acceptodds
Under review as a conference paper at ICLR 2027

FoveaMamba: Structure-Conditioned State Evolution for Remote Sensing Image Super-Resolution

Abstract

Remote-sensing image super-resolution requires recovering fine geometric details while modeling long-range spatial dependencies. An HR-reference-based structure-stratified analysis reveals substantial SSIM degradation from low- to high-structure regions across representative SR methods. We propose FoveaMamba, a structure-conditioned state-space network that incorporates multi-scale two-dimensional context into selective state evolution while retaining fixed directional traversal. Specifically, its Context-Guided State Space Module (CGSSM) uses the extracted context to modulate the parameterization of the discretization step . A stabilized high-frequency residual pathway provides complementary local detail compensation alongside long-range state-space propagation. Under a unified AID-training protocol, FoveaMamba achieves the highest PSNR on all six evaluated remote-sensing datasets and the highest SSIM on five. Structure-stratified comparisons indicate larger paired SSIM gains over selected baselines in high-structure regions, while internal mechanism analyses reveal greater network-averaged discretization-step modulation in these regions. Following DIV2K fine-tuning, FoveaMamba also attains the highest PSNR on four of the five evaluated natural-image benchmarks. Source code and trained models will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
remote sensing image super-resolution, state space models, Mamba, structure-aware learning, structure-conditioned state evolution

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

This paper presents FoveaMamba, a state-space network for remote sensing image super-resolution that aims to improve reconstruction in structurally complex regions. The motivation comes from a Sobel-based analysis showing that reconstruction quality varies with local structural complexity. The proposed architecture combines Pyramid Context Dilation (PCD) for multi-scale context extraction with a Context-Guided State Space Module (CGSSM), which uses this context to modulate Mamba’s discretization step. A high-frequency compensation module, MHI-Stable, and a Laplacian consistency loss are also introduced to support detail recovery.

Critical strengths
  1. Rather than treating reconstruction errors as spatially uniform, the paper examines performance across regions grouped by Sobel gradient energy. This analysis provides a useful starting point for studying structure-dependent reconstruction quality.
  2. The authors compare against CNN-, Transformer-, and Mamba-based methods on multiple remote sensing datasets. The structure-stratified evaluation and cross-domain adaptation experiments provide additional perspectives beyond aggregate PSNR and SSIM.
  3. The analysis of discretization steps, structure-stratified statistics, and bootstrap confidence intervals is useful. Although these results do not yet establish the claimed mechanism, the effort to inspect internal behavior is a positive aspect of the work.
Critical weaknesses
  1. Much of the architecture builds on familiar components, including multi-scale convolutions, gating, and high-frequency compensation. Context-dependent modulation of the discretization step appears to be the main contribution, but the paper does not adequately compare it with simpler alternatives, such as feature gating or modulation of other state-space parameters. It is therefore difficult to determine whether the proposed design offers a specific advantage over more conventional conditioning strategies.
  2. In Table 2, introducing CGSSM reduces PSNR, and subsequently adding MHI-Stable lowers both PSNR and SSIM. The full model performs best only after the Laplacian consistency loss is included. This leaves open whether the improvement comes primarily from the architecture, the loss, or their interaction. Comparisons under identical loss settings are needed, particularly for MHI-Stable. The choice of a Laplacian loss weight of 0.1 also requires justification and a sensitivity analysis.
  3. Although the paper includes several architectural families, the recent comparisons rely heavily on LSM, which targets general image super-resolution. Relevant diffusion-based remote sensing SR methods, such as EDiffSR, FastDiffSR, and DLMSR, are not included. Where their degradation settings and evaluation protocols are compatible, such comparisons would help position the method more convincingly. Any incompatibilities should be explained.
  4. Relative to LSM, the reported PSNR gains are only 0.01–0.04 dB, while SSIM is lower on some datasets. Meanwhile, the parameter count increases from 12.919M to 27.696M. Without repeated runs or uncertainty estimates for these differences, it is unclear whether the gains are reproducible. Inference time, FLOPs, and peak memory would also help assess whether the additional cost is justified.
  5. The description of how context gates are mapped to the four scanning directions is not sufficiently clear. If the gates are spatially indexed, their correspondence with each scan ordering needs to be specified; broadcasting alone does not establish correct alignment. The reported negative local Sobel–modulation correlations also need interpretation in relation to the proposed mechanism. Correlation statistics by themselves do not show that the modulation improves reconstruction. Controls using constant or spatially shuffled gates would provide a more direct test.
Questions for authors
  1. Can the authors compare the baseline and proposed modules under matched loss settings, including a baseline trained with the Laplacian consistency loss? This would help separate the architectural contribution from the effect of the training objective.
  2. Are the 0.01–0.04 dB gains over LSM consistent across independent training runs? What are the corresponding inference-time and memory costs, and is there a smaller FoveaMamba variant that offers a more favorable trade-off?
Reviewer_WBPDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes FoveaMamba, a Mamba-based architecture for remote sensing image super-resolution. The authors introduce a Context-Guided State Space Module (CGSSM) that incorporates multi-scale spatial context into the discretization-step parameterization of selective state-space models. Specifically, Pyramid Context Dilation (PCD) extracts multi-scale features, which are used to modulate the input-dependent discretization step, and the MHI-Stable branch provides high-frequency compensation.

Critical strengths
  1. The paper is well written and easy to follow. The motivation and overall framework are clearly presented, and the proposed modules are explained in detail with figures and mathematical formulations.
  2. The paper provides extensive experimental evaluations against a diverse range of competitive baselines across multiple datasets.
Critical weaknesses
  1. The novelty of the paper is limited. The context modulation in mamba architecture is not very novel, the only difference in this paper is to change the position of modulation without deep analysis of reasons. The MHI-Stable module can be regarded as a variant of residual connection, which is a relatively common design in neural network.
  2. The improvement of results is marginal. The performance gains over existing state-of-the-art methods are relatively little (differing only in the last few decimal places), raising concerns about the practical significance of the proposed approach. And we also cannot find significant differences in visualization results.
Questions for authors

1.In Eq. (5), the context gate is flattened in raster order and shared across the four directional scan sequences without direction-specific reordering. The spatial correspondence mismatches between the context gate and the features processed by the vertical and reversed scans. Could you please explain it? 2.The current ablation study does not sufficiently isolate the individual contributions of CGSSM and MHI-Stable. In particular, adding MHI-Stable to CGSSM under pixel-only supervision decreases both PSNR and SSIM. Could the authors provide a CGSSM+Edge configuration to establish whether MHI-Stable provides an independent improvement under matched supervision? 3.In Section 3.2, the authors introduce multiplicative modulation of the pre-Softplus discretization-step parameter (dtd_t) using multi-scale spatial context. Could the authors provide a stronger theoretical or empirical justification for choosing multiplicative modulation rather than additive conditioning? Furthermore, why is the conditioning applied exclusively to the discretization step (Δ\Delta), rather than to other input-dependent SSM parameters such as (BtB_t) and (CtC_t)? Ablation studies comparing these alternative conditioning strategies would help substantiate the proposed design choices.

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

FoveaMamba is proposed for the remote sensing image super-resolution. So for a selective state-space model the multi-scale spatial context modulates the discretization step, while the parallell high-frequency compensation branch is then supporting the local detail recovery. Uses pixel and laplacian consistency losses, the model states improved reconstruction across 6 different remote-sensing datasets and competitive after natural-image finetuning.

Critical strengths

First strength: You have done thorough cross-dataset evaluation. A single AID-trained checkpoint achieves the highest PSNR across the different remote-sensing datasets, including five of where they are evaluated without finetuning, supporting generalization. Second strength: The analysis is structure focused. The shared HR-defined partitions and the paired statistical comparisons examine the complex-regions performance, where we see useful evidence beyond the dataset average scores.

Critical weaknesses

First weakness: the significance of the gains are not so certain, as the reported improvements are barely over the strongest competitor. No seed variability, so that makes the reliability lower. Second weakness: There is not enough validation of the proposed mechanisms as the gate for the context is shared across the scan directions and without corresponding reordering, requiring justification for the spatial correspondence.

Questions for authors
  1. Are the small gains reproducible across seeds?
  2. Why is the context gate broadcast across different scan orders without reordering, and how the directionally aligned gating impact performance?
  3. Can you compare the CGSSM-Edge against the compelte model to isolate MHI's contribution over equal supervision?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.