{"id":"152f3347-e7bd-4638-9c45-569298d5b6a8","arxiv_id":"2506.19263","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A three-plane selective scan module improves change detection F1 by up to 3.7 points over Mamba baselines on five remote sensing datasets.","lead":"Researchers propose a Mamba-based module that scans remote sensing image features along height, width, and channel axes to better detect changes between image pairs. The method reports the highest F1 scores on five change detection benchmarks, with gains of 0.2 to 3.7 percentage points over prior Mamba methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ablation evidence for the central 3D-SSM claim is not statistically grounded: the SS2D baseline differs between Table III and Table IV on SYSU-CD, and all reported gains are single-run F1 differences of at most 0.34 points.","rationale":"In good faith, the paper proposes a plausible architectural extension and provides a code link, five datasets, and extensive ablations. However, the central claim is empirical: 3D-SSM must outperform SS2D because of the extra scan planes. The reader's weakest assumption concerned whether channel-axis scanning is mechanistically useful; that is a fair concern, but the more load-bearing condition is that the measured effect is real and reproducible. The manuscript never reports experimental variance, and the SS2D baseline itself is not stable across tables: Table III gives F1=84.56 on SYSU-CD while Table IV's HW-only row gives F1=84.79 for what should be the same configuration. This internal inconsistency means the 0.34-point F1 gain in Table III is within the range of baseline drift. Table IV's text also contradicts its own numbers regarding HW+WC and HC+WC, further undermining confidence in the ablation narrative. If a repeated-seed experiment shows a robust, statistically significant improvement of 3D-SSM over SS2D on both datasets, the concern is resolved and the central claim stands. If the improvement vanishes or is smaller than the baseline drift, the central claim fails. Because this is checkable and the reader already conditioned acceptance on reproducibility-related issues, the verdict should remain CONDITIONAL (i.e., no change from the reader's recommendation).","tokens_in":18157,"tokens_out":9419,"duration_ms":99476,"concrete_test":"Run the SS2D and 3D-SSM configurations on WHU-CD and SYSU-CD with at least 5 independent random seeds each, using the same pretrained VMamba-S backbone, batch size, number of epochs, and the exact released code; report mean plus/minus standard deviation of F1 and OA. Perform a paired significance test (e.g., bootstrap or Wilcoxon signed-rank) between Table III's SS2D and 3D-SSM, and between Table IV's HW-only and all-three-perspective conditions. Also verify whether Table III's SS2D and Table IV's HW-only are meant to be the same configuration; if they still differ on SYSU-CD (84.56 vs. 84.79), one ablation has an undocumented difference. If the 3D-SSM gain is not significant at p<0.05 or does not exceed the observed baseline drift, the central claim that 3D scanning improves change detection is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that scanning along HW, HC, and WC planes improves change detection over HW-only scanning (SS2D). The direct evidence is Table III (3D-SSM vs. SS2D: 95.30 vs. 95.05 on WHU-CD, 84.90 vs. 84.56 on SYSU-CD) and Table IV (all three vs. HW-only: 95.30 vs. 95.05 and 84.90 vs. 84.79). These gains are 0.11-0.34 F1 points, with no error bars, no repeated runs, and no significance test. More concerning, the supposedly identical SS2D configuration differs between the two tables on SYSU-CD: Table III reports F1=84.56, OA=92.96 for SS2D, while Table IV's HW-only row reports F1=84.79, OA=93.08. This 0.23-point baseline drift is almost as large as the 0.34-point improvement claimed in Table III, so the comparison is not controlled. Table IV's narrative also contradicts its own numbers: it says combining HW with either HC or WC 'resulted in reduced F1 scores,' but HW+WC improves F1 over HW-only on both datasets (95.22 vs. 95.05 on WHU-CD; 84.85 vs. 84.79 on SYSU-CD), and it says HC+WC improves F1 by 0.17% and 0.06% over SS2D, while the table shows 95.05 vs. 95.05 and 84.70 vs. 84.79. These inconsistencies do not by themselves prove the claim false, but they make the central causal assertion untestable from the manuscript: the reported benefit of 3D scanning could be run-to-run variation or an undocumented protocol difference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes 3D-SSM, a selective scan module that, in addition to the standard HW spatial scan, also scans feature tensors along the HC and WC planes. The module is embedded in a change-detection framework together with a spatiotemporal interaction module (SIM) and a multi-branch feature extraction module (MBFEM), and is compared against 13 prior methods on five remote sensing change detection benchmarks. The central claim is that scanning along all three planes improves change detection over the HW-only SS2D scan.","tokens_in":18510,"tokens_out":7217,"duration_ms":69171,"significance":"If the effect is real, 3D-SSM is a simple and potentially useful extension of selective scanning for change detection. The paper's strengths are its clear architecture description, the release of code, and a broad experimental campaign with five datasets and many ablations. However, the central claim rests on small, single-run F1 differences, and the ablation tables contain internal inconsistencies and an uncontrolled baseline. The significance of the paper therefore depends on whether these issues can be resolved with corrected and statistically grounded experiments.","major_comments":[{"comment":"The narrative and the table contradict each other. The text states that combining HW with either HC or WC 'resulted in reduced F1 scores,' but Table IV shows HW+WC improves F1 over HW-only on both datasets (95.22 vs. 95.05 on WHU-CD; 84.85 vs. 84.79 on SYSU-CD). The text also states that HC+WC improves F1 by 0.17% and 0.06% relative to SS2D, whereas the table shows 95.05 vs. 95.05 on WHU-CD and 84.70 vs. 84.79 on SYSU-CD. Because this table is the direct evidence for the three-plane scanning design, the internal inconsistency makes the stated ablation conclusions untestable.","section":"§IV-C2, Table IV"},{"comment":"The SS2D baseline is not stable across the two ablation tables. On SYSU-CD, Table III reports SS2D F1=84.56 and OA=92.96, while Table IV reports the HW-only configuration as F1=84.79 and OA=93.08. This 0.23-point F1 drift between supposedly identical configurations is close to the 0.34-point F1 improvement attributed to 3D-SSM in Table III. The comparison is therefore not controlled, and the reader cannot attribute the reported gains to 3D-SSM rather than to run-to-run variation or an undocumented protocol change.","section":"§IV-C1 and §IV-C2, Tables III and IV"},{"comment":"All key comparisons are single runs with no error bars, multiple seeds, or significance tests. The decisive differences, such as 3D-SSM vs. SS2D (+0.25 F1 on WHU-CD and +0.34 F1 on SYSU-CD in Table III; +0.25 and +0.11 in Table IV), are small relative to the baseline drift visible between Tables III and IV. The paper should provide repeated-run statistics or an equivalent significance analysis before claiming that 3D-SSM is superior to SS2D.","section":"§IV-C, Tables III-V"},{"comment":"The within-backbone comparison of 3D-SSM against SS2D is not capacity-matched. Replacing SS2D with 3D-SSM in the VSSB increases parameters from 10.63M to 73.34M and FLOPs from 281.17G to 503.21G at batch size 8, yet the reported gain is 0.74 F1 on WHU-CD. Such a large increase in model capacity can account for the improvement, so this experiment does not isolate the effect of scanning along the HC and WC planes.","section":"§IV-C6, Table VIII"}],"minor_comments":[{"comment":"The abbreviation 'VEM' appears in the equation while the text defines 'VME'; please unify the terminology and define the acronym at first use.","section":"Eq. (2)"},{"comment":"The text states that F1_G is used to enhance F2_G and F2_L, but the equations show F1_G being gated by F2 features; please align the text and the equations.","section":"§III-C, Eqs. (8)-(9)"},{"comment":"The loss is called cross-entropy in Eq. (18) but 'BCE loss' in the Table VII discussion; please use one consistent term.","section":"§III-E, Table VII"},{"comment":"The 'scaling' branch shown in Fig. 2(d) is not described in the equations; please clarify how the scaling branch contributes to the MBFEM output.","section":"§III-D, Eqs. (13)-(15)"},{"comment":"The phrase 'spatialtemporal interaction module' should be 'spatiotemporal interaction module'.","section":"Contributions"},{"comment":"The relation to 3DSS-Mamba [51], which also scans spatial and spectral dimensions, should be stated more precisely so that the novelty of scanning the HC and WC planes of a 2D feature map is clear.","section":"§II-C"},{"comment":"For DSIFN-CD, the test set contains only 192 image pairs; please state whether the reported metrics are computed on that small test set and comment on the resulting variance.","section":"§IV-A2, Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially publishable after a major revision. I would require the authors to (1) rerun the core ablations with multiple seeds and report mean and standard deviation, (2) reconcile the SS2D baseline between Tables III and IV, and (3) provide a capacity-matched comparison of SS2D vs. 3D-SSM inside the backbone. Given the small effect sizes, the central claim is not established by the present evidence. The release of code is a positive signal, but the repository should be checked to confirm the exact configurations used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper as a plausible but shaky architecture proposal. The authors replace the usual 2D selective scan in a Mamba backbone with a three-plane scan (HW, HC, WC) and add a spatiotemporal interaction module (SIM) and a multi-branch decoder (MBFEM). They report top F1 on five standard change-detection datasets and a code link. The architecture is clearly described and the ablation set is broad. Honest limitation section, too. That part is fine.\n\nThe real problem is the evidence for the central claim. The direct comparison of 3D-SSM versus SS2D is Table III: gains of 0.25 F1 on WHU-CD and 0.34 on SYSU-CD. Table IV repeats the HW-only configuration and reports the same WHU value but 84.79 instead of 84.56 on SYSU-CD. Same configuration, different numbers. That is a baseline drift almost as large as the claimed gain. Also, the Table IV narrative contradicts its own table: it says HW+HC or HW+WC reduce F1, but HW+WC improves F1 over HW-only on both datasets, and the HC+WC claim does not match the table either. No error bars, no repeated runs, no significance tests anywhere. For these benchmarks, 0.2-0.3 F1 differences are within run-to-run noise. The larger claims against M-CD (up to 3.74 F1 on DSIFN-CD) are more credible but still single runs.\n\nTwo smaller issues: the closest prior channel-axis scanning work, 3DSS-Mamba [51], is cited but never compared against, and the SIM equations have a typo (the second argument in Eq. 8 should likely be F1_L, not F1_G). Minor, but easy to fix.\n\nCredit where due: the full model beats the strongest Mamba baseline on all five datasets, the backbone swap in Table VIII shows a +0.74 F1 improvement from replacing SS2D with 3D-SSM, and the design is explained well enough that someone could reimplement it. The inconsistency between Table III and Table IV is the load-bearing flaw: it makes the main causal assertion untestable as written. That does not mean the idea is wrong, just that this paper does not yet demonstrate it.\n\nWho is this for? People working on Mamba-based change detection who want a concrete alternative scan strategy. It is not paradigm-shifting. I would send it to peer review because the architecture is plausible and the experimental scope is wide, but the authors need to redo the ablations with multiple seeds, fix the baseline drift, and add the missing 3DSS-Mamba comparison. My verdict: engage, but do not trust the headline numbers until those are addressed.","headline":"Useful but statistically fragile Mamba variant for change detection; the core 3D-scan benefit rests on 0.2-0.3 F1 differences that the paper's own tables do not consistently reproduce.","tokens_in":19087,"tokens_out":2155,"would_cite":false,"duration_ms":23873,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mamba-based change detector scans feature maps along height, width, and channel planes, reporting the best F1 on five remote sensing change-detection benchmarks.","keywords":["remote sensing change detection","Mamba","selective scan","3D-SSM","state space model","spatiotemporal interaction","multi-branch feature extraction","bi-temporal images"],"falsifier":"Train the model on WHU-CD and SYSU-CD with the HC and WC scan branches replaced by linear channel-mixing layers matched for parameter count and FLOPs, keeping the HW scan and all other modules unchanged; if the F1 score does not fall below the reported 95.30 and 84.90, the three-plane scanning geometry itself is not what carries the improvement.","tokens_in":17913,"feed_emoji":"🛰️","tokens_out":13107,"duration_ms":119055,"temperature":0.7,"pith_summary":"The paper claims that a Mamba-based change-detection model gains by running its selective scan along the channel axis as well as the two spatial axes. Mamba is a state-space sequence model whose selective scan processes a flattened sequence in linear time; earlier Mamba change detectors scanned only the height–width plane. The authors propose a 3D selective scan module (3D-SSM) that flattens a feature tensor along the HW, HC, and WC planes, scans each plane, and sums the three outputs. On top of it they build a spatiotemporal interaction module (SIM) for fusing bi-temporal features and a multi-branch decoder. Their experiments report the best F1 score among the compared methods on all five benchmark datasets, so the claim, if correct, is that channel-aware scanning is a practical gain for detecting changes in remote sensing imagery.","feed_headline":"3D selective scan tops five change-detection benchmarks","feed_subtitle":"Scanning height, width, and channel planes helps Mamba models catch subtle changes in satellite images.","key_machinery":"The central object is the 3D selective scan module (3D-SSM), a variant of the Mamba selective scan that treats the channel axis as a scan direction alongside height and width. It carries the argument by replacing the HW-only SS2D scan: for each input tensor it forms sequences along the HW, HC, and WC planes, runs the bidirectional selective scan (the VME operation of Vision Mamba) on each, and sums the results, so that every output position receives context from both spatial neighbors and channel neighbors. Two further components hang off it: the spatiotemporal interaction module (SIM), which uses 3D-SSM to extract global features from both temporal images and combines them with local convolutional features through dynamic gating before taking an absolute difference; and the multi-branch feature extraction module (MBFEM), which fuses 3D-SSM with FFT and convolutional branches in the decoder. The module is the only part of the architecture that is novel at the scan level.","core_discovery":"On the paper's own terms, the central discovery is that 3D-SSM captures global information from both the spatial plane and the channel perspective, giving a more comprehensive feature representation than the standard SS2D scan. For an input feature $F_{\\mathrm{in}} \\in \\mathbb{R}^{H \\times W \\times C}$, the module flattens it into three sequences: $X \\in \\mathbb{R}^{(H \\times W) \\times C}$ over the HW plane, $X_H \\in \\mathbb{R}^{(W \\times C) \\times H}$, and $X_W \\in \\mathbb{R}^{(H \\times C) \\times W}$. Each sequence goes through the bidirectional VME scan, and the outputs combine as $F_{3D} = F_{\\mathrm{out}} + F_{H}^{\\mathrm{out}} + F_{W}^{\\mathrm{out}}$. The authors build the SIM on this module to exchange global and local information between the two time points, and the MBFEM decoder combines convolution, FFT, and 3D-SSM branches. The reported F1 is the best among all compared methods on all five datasets; the ablations show the full three-plane combination is needed, since adding HC or WC to HW alone can lower F1.","pith_inferences":["A testable prediction from the paper's Table IV is that channel-plane scanning is only useful when anchored by the spatial scan: the HC and WC branches may be aligning channel structure within a spatially organized context rather than providing independent information, and a per-stage ablation of which scan directions help at which resolution would clarify this.","If the mechanism is about spectral change, the method's margin over SS2D should grow on multi-spectral or hyper-spectral change-detection data, where change often appears as a spectral shift rather than a spatial edge.","Because the parameter count is modest but FLOPs are high, a matched-FLOP comparison against a wider SS2D model is the cleanest check of whether the three-plane scan adds representational power or just capacity."],"forward_implications":["Mamba-based change detectors can treat the channel axis as a first-class scan direction without changing the linear-time complexity of the selective scan.","The full three-plane scan is the operative recipe: on the two ablation datasets, HW+HC and HW+WC two-plane variants can reduce F1 below HW alone, while the three-plane combination improves it.","Pairing the 3D scan with FFT and convolution branches in the decoder is what the authors use to get both global context and boundary detail.","Reported F1 gains over the strongest Mamba baseline are 0.39% on WHU-CD, 0.31% on LEVIR-CD, 0.21% on CDD, 1.89% on SYSU-CD, and 3.74% on DSIFN-CD."],"supporting_citations":[{"why":"Supplies the S6 selective scan primitive that 3D-SSM runs on each flattened plane.","marker":"[24]"},{"why":"Supplies the bidirectional VME scan operation reused in Eq. (2) to process each flattened sequence.","marker":"[27]"},{"why":"Provides the VMamba backbone and the SS2D spatial-scan baseline that 3D-SSM is designed to replace.","marker":"[28]"},{"why":"M-CD, the strongest Mamba-based baseline, which the proposed method outperforms in F1 on all five datasets.","marker":"[30]"},{"why":"Prior work scanning 3D tokens along spatial and spectral dimensions, the direct motivation for channel-axis scanning.","marker":"[51]"},{"why":"Provides the LEVIR-CD benchmark on which the method reports the best F1.","marker":"[59]"},{"why":"Provides the WHU-CD benchmark used for the main ablations.","marker":"[60]"},{"why":"Provides the CDD benchmark on which the method reports the best F1.","marker":"[61]"},{"why":"Provides the SYSU-CD benchmark used for the main ablations.","marker":"[62]"},{"why":"Provides the DSIFN-CD benchmark on which the method reports the largest F1 gain.","marker":"[63]"}],"fun_headline_variants":["3D scan catches subtle changes across five datasets","Channel-aware 3D scan boosts change detection","New module scans all planes for sharper change maps","3D-SSM: three-plane scan tops change detection","Scanning channels improves remote sensing change detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that running the selective scan along the height–channel and width–channel planes creates useful cross-channel dependencies, rather than simply adding parameters and computation.","fun_headline_variants_meta":{"raw":{"variants":["3D scan catches subtle changes across five datasets","Channel-aware 3D scan boosts change detection","New module scans all planes for sharper change maps","3D-SSM: three-plane scan tops change detection","Scanning channels improves remote sensing change detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000524,"raw_usage":{"total_tokens":2560,"prompt_tokens":1000,"completion_tokens":1560,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1488}},"tokens_in":616,"tokens_out":1560,"duration_ms":12332,"temperature":1.0,"reasoning_tokens":1488,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:06:54.412024+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model on WHU-CD and SYSU-CD with the HC and WC scan branches replaced by linear channel-mixing layers matched for parameter count and FLOPs, keeping the HW scan and all other modules unchanged; if the F1 score does not fall below the reported 95.30 and 84.90, the three-plane scanning geometry itself is not what carries the improvement.","supporting_citations":[{"cited_title":"Vmamba: Visual state space model,","cited_arxiv_id":null,"evidence_quote":"Provides the VMamba backbone and the SS2D spatial-scan baseline that 3D-SSM is designed to replace."},{"cited_title":"3DSS-Mamba: 3D- Spectral-Spatial Mamba for Hyperspectral Image Classification,","cited_arxiv_id":null,"evidence_quote":"Prior work scanning 3D tokens along spatial and spectral dimensions, the direct motivation for channel-axis scanning."},{"cited_title":"Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set,","cited_arxiv_id":null,"evidence_quote":"Provides the WHU-CD benchmark used for the main ablations."},{"cited_title":"Change detection in remote sensing images using conditional adver- sarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the CDD benchmark on which the method reports the best F1."},{"cited_title":"A deeply supervised attention metric-based network and an open aerial image dataset for remote sensing change detection,","cited_arxiv_id":null,"evidence_quote":"Provides the SYSU-CD benchmark used for the main ablations."}],"review_version":1}