acceptodds
Under review as a conference paper at ICLR 2027

RC-Mamba: A Group-Conditional Bayesian Risk-Controlled Mamba for Multi-Source Remote Sensing Classification

Abstract

Multi-source remote sensing classification integrates hyperspectral, SAR, and LiDAR observations for accurate land-cover recognition. Recent state-space models (SSMs) with multi-scale Mamba fusion have achieved competitive accuracy, yet they remain purely discriminative and lack distribution-free reliability under spectral-spatial non-stationarity, cross-modal heterogeneity, and class imbalance. To address this problem, we propose RC-Mamba, a reliability-calibrated Mamba framework for multi-source remote sensing classification. RC-Mamba couples a multi-scale Mamba backbone with three reliability modules: Group-Conditional Bayesian Risk Thresholding (G-BRT), which learns group-aware risk thresholds beyond a single global threshold; Threshold-Coupled Margin Regularization (TCMR), which injects these thresholds into discriminative training to enlarge confidence margins of critical samples; and Risk-Aware Selective Refinement (RASR), which refines only calibrated high-risk samples at test time. Experiments on Berlin, Augsburg, Houston2013, and Houston2018 show that RC-Mamba improves the reproduced MSFMamba baseline by – OA points over three seeds while exposing a group-aware reliability interface for independent-reference evaluation. The audited reliability diagnostic further shows why marginal coverage alone is insufficient and motivates explicit worst-group evaluation. These results establish RC-Mamba as a risk-aware SSM framework for multi-source remote sensing imagery.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-source remote sensing classification, State-space models for multi-source remote sensing, Distribution-free uncertainty calibration, Conformal prediction as Bayesian quadrature, Group-conditional HPD-Dirichlet conformal thresholding, Conformal margin regularization, Risk-aware selective refinement

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_crjcOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

RC‑Mamba targets multi‑source remote‑sensing classification. Built upon the MSFMamba backbone, it establishes a complete “risk measurement‑learning‑handling” pipeline via G‑BRT group‑conditional Bayesian risk thresholding, TCMR threshold‑coupled margin regularization, and RASR selective refinement. It achieves balanced and controllable prediction risk and reliability across different groups while boosting classification accuracy.

Critical strengths

This work builds a full end‑to‑end pipeline covering group‑wise risk measurement, model training constraints, and residual‑risk handling at inference time. Existing conditional conformal and conformal risk‑control methods mostly perform post‑hoc calibration after training and can only adjust thresholds for a fixed pre‑trained model. In contrast, RC‑Mamba derives group‑aware risk standards from G‑BRT, converts group‑level risk violations into differentiable loss signals through TCMR and feeds them back into the Mamba backbone. Consequently, group‑wise reliability constraints are directly embedded into representation learning. Finally, RASR applies selective test‑time augmentation exclusively to high‑risk samples, avoiding the heavy computational overhead incurred by full‑dataset TTA.

Critical weaknesses

Limited theoretical guarantees of G‑BRT inherited from CPBQ Bayesian‑quadrature G‑BRT constructs risk upper bounds using Dirichlet sampling following CPBQ. CPBQ only delivers conservative upper bounds under the Bayesian posterior perspective rather than strict frequentist finite‑sample guarantees offered by conventional conformal prediction. Many statements in the manuscript risk conflating these two paradigms and give the impression that “strict group‑wise risk control” is attained. Nevertheless, G‑BRT provides no rigorous frequentist group‑risk guarantees and merely yields Bayesian conservative estimates. The boundary of its theoretical guarantees is insufficiently clarified.

Inadequate reporting of statistical significance Though 3 random seeds are adopted in experiments, only mean values are presented in tables without standard deviations. It cannot be verified whether the OA improvements of 0.08‑0.84 points brought by RC‑Mamba over MSFMamba are statistically significant. Minor gains on some datasets may plausibly stem from random fluctuation.

Insufficient ablation studies Current ablations only examine progressive module stacking. Several critical ablations are missing: ① Ablation on group tile size: how tile dimensions affect overall accuracy (OA) and group‑wise coverage; ② Sweep ablation for TCMR hyper‑parameters (μ\mu,$\gamma$); ③ Ablation for the EMA smoothing coefficient (ρ\rho) to verify the necessity of smoothing; ④ Analysis of how reference cache‑set size influences G‑BRT threshold estimation.

Incomplete baselines for reliability evaluation Only post‑hoc baselines including Split‑CP, CRC and RCPS are compared. Training‑aware conformal‑learning approaches (e.g., Learning Optimal Conformal Classifiers) are absent, as are major baselines from conditional conformal prediction (Conditional CP).

Questions for authors
  1. Please clearly elaborate the distinction between the Bayesian posterior upper bound of G‑BRT and frequentist finite‑sample group‑risk guarantees. Are there theoretical scenarios in which the predefined group‑risk target (α\alpha) can be violated even under recommended hyper‑parameter settings of (α\alpha,β\beta)? Please explain. Should relevant statements across the manuscript be revised to prevent over‑claiming G‑BRT’s theoretical guarantees?
  2. Please supplement standard deviations across multiple random seeds for all metrics to justify the statistical significance of gains in both classification accuracy and reliability. Explicitly mark cases where performance improvements are statistically insignificant for certain datasets.
  3. Ablation regarding group tile size is missing. Tile size is a manually configured critical grouping hyper‑parameter; please analyze its impacts on OA and group‑level coverage. Hyper‑parameter sweep ablations for TCMR are required, since these hyper‑parameters directly govern regularization strength and training dynamics; please report sweep results and justify your hyper‑parameter choices. An ablation for EMA coefficient should be conducted to validate whether EMA smoothing is indispensable, by comparing threshold estimation quality and final reliability metrics with and without EMA. Please also provide an ablation for the G‑BRT reference cache‑set size and analyze how sample volume in the cache affects threshold estimation and downstream model performance.
  4. Why are training‑aware conformal methods and conditional conformal prediction not included in reliability comparisons? Please add experimental results for these baselines and demonstrate the advantages and necessity of the proposed joint‑training pipeline (G‑BRT + TCMR) against existing training‑aware / conditional conformal alternatives.
Reviewer_T1kROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the discrepancies in prediction reliability across different spatial regions or sensing domains in multi-source remote sensing classification, this paper proposes the reliability calibration framework RC-Mamba. This framework, based on the existing MSFMamba framework, introduces Group Conditional Bayesian Risk Threshold Estimation (G-BRT), Threshold Coupled Interval Regularization (TCMR), and Risk-Aware Selective Refinement (RASR), combining group-aware risk measurement, training optimization, and test-time selective enhancement. Experiments on four multi-source remote sensing datasets report a small improvement in classification performance compared to the baseline, as well as improvements in some empirical reliability metrics.

Critical strengths
  1. There are clear practical implications for focusing on differences in reliability between groups that may be masked by overall accuracy in multi-source remote sensing classification. Combining group-perceived risk measurement, training regularization, and test-time selective refinement, the overall process is clear.

  2. Experiments were carried out on four data sets covering HSI-SAR and HSI-LiDAR. In addition to classification indicators, coverage, prediction set size, coverage differences between groups and controlled distribution shift results were also reported, with a more comprehensive evaluation dimension.

Critical weaknesses
  1. The correspondence between the problem definition and the method design is insufficient. The abstract emphasizes spectral-spatial nonstationarity, cross-modal heterogeneity, and class imbalance, but the method primarily sets thresholds through spatial blocks or sensor domain grouping, without fully explaining how this grouping addresses these challenges, especially minority-class and cross-modal degradation.

  2. The object and scope of reliability guarantees are not clearly defined. The method measures the risk of the true class probability falling below a threshold, which is not equivalent to the final single-label classification error rate. The Bayesian posterior risk justification in the appendix also fails to adequately explain how it supports the distribution-independent reliability emphasized in the abstract, and whether relevant guarantees are retained after training updates, EMA threshold smoothing, global threshold backoff, and refinement at test time.

  3. The necessity of each module and its advantages over simpler solutions are not fully established. TCMR improves the true class confidence of training samples but does not automatically imply improved test reliability; furthermore, there is a lack of specific comparisons with solutions such as group condition calibration and regularization of ordinary hard samples. The three modules also do not demonstrate a clear Mamba-specific dependency, and their special value in conjunction with this backbone is not sufficiently understood.

  4. The risk identification in RASR differs from the previous two modules. G-BRT and TCMR use true class probabilities, while RASR uses maximum predicted probabilities, potentially missing high-confidence errors. The paper lacks a direct assessment of the quality of high-risk sample identification and does not adequately demonstrate that flipping enhancement effectively addresses the residual risk discussed.

  5. The ablation experiments suffer from unclear attribution. According to the method description, adding only G-BRT and disabling TCMR and RASR should not directly change classification predictions, but the accuracy changes shown in Table 2 require explanation. Furthermore, the accuracy gain of TCMR is small, and ablation does not report group reliability metrics separately, making it difficult to confirm its core contribution.

  6. The experimental results still offer limited support for some claims. The main tables do not report standard deviations or confidence intervals, making it difficult to assess the stability of small performance gains; the decrease in average class accuracy in Augsburg does not directly support a reduction in class imbalance. Existing offset experiments mainly employ class resampling and block removal, which are insufficient for verifying modal degradation and other scenarios; the efficiency advantage of selective refinement also lacks support from actual measurement time and resource overhead.

Questions for authors

Please refer to the weaknesses.

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

This paper proposes RC-Mamba, a reliability-aware framework for multi-source remote sensing classification. The proposed framework couples a multi-scale Mamba backbone with three reliability modules: Group-Conditional Bayesian Risk Thresholding (G-BRT), which estimates group-specific confidence thresholds; Threshold-Coupled Margin Regularization (TCMR), which injects these thresholds into discriminative training; and Risk-Aware Selective Refinement (RASR), which selectively refines low-confidence predictions through test-time augmentation. The reported results on four benchmarks suggest improvements in both classification performance and reliability compared with the reproduced MSFMamba baseline.

Critical strengths

1.The paper addresses an important limitation of traditional discriminative classifiers: high overall accuracy does not necessarily imply consistent predictive reliability across heterogeneous spatial regions. Rather than applying confidence calibration solely as a post-processing step, RC-Mamba establishes a practical paradigm that connects group-wise risk measurement, threshold-aware representation learning, and selective test-time refinement.

Critical weaknesses

1.According to the experimental results reported in the original paper for the baseline method MSFMamba, it outperforms MACN on the AUGSBURG dataset, however, in the current manuscript, the performance of MSFMamba is notably lower than that of MACN. It is worth noting that the experimental results for other comparative methods remain consistent across both papers. This discrepancy makes the reported performance of MSFMamba across the four datasets particularly puzzling. The authors are advised to verify these experimental results and provide further clarification regarding the reasons for this discrepancy, in order to ensure the accuracy and reproducibility of the findings.

2.Both Figure 3 in this manuscript and Figure 11 in the original MSFMamba paper display classification results for various methods on the Houston 2013 dataset, however, some of the result images appear to be identical, despite being attributed to different methods. For instance, the classification result image for FusAtNet in this manuscript is identical to the result image for TBCNN in the original MSFMamba paper. A similar phenomenon appears to exist for the classification results on other datasets presented in the appendix. The authors are advised to verify the correspondence between each result image and its respective method to ensure the accuracy and reliability of the visualizations.

Questions for authors

1.Were the experimental results for the various comparison methods presented in this paper obtained through independent replication, or were they cited directly from the relevant literature? Notably, the results reported for nearly all comparison methods align with those presented in the original MSFMamba paper, whereas the results reported for the MSFMamba method itself show significant discrepancies. The authors should provide a reasonable explanation for this phenomenon.

2.Do the classification result visualizations presented in the paper genuinely originate from the actual experimental outputs of the respective comparison methods? The authors should clarify why, for the same dataset, certain classification result visualizations in this paper are identical to those reported in the original MSFMamba paper, yet are attributed to different comparative methods.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.