acceptodds
Withdrawn from ICLR 2027

KGLP-Net: Kalman-Reliable Global-Local Learning With Group Fisher Pruning for Hyperspectral Image Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing global-local HSI classifiers rarely account for the relative uncertainty of complementary information sources, while model compression typically ignores the uncertainty estimated during fusion. To address this limitation, we propose **Kalman-Reliable Global–Local Learning with Group Fisher Pruning (KGLP-Net)**, which couples uncertainty-aware fusion with reliability-guided structured pruning. KGLP-Net employs an asymmetric dual-view architecture, where a Global Mamba Encoder captures long-range contextual dependencies and a Local CNN Encoder preserves fine-grained target-neighborhood information, providing complementary representations with distinct uncertainty characteristics. **Kalman-Reliable Fusion (KRF)** treats the global and local representations as a prior and an observation, respectively, and performs an uncertainty-guided, Kalman-inspired posterior update to jointly estimate a fused posterior representation and its uncertainty. **Kalman-Reliable Group Fisher Pruning (KRGF)** further transfers the posterior uncertainty into group-level reliability via normalized posterior precision and integrates it with Group Fisher importance, allowing the uncertainty to directly inform structured pruning. On all four public HSI datasets, Compact48, the KRGF-compressed model with hidden width 48, achieves the highest OA among ten representative competing classifiers. Compared with Dense64, it reduces parameters, FLOPs, and inference time by 38.5%, 34.8%, and 37.6% on average, respectively, while incurring only a 0.03 percentage point decrease in mean OA.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperspectral Image Classification, Global–Local Representation Learning, Uncertainty-Aware Fusion, Reliability-Guided Structured Pruning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fTUgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Overall, I find the paper technically solid and clearly presented. The motivation is easy to follow, and the interaction among the global–local representation learning, Kalman-inspired fusion, and reliability-guided pruning is thoughtfully designed. In particular, the paper establishes a coherent connection between uncertainty-aware feature fusion and structured compression, and the empirical results show a favorable accuracy–efficiency trade-off across the evaluated HSI classification benchmarks.

My main concern is therefore not the technical correctness of the proposed design, but rather the broader significance and methodological novelty of the contribution. Hyperspectral image classification is a relatively mature and specialized task, and although the proposed modules are combined in an elegant way, I am not fully convinced that the paper introduces sufficiently new ideas that would have substantial impact beyond this particular problem setting. The current results demonstrate that the proposed framework is effective, but stronger evidence of broader applicability or more general methodological insight would make the contribution considerably more compelling.

In particular, I would encourage the authors to further clarify what aspects of the proposed uncertainty-guided fusion and pruning framework are expected to generalize beyond the current KGLP-Net architecture and HSI classification setting. For example, demonstrating the reliability-guided pruning principle on additional architectures or tasks would help support the broader claim that uncertainty-derived reliability can serve as a general principle for structured model compression.

Overall, I think the paper contains several well-designed components and strong empirical results, but I am not yet convinced that its methodological contribution and broader impact are sufficiently substantial for acceptance at ICLR.

Critical strengths
  1. Clear motivation and coherent overall design.
    The paper provides a clear motivation for jointly considering global–local complementarity, uncertainty-aware fusion, and model compression. The three major components—global/local representation learning, Kalman-Reliable Fusion (KRF), and Kalman-Reliable Group Fisher Pruning (KRGF)—are connected in a logically consistent manner rather than being introduced as isolated modules.

  2. Thoughtful coupling between fusion and pruning.
    A particularly interesting aspect is that the uncertainty estimated during feature fusion is not discarded after representation learning. Instead, the posterior uncertainty is converted into a reliability signal and further incorporated into Group Fisher-based structured pruning. This creates a meaningful connection between representation reliability and model compression.

  3. Strong accuracy–efficiency trade-off.
    Compact48 achieves the best overall accuracy among the compared methods on all four evaluated datasets, while substantially reducing parameters, FLOPs, and inference time relative to Dense64. On average, the model reduces parameters by 38.5%, FLOPs by 34.8%, and inference time by 37.6%, with only a 0.03 percentage-point decrease in mean OA.

  4. Reasonably comprehensive ablation studies.
    The paper separately evaluates the contributions of the asymmetric global–local representation, KRF, and KRGF. In particular, the experiments comparing Sum, Concat, Gate, and KRF, as well as Group Fisher, posterior reliability, and their combination, help demonstrate that the proposed components provide complementary benefits.

  5. Good experimental coverage within the HSI classification setting.
    The evaluation includes four datasets with different scene characteristics, ten independent runs, class-wise results, qualitative classification maps, patch-size sensitivity, and compression trajectories. This provides relatively solid empirical evidence that the reported improvements are not limited to a single dataset or isolated configuration

Critical weaknesses
  1. The paper does not provide sufficiently convincing insight into why the three major components are particularly well suited to hyperspectral data.
    The proposed global–local representation learning, Kalman-Reliable Fusion, and reliability-guided pruning are individually reasonable, but their connection to the unique characteristics of HSI data is not analyzed in sufficient depth. Beyond explicitly modeling spectral information in the feature-extraction stage, it is not clear which parts of the proposed framework fundamentally rely on hyperspectral properties. In other words, if the same design were applied to RGB (maybe other modality) classification, would it work equally well? I would like to see a clearer discussion or additional experiments showing what makes these components particularly appropriate for hyperspectral imagery rather than being generally applicable architectural choices.

  2. The methodological contribution is relatively incremental for a mature task.
    Hyperspectral image classification is already a well-established research problem, and I do not think this work introduces a sufficiently novel perspective or a clear paradigm shift to the task. Rather than bringing a substantially new concept from the broader ML/CV community into HSI classification, the paper mainly advances performance through improved global–local modeling, uncertainty-aware fusion, and coordinated structured pruning. The integration is thoughtful and effective, but the overall contribution remains evolutionary rather than fundamental.

    From my personal perspective, this type of work may be a more natural fit for flagship domain-oriented journals such as TGRS or TIP. Nevertheless, I do recognize the technical contribution and the care taken in integrating the proposed components.

  3. The paper emphasizes improving the state of the art more than establishing a broadly useful design principle.
    Personally, I would find the work more compelling if the authors systematically demonstrated that a specific weakness has been consistently overlooked by prior HSI classification methods, and that the proposed modification—even if relatively small—can reliably improve a range of existing approaches. Such a study could provide a clearer takeaway for practitioners and reduce the development burden for the broader community.

    In its current form, the paper mainly presents a new end-to-end framework and demonstrates that it performs well. I would prefer stronger evidence that the proposed ideas can serve as reusable principles beyond KGLP-Net itself, especially since the paper also suggests that uncertainty-derived reliability may represent a more general principle for lightweight neural network design.

Questions for authors

Please provide a strong and detailed rebuttal addressing the weaknesses raised above.

Reviewer_r2G9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KGLP-Net, a global-local feature fusion and structured pruning framework for hyperspectral image classification. The authors argue that global information provides extensive spatial context but is prone to introducing irrelevant background; local information focuses on details near target pixels but may lack sufficient context. Therefore, this paper aims to dynamically adjust the fusion weights based on the uncertainty of both types of information and further utilize the fused reliability for model compression.

The method consists of three parts. First, the Global Mamba Encoder processes larger image patches, extracting long-range spatial and spectral information; the Local CNN Encoder processes smaller, central image patches, extracting fine-grained features near the target. Then, Kalman-Reliable Fusion (KRF) predicts the uncertainties of the two paths based on the differences and interactions between global and local features, and performs dynamic fusion through a Kalman-like update, while simultaneously calculating the posterior uncertainty of the fused features. Finally, Kalman-Reliable Group Fisher Pruning (KRGF) transforms the posterior uncertainty into channel reliability and combines it with Group Fisher importance to select channels to retain and reconstruct a compact model.

The training process consists of three stages: first, a full model with a hidden width of 64 is trained; then, based on the KRGF score, the model is progressively compressed to a width of 48, 40, or 32; finally, the compressed model is fine-tuned.

The overall logic is as follows:

Global and local feature extraction

→ Relative uncertainty estimation

→ Kalman-style dynamic fusion

→ Posterior reliability calculation

→ Channel pruning combined with Group Fisher

→ Obtaining a lightweight classification model

The core feature of this method is linking "uncertainty-aware feature fusion" and "reliability-guided model pruning." However, the paper mainly demonstrates the role of these uncertainties indirectly through classification and pruning performance, and has not yet fully verified whether they are truly reliable and calibrated uncertainties.

Critical strengths

The research problem is significant. Hyperspectral classification models need to utilize both a wide range of context and central neighborhood details, while also considering model efficiency in resource-constrained scenarios.

The framework logic is relatively complete. This paper organizes global Mamba, local CNN, dynamic fusion, and structured pruning into a continuous process, rather than simply splicing together multiple unrelated modules.

The global and local branches are complementary. Table 2 shows that Local-only and Global-only perform differently on different datasets, while Dual-view achieves higher average OA on all four datasets.

The paper also reports the accuracy, number of parameters, FLOPs, and inference time for Dense64, Compact48, Compact40, and Compact32, allowing observation of performance changes under different compression levels.

The experiments include four publicly available hyperspectral datasets, and the mean and standard deviation are reported through ten replicate experiments, providing a more comprehensive validation range than experiments on a single dataset.

Critical weaknesses
  1. The paper directly interprets the positive tensors Pg\mathbf{P}_g and Rl\mathbf{R}_l output by the network as uncertainties between global priors and local observations, and further interprets the reciprocal of P+\mathbf{P}^{+} as posterior accuracy and reliability. However, these quantities are primarily learned indirectly through the classification objective, and the paper fails to provide direct evidence for the following:
  • The correlation between uncertainty and sample error rate, prediction confidence, or actual feature perturbation;

  • Whether uncertainty increases accordingly with increasing noise, occlusion, or distribution shift;

  • Whether the reliability scores are calibrable or can be used to distinguish between correct and incorrect predictions;

  • Whether interpreting these quantities as “uncertainty” provides additional information compared to ordinary dynamic gating scores.

Therefore, the existing experiments demonstrate that these learned weights are helpful for fusion and pruning, but are insufficient to prove that they have a rigorous or calibrated statistical uncertainty meaning. It is recommended that the authors supplement with experiments on controlled perturbations, error detection, risk-coverage curves, or reliability correlation, and clarify in the description whether the method involves rigorous probabilistic modeling or a deterministic gating mechanism inspired by Kalman forms.

  1. Therefore, the existing experiments demonstrate that these learned weights are helpful for fusion and pruning, but are insufficient to prove that they have a rigorous or calibrated statistical uncertainty meaning. It is recommended that the authors supplement with experiments on controlled perturbations, error detection, risk-coverage curves, or reliability correlation, and clarify in the description whether the method involves rigorous probabilistic modeling or a deterministic gating mechanism inspired by Kalman forms.

  2. KRF is only inspired by the Kalman update idea and does not strictly belong to Kalman filtering. Meanwhile, the scoring function Qc=IcρcQ_c=I_c\rho_c is primarily an empirical multiplicative combination, and its necessity and optimality have not yet been proven through theoretical derivation or comparison of different combination methods.

  3. The paper has not directly proven that the learned uncertainty accurately reflects the true reliability of global features, local features, and channel information. Therefore, existing experiments mainly demonstrate the effectiveness of the method in terms of performance, but are insufficient to fully prove the validity of the proposed "reliability-driven fusion and pruning" mechanism.

Questions for authors
  1. The hidden width uses DD, and Dropout uses a similar D\mathcal{D}. This notation is not incorrect, but its readability is poor. It is recommended to change it to Dropout⁡(⋅)\operatorname{Dropout}(\cdot) and report the specific dropout probability.

  2. The paper randomly divides training, validation, and test pixels within the same hyperspectral image, using a maximum 25×2525\times25 center image block as input. The input blocks corresponding to adjacent center pixels will highly overlap. For example, two horizontally adjacent 25×2525\times25 image patches have a region overlap of 96%. Therefore, although the labels on the training and test sets are mutually exclusive, their input contexts may be highly similar. It is suggested that the authors supplement with spatially mutually exclusive partitioning, buffer partitioning, or cross-region testing to verify whether the method remains effective under conditions of low spatial correlation.

Reviewer_YNF6OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces KGLP-Net. First, it addresses an issue in the global-local dual-branch complementary design for hyperspectral image classification, where the confidence modeling of the two branches (like addition or concatenation) during the fusion stage is incomplete, and proposes uncertainty-aware global-local representation learning. To boost the model's efficiency, this article further proposes a confidence-guided pruning strategy. Experiments on multiple datasets have shown that KGLP-Net is effective and efficient.

Critical strengths
  1. KGLP-Net achieved the best classification results on multiple datasets with fewer parameters and computations.
Critical weaknesses
  1. The main point of this paper is modeling the uncertainty of the two branches, but there's neither an explicit loss function guiding the uncertainty learning nor visual evidence showing that the learned uncertainty actually helps the two streams complement each other better.
  2. This paper didn’t conduct cross-scenario generalization experiments, so the metrics might be a bit optimistic.
Questions for authors
  1. Can authors provide some visualization about uncertainty in two branch? This might help us better understand the role of uncertainty modeling.
  2. Can authors provide some cross-scenario classification experiments ? (For example, in WHU-OHS)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.