Pith. sign in

REVIEW 4 major objections 6 minor 16 references

AtrousMamaba: An Atrous-Window Scanning Visual State Space Model for Remote Sensing Change Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that Mamba can capture fine local detail, not just long-range context, if its selective scans are reordered into atrous windows, and that this yields top accuracy on six change-detection datasets.

desk verdict A genuinely new Mamba scan variant for change detection with a solid six-dataset evaluation, but a glaring Table 2 anomaly and an under-isolated ablation prevent the headline claim from being verifiable as printed. read the letter →

arxiv 2507.16172 v1 pith:GBR5U3HC submitted 2025-07-22 cs.CV

classification cs.CV
keywords changedetectionMambavisualstatespacemodelatrouswindowscanningremotesensingimagerybinarysemanticsiamesenetwork
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

AtrousMamba is built to answer a specific open question: can Mamba, a state-space model known for linear-complexity long-range modeling, extract fine-grained local features as well as convolutional networks do? The paper argues yes, provided the scanning order is changed. Instead of flattening whole rows or columns, the atrous window scan splits the feature map into four groups of windows at different dilation rates, unfolds each group into sequences, and runs them through four parallel selective-scan (S6) blocks, then reweights the channels with a lightweight attention merge. On this foundation the authors construct two end-to-end networks, one for binary and one for semantic change detection, and report that they outperform CNN-, Transformer-, and Mamba-based methods on six benchmark datasets. The claim, if correct, is that a single Mamba architecture can simultaneously preserve fine boundary detail and global context across both change-detection subtasks.

What carries the argument

The load-bearing component is the atrous window scan module (AWSM), made of AWScan and AWSMerge. AWScan uses several dilation rates to cut the feature map into four groups of windows of different sizes, unfolds the patches inside each window into a sequence along a horizontal path, and feeds the four groups to four independent S6 selective-scan blocks in parallel; AWSMerge pools each channel globally and applies two fully connected layers to reweight the channels before recombination. The intended effect is to shorten the distance between adjacent tokens within the same semantic region, so local features propagate cleanly, while the different window scales let the receptive field expand progressively to cover global context.

What would settle it

Train the same AWMambaBCD decoder with a fixed-size (non-atrous) window scan substituted for the atrous scan while keeping AWSMerge and the four parallel S6 branches identical; if accuracy on CLCD, SYSU-CD, and WHU-CD does not drop, the progressive atrous expansion is not the cause of the reported gains. This directly tests the mechanism that the paper's Table 5 leaves entangled.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the spatial discontinuities introduced by standard Mamba scans are not inevitable: an atrous window scan that processes nearby tokens together lets a selective state-space model keep the local precision of a CNN while retaining the global receptive field Mamba is known for. The atrous window scan visual state space (AWVSS) module partitions the feature map into four window groups with dilation rates such as 2, 5, 7, and 9, unfolds each group into horizontal token sequences, and processes the four sequences in parallel with independent S6 blocks; the atrous window scan merge (AWSMerge) then recalibrates channels. Tables 1 and 2 report the resulting AWMambaBCD and AWMambaSCD models as the best performers on their six datasets, and the ablation in Table 5 shows the atrous window scan improving over the cross-scan baseline on CLCD, SYSU-CD, and WHU-CD. The authors read these results as evidence that Mamba not only captures long-range dependencies in visual data but also preserves fine-grained local details.

Load-bearing premise

The paper assumes that the accuracy gain comes from the atrous window ordering of tokens, but the ablation does not separate that ordering from the added channel-reweighting merge and the four parallel scan branches, so those auxiliary components could be the actual source of the improvement.

Editorial extensions

If this is right

  • If the central claim holds, Mamba-based decoders are viable for dense prediction tasks that demand sharp boundaries, not only for tasks that reward long-range context.
  • The same AWVSS decoder works in both binary and semantic change detection networks, so a shared Mamba framework can cover both subtasks instead of requiring separate scanning designs.
  • The Table 5 ablation shows the atrous window scan beating the cross-scan baseline on all three binary datasets, meaning scan ordering alone is a productive axis for improving Mamba-based change detection.
  • Because the reported parameter counts and FLOPs stay close to the Mamba baseline and well below Transformer comparators, the local-detail gain does not sacrifice the linear-complexity advantage.

Reading between the lines

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

  • A direct way to isolate the mechanism would be to keep AWSMerge and the four parallel S6 branches fixed and swap only the atrous window partition for a fixed-size or random window partition; the paper's Table 5 varies the scanning strategy and the merge module together, so this swap would test whether progressive dilation is what matters.
  • If the atrous ordering is the true cause, the same scan should transfer to other dense prediction tasks such as semantic segmentation of remote sensing imagery, and to SSM backbones other than VMamba; the paper only evaluates change detection.
  • Because the six datasets span ground resolutions from 0.2 m to 30 m, the fixed dilation rates (2, 5, 7, 9) may need to scale with sensor resolution; a learnable or resolution-aware rate schedule is a natural next step.
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

4 major / 6 minor

Summary. The manuscript proposes AtrousMamba, a visual state space model for remote sensing change detection. The method partitions a feature map into four groups of windows with different atrous rates, scans each group horizontally through parallel S6 modules, and merges the results with a channel-attention module called AWSMerge. On top of this module the authors build two end-to-end frameworks, AWMambaBCD for binary change detection and AWMambaSCD for semantic change detection, using a VMamba-based siamese encoder. The paper reports experiments on six datasets (CLCD, SYSU-CD, WHU-CD, SECOND, Landsat-SCD, JL1-SCD) and claims the proposed frameworks outperform CNN-based, Transformer-based, and previous Mamba-based methods.

Significance. If the empirical claims hold, the atrous-window scanning idea would be a useful contribution to the line of work on adapting Mamba to dense prediction, and the paper usefully covers both binary and semantic change detection. The design is clearly motivated and the coverage of six benchmark datasets is a strength. However, the central claim of universal superiority is not verifiable as printed: Table 2 contains an internal inconsistency that puts the reported numbers in doubt, the ablations do not isolate the scanning strategy from the added auxiliary components, and the reported gains are small with no error bars, no multiple runs, and no released code. The paper therefore cannot, in its current form, support the mechanistic conclusion that atrous window scanning is what preserves fine-grained local details.

major comments (4)
  1. [Table 2 and §4.6.2] As printed, Table 2 is internally inconsistent: in the MambaSCD_T row, after the parameter/FLOP values 34.69/26.66, the nine numeric entries are 63.04, 72.87, 94.80, 87.63, 88.09, 57.24, 90.21, 88.55, 64.46. Under the column alignment used for every other row, the third value (94.80) would be the SECOND SeK, which is impossible because all other SECOND SeK values in the table are between 16.09 and 24.95 and because §4.6.1 states that AWMambaSCD_S reaches the highest SECOND SeK of 24.95. If instead 94.80 is read as the Landsat-SCD F1, then the baseline MambaSCD_T (94.80) exceeds the proposed AWMambaSCD_S (89.03) on Landsat-SCD, directly contradicting §4.6.2's claim that AWMambaSCD_S achieves the highest F1 on that dataset. Either interpretation makes the table unusable as evidence for the headline claim that the proposed framework outperforms all compared Mamba-based methods. Please correct the table and re-verify every reported cell against the raw experiment logs.
  2. [§5.2, Table 5] The scanning-strategy ablation does not isolate the atrous window scan: the 'Atrous Window Scan' row replaces the baseline CSM with the full AWVSS module, which simultaneously introduces four parallel S6 branches and the AWSMerge channel-attention module. The reported improvements in IoU and F1 could therefore come from the added capacity and the attention mechanism rather than from the atrous window ordering. To support the mechanistic claim that atrous window scanning itself preserves local detail, please add an ablation that holds the auxiliary components fixed, for example CSM with four parallel branches plus the same AWSMerge, or atrous window scanning without AWSMerge.
  3. [§5.1, Tables 3–4] The 'different backbone architectures' comparison is confounded: the Mamba row uses the full AWMambaBCD_S / AWMambaSCD_S method, whereas the ResNet50 and Swin-Small rows appear to use only the backbone without the proposed decoder components, based on the text and the table layout. The conclusion that 'Mamba outperforms both CNN-based and Transformer-based backbone networks' therefore conflates the backbone choice with the decoder design. Please report ResNet50 and Swin-Small with the same decoder components, or state explicitly what decoder is used for each row, and compare under matched training budgets.
  4. [§4.5, §4.6] The claimed superiority is based on single point estimates, and most improvements over the closest baselines are small (roughly 0.1 to 2.8 IoU points). No error bars, no multiple seeds, no significance tests, and no code are provided, and the paper says only that code 'will be made publicly available' without a link (Data availability says data are 'on request'). Under these conditions, the evidence as printed is not sufficient to establish that the reported gains are reliable. Please report mean and standard deviation over at least three independent runs, provide a public code repository, and explicitly discuss which differences are within run-to-run variability.
minor comments (6)
  1. [Title and Fig. 1 caption] The title and the caption of Fig. 1 spell 'AtrousMamaba' (missing one 'b') inconsistently with the body text and the rest of the paper; please unify the spelling.
  2. [§3.2.1, §3.2.2] The text refers to 'Section D' for the detailed decoder architecture, but no such section appears in the manuscript; either include the omitted decoder details as an appendix or remove the cross-reference.
  3. [§3.5.2, Eq. (17)] The contrastive semantic change loss in Eq. (17) is not fully specified: the notation X1 and X2 is introduced only in prose, and there is no margin, temperature, or distance threshold to define when a pair is pushed apart or pulled together; please clarify the exact loss formulation and how the ground-truth label Yc enters it.
  4. [§4.4, §4.5] The baseline is spelled 'CD-Lamba' in §4.4 and 'CD-Lambda' in §4.5; please standardize the name to match the cited reference.
  5. [Data availability and §4.2] The data availability statement says 'Data will be made available on request', while §4.2 states that source code will be made publicly available; please provide a repository link or explain the discrepancy.
  6. [Abstract] There is a capitalization inconsistency in the abstract ('In this paper, We propose'); please proofread for minor grammatical issues throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims rest on external benchmark measurements, not on fitted inputs or self-referential derivation.

full rationale

AtrousMamba is an empirical architecture paper. Its central claim is that the proposed framework outperforms CNN-, Transformer-, and Mamba-based methods on six remote sensing change detection benchmarks. The support for this claim consists of measured accuracy metrics (F1, IoU, OA, mIoU, SeK) obtained by training and evaluating the model and external baselines on public datasets. There is no derivation of predicted quantities from fitted parameters, no parameter that is renamed as a prediction, and no equation in Section 3 whose output is identical to an input by construction. The scanning strategy, AWVSS, is a design choice rather than a theoretical derivation, and its evaluation is an empirical ablation against the Cross-Scan Module baseline. Concerns that the ablation does not fully isolate the scanning mechanism from the added attention and parallel branches are attribution or experimental-design concerns, not circularity: the paper does not define the scanning strategy in terms of its measured accuracy gain. Likewise, the apparent inconsistency in Table 2 is a correctness and reproducibility issue, not a circular step, because the reported numbers are externally benchmarked results rather than outputs of a self-referential argument. No self-citations are used as load-bearing support, and no uniqueness theorem or prior result by the same authors is invoked to forbid alternative explanations. The paper is self-contained as an empirical study, so no circular reasoning is present.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on hand-chosen architectural hyperparameters (dilation rates 2, 5, 7, 9; four scan groups; unreported loss weights), on the borrowed VMamba backbone, and on the implicit assumption that the new scanning pattern rather than the added attention module causes the improvement. No external theoretical or independent empirical evidence is provided for these choices.

free parameters (3)
  • Atrous dilation rates = 2, 5, 7, 9
    Four rates are hand-picked as examples and used in all experiments; no sensitivity study or tuning is reported.
  • Number of atrous window groups = 4
    The design always uses four groups; other counts are not ablated, so the value is a free design choice.
  • Loss combination weights λ1, λ2, λ3 = Not reported
    The total loss in Eq. (18) includes three weight hyperparameters, but their values are never given, making the training objective incompletely specified.
assumptions (3)
  • domain assumption Mamba/VMamba backbones provide effective features for change detection
    The encoders are borrowed from VMamba without independent verification on these datasets; the paper only compares decoders on top of this backbone.
  • ad hoc to paper The four parallel S6 scans can be merged by channel attention without loss of spatial coherence
    No theory or experiment isolates the contribution of the parallel-scan design from the merge module; this is a design choice specific to the paper.
  • domain assumption S6 selective scan remains stable when applied to short window patch sequences
    The paper applies S6 to short sequences inside windows, while the original Mamba was designed for long sequences; no analysis of short-sequence behavior is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AtrousMamaba: An Atrous-Window Scanning Visual State Space Model for Remote Sensing Change Detection." pith.science (2026). https://pith.science/paper/GBR5U3HC

@misc{pith2026250716172,
  author       = {Pith},
  title        = {Pith review of: AtrousMamaba: An Atrous-Window Scanning Visual State Space Model for Remote Sensing Change Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GBR5U3HC}},
  note         = {Machine review of arXiv:2507.16172}
}
read the original abstract

Recently, a novel visual state space (VSS) model, referred to as Mamba, has demonstrated significant progress in modeling long sequences with linear complexity, comparable to Transformer models, thereby enhancing its adaptability for processing visual data. Although most methods aim to enhance the global receptive field by directly modifying Mamba's scanning mechanism, they tend to overlook the critical importance of local information in dense prediction tasks. Additionally, whether Mamba can effectively extract local features as convolutional neural networks (CNNs) do remains an open question that merits further investigation. In this paper, We propose a novel model, AtrousMamba, which effectively balances the extraction of fine-grained local details with the integration of global contextual information. Specifically, our method incorporates an atrous-window selective scan mechanism, enabling a gradual expansion of the scanning range with adjustable rates. This design shortens the distance between adjacent tokens, enabling the model to effectively capture fine-grained local features and global context. By leveraging the atrous window scan visual state space (AWVSS) module, we design dedicated end-to-end Mamba-based frameworks for binary change detection (BCD) and semantic change detection (SCD), referred to as AWMambaBCD and AWMambaSCD, respectively. Experimental results on six benchmark datasets show that the proposed framework outperforms existing CNN-based, Transformer-based, and Mamba-based methods. These findings clearly demonstrate that Mamba not only captures long-range dependencies in visual data but also effectively preserves fine-grained local details.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 2 canonical work pages

  1. [5]

    arXiv preprint arXiv:2111.00396

    Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396. Huang, T., Pei, X., You, S., Wang, F., Qian, C., Xu, C.,

  2. [7]

    arXiv preprint arXiv:2401.10166

    Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B., Oct

  3. [8]

    Semantic change detection using a hierarchical semantic graph interaction network from high-resolution remote sensing images. ISPRS J. Photogramm. Remote Sens. 211, 318-335. Ma, J., Li, F., Wang, B., 2024a. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722. Ma, X., Zhang, X., Pun, M.-O., 2024b. RS3...

  4. [9]

    arXiv preprint arXiv:2402.02491

    Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491. Shi, Q., Liu, M., Li, S., Liu, X., Wang, F., Zhang, L.,

  5. [11]

    ADS-Net: An Attention-Based deeply supervised network for remote sensing image change detection. Int. J. Appl. Earth Obs. Geoinf. 101, 102348. Wang, L., Li, D., Dong, S., Meng, X., Zhang, X., Hong, D., 2024a. PyramidMamba: Rethinking Pyramid Feature Fusion with Selective Space State Model for Semantic Segmentation of Remote Sensing Imagery. arXiv preprint...

  6. [13]

    arXiv preprint arXiv:2406.02395

    GrootVL: Tree Topology is All You Need in State Space Model. arXiv preprint arXiv:2406.02395. Yang, K., Xia, G.-S., Liu, Z., Du, B., Yang, W., Pelillo, M., Zhang, L.,

  7. [14]

    arXiv preprint arXiv:2406.04207

    CDMamba: Remote sensing image change detection with mamba. arXiv preprint arXiv:2406.04207. Zhang, M., Shi, W.J.I.T.o.G., Sensing, R., 2020b. A feature difference convolutional neural network-based change detection method. IEEE Trans. Geosci. Remote Sens. 58, 7232-7246. Zhang, R., Zhang, H., Ning, X., Huang, X., Wang, J., Cui, W.,

  8. [15]

    arXiv preprint arXiv:2404.02668

    RS-Mamba for Large Remote Sensing Image Dense Prediction. arXiv preprint arXiv:2404.02668. Zhao, S., Zhang, X., Xiao, P., He, G.,

Show all 16 references
  1. [16]

    ChangeMask: Deep multi-task encoder- transformer-decoder architecture for semantic change detection. ISPRS J. Photogramm. Remote Sens. 183, 228-239. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X., 2024a. Vision mamba: Efficient visual representation learning with bi...

  2. [2018]

    IEEE, pp

    Fully convolutional siamese networks for change detection, 2018 25th IEEE International Conference on Image Processing (ICIP). IEEE, pp. 4063-4067. Daudt, R.C., Le Saux, B., Boulch, A., Gousseau, Y.,

  3. [2020]

    arXiv preprint arXiv:2010.11929

    An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. Fang, S., Li, K., Shao, J., Li, Z.,

  4. [2021]

    IEEE Trans

    Remote sensing image change detection with transformers. IEEE Trans. Geosci. Remote Sens. 60, 1-14. Chen, H., Song, J., Han, C., Xia, J., Yokoya, N., 2024a. ChangeMamba: Remote Sensing Change Detection with Spatio-Temporal State Space Model. arXiv preprint arXiv:2404.03425. Ch...

  5. [2022]

    arXiv preprint arXiv:2208.04933

    Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933. Song, D., Dong, Y., Li, X.,

  6. [2023]

    arXiv preprint arXiv:2312.00752

    Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752. Gu, A., Goel, K., Ré, C.,

  7. [2024]

    arXiv preprint arXiv:2403.09338

    Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338. Ji, S., Wei, S., Lu, M.,

  8. [2025]

    arXiv preprint arXiv:2501.15455

    CD-Lamba: Boosting Remote Sensing Change Detection via a Cross-Temporal Locally Adaptive State Space Model. arXiv preprint arXiv:2501.15455. Xiao, Y., Song, L., Huang, S., Wang, J., Song, S., Ge, Y., Li, X., Shan, Y.,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.