acceptodds
Under review as a conference paper at ICLR 2027

Morphology-Guided Sparse State-Space Routing for SAR Image Restoration

Abstract

Radio-frequency interference corrupts synthetic aperture radar (SAR) echoes and produces structured artifacts during coherent focusing, making pre-focusing recovery important for SAR image restoration (SAR-IR). However, fixed state-space scan geometries limit adaptation to diverse interference morphologies, while dense multi-directional scanning incurs redundant computation. We propose MorphRoute-Mamba, a time-frequency SAR-IR framework that learns morphology-guided routing over predefined directional state-space experts. Guided by full record spectral context and Fast-SCNN-predicted interference masks, the framework selects a single expert per block and adaptively fuses its output with context from reliability-ranked anchors. This combines directional state-space propagation with reliability-guided contextual access. For efficient deployment, validation-guided route compilation fixes block-wise scan choices while retaining input-dependent conditioning and context fusion. Simulated experiments demonstrate reconstruction gains across the time-frequency, echo, and focused image domains, while measured SAR scenes provide complementary evidence of interference suppression and structural preservation. Component ablations support the proposed design, and deployment profiling demonstrates a favorable reconstruction-latency trade-off.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
SAR image restoration, state-space models, sparse directional routing, radio-frequency interference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GBTZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies radio-frequency interference (RFI) suppression for SAR at the pre-focusing stage. Rather than restoring an already focused SAR image, the method operates on complex echoes represented in the time-frequency domain. It additionally uses the full-record spectrum and an estimated interference mask to condition restoration. The main proposed idea, MorphRoute, is to associate the morphology of interference in the time-frequency plane with the propagation geometry of a state-space model. The network provides a bank of predefined scan patterns and selects one scan expert at each block according to the input conditions. Features from locations considered reliable are also retrieved and incorporated into the restoration process. For deployment, the input-dependent scan choices can be compiled into a fixed block-wise configuration based on validation-set statistics. The paper evaluates reconstruction across the time-frequency representation, recovered echo, and focused SAR image, and reports results on both simulated and measured data. The complete system achieves strong restoration performance and favorable computational efficiency compared with several restoration baselines.

Critical strengths

The pre-focusing formulation is well motivated. Processing complex SAR echoes before image formation is substantially more specific to the underlying sensing problem than treating a focused SAR amplitude image as a generic image-restoration input. The evaluation is also carried through the time-frequency, echo, and focused-image domains, which provides a coherent test of whether improvements survive the SAR processing chain. The motivating observation is interesting: different RFI types exhibit different geometrical structures in the time-frequency plane, while sequential state-space processing imposes a particular geometry of information propagation. Relating these two geometries is a potentially useful direction. The paper also contains meaningful evidence that SAR-specific conditioning matters. In particular, removing the interference-mask information only from the routing pathway degrades performance while retaining the other mask-dependent components. The appendix further shows that different measured interference inputs can lead to different expert-selection patterns. Thus, I do not think the issue is that interference information has no effect on routing.

Critical weaknesses

My main concern is that the experiments do not establish the paper's central mechanistic claim as strongly as the presentation suggests.

The key hypothesis is not merely that interference information is useful, or that different scan orders can affect restoration. The more interesting claim is that different interference morphologies call for different propagation geometries, and that the proposed router exploits this correspondence. I find the evidence for this stronger claim insufficient.

First, the proposed scan bank is largely domain-inspired rather than derived from the stated mechanism. The frequency, time, stripe, and positive/negative ridge traversals are plausible choices given common RFI patterns, but the paper does not derive why these particular seven geometries form an appropriate set, why these discretized directions are sufficient, or why a compound interference pattern should be represented by selecting a single scan expert. As a result, the method can also be interpreted more conservatively as defining a domain-inspired architecture search space and learning which member to use.

Second, there is a gap between the theoretical analysis and the actual routing algorithm. The theoretical result explains why changing serialization order changes causal accessibility and propagation distance between positions. This supports the general statement that scan order can matter. However, it does not provide a criterion mapping a particular interference support or morphology to an optimal scan direction. The learned router does not explicitly optimize the theoretical perturbation bound, nor is there evidence that its selected directions agree with directions predicted by that analysis. Thus, the theory and the router remain related at the level of intuition, but the paper does not establish that the router succeeds for the mechanism described by the theory.

The static-compilation result makes this issue particularly important. Replacing the input-dependent expert identity with a fixed block-wise route decreases focused-image PSNR only from 54.95 to 54.74 dB. Importantly, this experiment does not remove all input adaptivity: spectral/mask conditioning, feature fusion, anchor retrieval, and the selective state-space operations remain input dependent. What it removes specifically is the input-dependent choice of propagation geometry. Therefore, the small 0.21 dB difference raises a direct question about the central claim. Much of the performance appears obtainable from a fixed heterogeneous scan architecture together with the other adaptive components, rather than from selecting different propagation geometries for different interference morphologies.

The scan-bank ablation does not fully resolve this issue. Replacing the seven-direction bank with axial scans changes both the available propagation geometries and the number of experts. The performance difference therefore cannot be uniquely attributed to the morphology-inspired non-axial directions. A same-capacity comparison against alternative or arbitrary scan banks would be more informative for establishing that the proposed geometries themselves are important.

Finally, attribution of the headline performance improvement is complicated by knowledge distillation and the other components of the system. For the same dynamic architecture, distillation improves focused-image PSNR from 52.14 to 54.95 dB. When MambaIRv2 is also trained with distillation, it reaches 54.37 dB, reducing the architectural performance gap to 0.58 dB. This remaining improvement may still be meaningful, particularly together with the reported efficiency advantages, but the much larger margin against the original baselines conflates architecture and training procedure. Moreover, the remaining difference still includes spectral conditioning, mask conditioning, conditional fusion, the scan bank, anchor retrieval, and routing. It therefore cannot be attributed specifically to morphology-dependent propagation.

For these reasons, I find the empirical/system contribution more convincing than the claimed mechanistic contribution. The experiments demonstrate a strong SAR-specific restoration system combining domain conditioning, multiple propagation geometries, contextual retrieval, and distillation. They do not yet demonstrate with comparable strength that interference morphology determines an appropriate propagation geometry, which is the more distinctive scientific claim of the paper.

Questions for authors

Can the authors compare morphology-conditioned routing against fixed, random, and learned-but-morphology-agnostic routing while keeping the expert count, parameter count, scan bank, and computational budget identical? This would more directly isolate the contribution of input-dependent propagation-geometry selection.

Can the theoretical analysis be used to predict a preferred scan direction for a given interference support? If so, how often does the learned router agree with this prediction? If not, what precisely connects the theoretical perturbation analysis to the routing rule beyond showing that scan order matters in general?

How should the authors interpret the 0.21 dB difference between dynamic and compiled routing with respect to the central morphology-routing hypothesis? Since the compilation fixes specifically the input-dependent expert identity while retaining the other adaptive components, this appears to suggest that most of the performance does not require sample-specific propagation geometry.

Can the authors evaluate the proposed seven-expert bank against alternative scan banks with the same number of experts and comparable computational cost? The current axial-only comparison confounds scan geometry with expert count. Under a common distillation protocol, what fraction of the remaining advantage over MambaIRv2 can be attributed specifically to routing rather than to the other SAR-specific conditioning and retrieval components?

Reviewer_x5rXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MorphRoute-Mamba for SAR radio-frequency interference removal and imaging. It combines time-frequency and frequency-domain features with a predicted interference mask. A sparse router selects one of seven Mamba scan experts, while a reliability-based anchor branch retrieves additional context outside the selected scan path.

Critical strengths

The scan order design for Mamba is well motivated on SAR restoration under interference. Different orders change which time-frequency locations are upstream of a target and the path through which information is propagated.

The time-frequency representation provides localization of temporally varying interference, while the full FFT provides higher-resolution spectral information.

Evaluation includes scene-disjoint simulated data and measured interference, including UWI and SFI2+SFTI.

Critical weaknesses

The main claim links interference morphology to adaptive scan geometry. However, the compiled static model fixes one expert per block for all inputs and interference categories, with only a small PSNR drop relative to dynamic routing. This suggests that a substantial part of the gain may come from adaptive conditioning, anchors, and fusion rather than per-input morphology-guided expert selection itself.

The paper does not compare hard Top-1 routing with soft, Top-2, uniform multi-expert, or fixed-route alternatives. In addition, Fast-SCNN mask accuracy is not reported, particularly for unseen measured interference. Since mask errors can simultaneously affect conditioning, routing, and anchor selection, their impact on final restoration should be investigated.

The proposed “anchor” reliability score is based on the predicted mask response and local feature consistency, but no analysis verifies that the selected anchors correspond to clean and informative SAR regions. Visualising the selected anchors and analysing their overlap with interference regions would help clarify what information the anchors actually represent.

The no-anchor ablation shows that the anchor mechanism improves performance, but does not establish that it specifically compensates for the limitation of single-direction Mamba. Comparisons among anchor-only, single- and multi-scan configurations would provide stronger evidence for the claimed complementarity.

The six compound training categories are constructed from combinations of the same five base interference types. Therefore, the experiments provide limited evidence of generalisation to genuinely unknown interference. Additional unseen-interferences would strengthen the generalisation claim.

The simulated data cover a wide ISNR range, but performance is not analysed across different interference-strength levels. An ISNR-stratified evaluation would help demonstrate robustness, particularly since weak interference may also lead to less reliable mask detection, routing, and anchor selection.

The proposed routing mechanism is strongly coupled to the Mamba backbone. It would be useful to compare the framework with alternative CNN or Transformer backbones to clarify whether the performance gain arises from the proposed SAR-specific conditioning and anchor mechanisms or specifically from addressing Mamba's scan-order dependency.

Questions for authors

Compare hard Top-1 routing with soft, Top-2, uniform multi-expert, and fixed-route alternatives to better demonstrate the benefit of morphology-guided expert selection.

Evaluate the accuracy and robustness of the predicted interference masks, particularly for unseen measured interference, and analyse how mask errors propagate to conditioning, routing, anchor selection, and final restoration performance.

Provide additional analysis and visualisation of the selected anchors to demonstrate whether they indeed correspond to reliable and informative SAR regions.

Further analyse the complementarity between anchor retrieval and single and multi-direction Mamba.

Provide performance across different ISNR levels and identify the conditions under which the method begins to fail or degrade significantly, particularly for weak interference where mask detection may become difficult.

Reviewer_8yZuOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In synthetic aperture radar (SAR) imaging, signals emitted by other radio devices in the same band leak into the received echoes as radio-frequency interference (RFI), producing bright streaks or widespread blurring in the focused image. This paper aims to remove such interference from the raw echoes before imaging. The echoes are first transformed into a time–frequency representation, in which interference appears as regular patterns (horizontal bands, vertical bands, diagonal lines, etc.), and a neural network, MorphRoute-Mamba, then suppresses the interference in this domain. The network has three main design elements: (1) two auxiliary inputs serve as prompts, namely the spectrum of the whole echo and a predicted interference-location mask; (2) seven "experts", each scanning along a different direction, with a router selecting one expert per image block according to the interference morphology; and (3) knowledge distillation from a larger Transformer teacher during training.

The training data are built by superimposing 11 types of simulated interference onto 73 interference-free measured scenes; the test set contains 11 simulated scenes and 4 scenes with measured interference. The proposed method achieves the best metrics in the echo, time–frequency and image domains, and the lowest inference latency.

Critical strengths
  1. The task and evaluation are clearly defined. Interference is suppressed in the time–frequency domain, and the improvements carry over consistently to the echo and image domains, across all 11 interference types.

  2. The experimental design is generally clean. Training, validation and test scenes do not overlap; the measured scenes are accompanied by interference-free revisit images of the same site as references; and the ablation study is fairly comprehensive.

Critical weaknesses

My main concern is that the paper may attribute gains that come from a training technique to the new architecture; the gain of the architecture itself and the role of its core design remain in question.

  1. The comparison in the main text may not be fair: the proposed method uses knowledge distillation, whereas the baselines do not. Knowledge distillation is a generic training technique that is independent of the network architecture. In the main text, the proposed method leads the second-best baseline by 2.84 dB; however, Appendix Table 15 shows that when the baselines are trained with the same distillation, MambaIRv2 reaches 54.37 dB, only about 0.5 dB below the proposed method, while the proposed method without distillation reaches only 52.14 dB, i.e., below the distilled baselines. The main text should report the distilled baselines and revise the conclusions accordingly.

  2. The performance of the teacher network is never reported. Distillation contributes 2.81 dB, yet the PSNR, parameter count and latency of the teacher are not given anywhere. If the teacher is clearly better, the paper needs to explain why the teacher is not used directly; if the teacher is worse than the student, the paper needs to explain why distillation helps.

  3. The claimed "dynamic selection of scan direction according to interference morphology" is not validated. The authors' own experiments show that freezing the expert of every block reduces PSNR only from 54.95 to 54.74 dB; that is, a network that does not look at the input at all achieves 99.6% of the performance. Fig. 13 likewise shows that most blocks almost always select the same direction. This contradicts the motivation illustrated in Fig. 2, where different interference types are said to prefer different directions. I suggest adding a control: a fixed-expert network without any router, trained from scratch.

These issues currently lead me to lean towards rejection. That said, I believe most of them can be addressed. If the authors can largely resolve the concerns above, I would be inclined to raise my score.

In addition, the formula font in Fig. 3 is too small and should be enlarged.

Questions for authors

See weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.