial-Experts State Space Model for Hyperspectral Image Classification. Information Fusion, 127: 103811

    2026

  16. [24]

    Yao, J.; Hong, D.; Li, C.; and Chanussot, J. 2024. SpectralMamba : Efficient Mamba for Hyperspectral Image Classification. arXiv:2404.08489

    2024 arXiv

  17. [25]

    Zhong, Y.; Hu, X.; Luo, C.; Wang, X.; Zhao, J.; and Zhang, L. 2020. WHU-Hi : UAV -Borne Hyperspectral with High Spatial Resolution ( H2 ) Benchmark Datasets and Classifier for Precise Crop Identification Based on Deep Convolutional Neural Network with CRF . Remote Sensing of E...

    2020

  18. [26]

    Zhu, Q.; Xu, M.; Ma, R.; Ran, L.; Xue, J.; and Guan, Q. 2026. Knowledge-Data-Model-Driven Multimodal Few-Shot Learning for Hyperspectral Fine Classification: Generalization Across Sensor, Category and Scene. ISPRS Journal of Photogrammetry and Remote Sensing, 233: 623--650

    2026

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.