{"id":"13764a2c-b921-47b5-945e-73052d519414","arxiv_id":"2607.23464","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual-domain Mamba network with edge-aligned directional scanning and NSCT subbands slightly beats recent PolSAR classifiers on four public datasets.","lead":"DA-Mamba combines direction-adaptive Mamba scanning with NSCT frequency decomposition and spatial Mamba for PolSAR land-cover classification. It reports small accuracy gains over recent baselines on standard radar benchmarks, mainly by better modeling anisotropic edges.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"Sub-0.4% SOTA margins rest on a single random split, and the authors' own ablation (Table 3) shows the full model losing to ablated variants on two of four datasets — both the superiority claim and the dual-domain attribution are statistically unsupported as reported.","rationale":"The reader identified exactly the right soft spot: single-split evaluation, missing error bars, sub-percent margins, and the Table 3 inconsistency where the full model is not best on San Francisco or Oberpfaffenhofen. I agree this is the load-bearing concern; I would only emphasize that the ablation reversal is the sharper half of it, because it undermines the architectural attribution (contributions 1–2), not merely the leaderboard ranking. Nothing here suggests bad faith — this is the standard empirical protocol in the subfield, and the paper is otherwise transparent about hyperparameters and does report sensitivity ablations. But the strong SOTA language (\"surpasses state-of-the-art\") is calibrated to margins the evidence cannot currently support. The remedy is cheap and standard: multi-seed/multi-split statistics on the contested datasets, plus an explicit statement of baseline provenance and patch-overlap handling. That is precisely what the reader's CONDITIONAL verdict conditions on, so I see no reason to move it: the claim is plausible, the design is reasonable, and the gap is evidentiary rather than conceptual. If the multi-seed test shows the full model's advantage is noise, CONDITIONAL should degrade toward REJECT of the SOTA claim (though the architecture itself may still be worth archiving); if the advantage survives, CONDITIONAL can rise to ACCEPT.","tokens_in":12506,"tokens_out":1921,"duration_ms":55265,"concrete_test":"Re-run the full model, the two ablated variants that beat it in Table 3 (San Francisco row 3: 99.70 vs 99.55 OA; Oberpfaffenhofen row 2: 95.83 vs 95.52 OA), and NGDiffSM under 5–10 independent random 5%/1% splits and seeds. Report mean ± std and a paired significance test per dataset. If the full model's edge over NGDiffSM and over its own ablations is not significant on at least 3 of 4 datasets, the SOTA and dual-domain-collaboration claims should be downgraded to \"competitive\"; if the full model wins consistently, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: (a) DA-Mamba-DDCL surpasses SOTA, and (b) the gains come from direction-adaptive scanning plus dual-domain collaboration. Both rest on numbers from a single random 5%/1% split with no repeated runs, no standard deviations, and no significance testing. The decisive margins over the strongest baseline (NGDiffSM) are +0.18% OA on Xi'an (Table 1) and +0.38% OA on Oberpfaffenhofen (Table 2) — well within the run-to-run variance typical for patch-based PolSAR classifiers at this training ratio, which the paper's own sensitivity analyses (Table 4: 8 vs 16 directions swings Xi'an OA by ~1.7%; Fig. 6: state dimension swings it ~1.1%) demonstrate is large relative to the claimed margin.\n\nMore damaging to claim (b): in Table 3 the full five-module model is not the best configuration on San Francisco (an ablated variant reaches OA 99.70 vs the full model's 99.55) or on Oberpfaffenhofen (a variant reaches 95.83 vs 95.52). The text concedes only that the full model \"remains competitive\" there, but contribution 2 and the abstract's \"dual-domain fusion strengthens discriminability\" claim require the complete design to be at least consistently non-inferior. With a single split, these reversals and the headline margins are indistinguishable from seed noise. A secondary fairness concern: the paper does not state whether baseline numbers were re-run under the identical split/protocol or transcribed from prior publications; with 64×64 overlapping patches and random pixel-level splits, any protocol mismatch (or uncorrected train/test patch overlap, which the paper does not address) would further erode the comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript proposes DA-Mamba-DDCL, a PolSAR image classification framework that couples (i) a spatial-domain 2D Mamba over raw polarimetric features with (ii) an NSCT-based frequency branch in which high-frequency directional subbands are serialized along their decomposition orientation (ζ_k(i,j) = i·cosθ_k + j·sinθ_k) and processed by a shared Mamba SSM, plus a lightweight low-frequency branch and a gated three-stream fusion. The authors report state-of-the-art results on four PolSAR benchmarks (Xi'an OA 97.17%, Oberpfaffenhofen OA 95.52%, with San Francisco and Flevoland relegated to supplementary), module ablations (Table 3), and sensitivity analyses over NSCT direction count, Mamba state dimension, and training ratio.","tokens_in":12904,"tokens_out":2790,"duration_ms":76578,"significance":"The direction-aware serialization idea is a clean and well-motivated match to NSCT structure: aligning the SSM scan order with each subband's principal orientation is a parameter-cheap modification with a clear physical rationale for anisotropic PolSAR scattering, and the SSF multi-directional fusion is a sensible complement. The paper is stronger than many architecture papers in this space in shipping a four-dataset evaluation, a five-configuration module ablation, and three hyperparameter sensitivity studies, and the authors commit to releasing code, which would make the pipeline (rotated-Sobel edge enhancement, permutation-based scans, gated fusion) reproducible in principle. If the reported gains are statistically robust, this would be a useful contribution to the growing Mamba-for-remote-sensing literature. However, as detailed below, the decisive margins over the strongest baseline are smaller than the run-to-run and hyperparameter variance the paper itself documents, so the central empirical claim is currently under-supported.","major_comments":[{"comment":"The headline superiority claim rests on a single random 5%/1% train/val split with no repeated runs, no standard deviations, and no significance testing. The decisive margins over the strongest baseline (NGDiffSM) are +0.18% OA on Xi'an and +0.38% OA on Oberpfaffenhofen. The paper's own sensitivity studies show variance of this order or larger from a single hyperparameter change: Table 4 shows direction count swinging Xi'an OA from 95.26% (4 directions) to 97.17% (8) to 95.48% (16), and Fig. 6 shows d_state moving Xi'an OA between 96.02% and 97.17%. Patch-based PolSAR classifiers at 5% labels routinely exhibit seed-to-seed OA fluctuations of several tenths of a percent. As reported, the SOTA margins are indistinguishable from seed noise. Please report mean ± std over at least 5 random splits/seeds for the proposed method and the top baselines, or an equivalent significance test.","section":"Experimental Setup / Tables 1–2"},{"comment":"The ablation results do not consistently support contribution 2 (the dual-domain collaborative design). The full five-module model is not the best configuration on San Francisco (OA 99.55 vs. 99.70 for the two-branch variant) or on Oberpfaffenhofen (OA 95.52 vs. 95.83 for the variant without one module). The text says only that the full model 'remains competitive' there, but the abstract and contribution 2 claim that dual-domain fusion 'strengthens discriminability' — a claim that requires the complete design to be at least non-inferior across datasets. Combined with Major Comment 1 (single split), these reversals may be noise, but then so may the gains. Either the ablation needs multi-run statistics, or the claims need to be qualified to 'best on Xi'an and Flevoland' with an honest discussion of where modules hurt.","section":"Ablation Analysis, Table 3"},{"comment":"The protocol for baseline numbers is not stated in the main text: it is unclear whether DFGCN, HybridCVNet, CV-MsAtViT, SpectralNet, S2Mamba, and NGDiffSM were re-trained under the identical 5%/1% split, patch size, and augmentation, or whether numbers were transcribed from prior publications under different protocols. This matters because SpectralNet and S2Mamba are hyperspectral methods whose adaptation to PolSAR inputs requires non-trivial choices, and because S2Mamba's Oberpfaffenhofen results (OA 74.91%, road class 8.06%) suggest a possible training failure rather than a fair comparison. Please state explicitly in the main text (not only the supplementary) that all baselines were re-run under the identical protocol, and describe the input adaptation for the HSI-derived baselines.","section":"Experimental Setup / Results"},{"comment":"With 64×64 patches and random pixel-level train/test partitioning, spatially overlapping patches straddling the train/test boundary are likely, which is known to inflate OA in patch-based PolSAR/HSI classification and can easily account for sub-0.5% effects of the size claimed here. The manuscript does not state whether any spatial-disjointness measure (buffer zones, block-wise splits, or non-overlapping patch extraction) was used. Please clarify the sampling procedure and, if overlapping patches are used, either add a spatially disjoint evaluation or explicitly acknowledge this as a limitation shared with the compared protocols.","section":"Experimental Setup"}],"minor_comments":[{"comment":"The abstract states 'Evaluated on three real-world PolSAR datasets,' but the Experimental Setup evaluates four (Xi'an, San Francisco, Flevoland, Oberpfaffenhofen). Please reconcile.","section":"Abstract"},{"comment":"Several references are duplicated with 'a'/'b' suffixes pointing to identical entries: Dong, Zhang, and Zou 2021a/2021b; Freeman and Durden 1998a/1998b; Zhou et al. 2016a/2016b; Lee, Grunes, and Kwok 1994a/1994b. Also, Yu et al. 2015 ('Learning cross space mapping via DNN using large scale click-through logs') is a web-search paper and is almost certainly the wrong citation for CNN-based PolSAR classification.","section":"References"},{"comment":"The SSF fusion equation introduces Softmax in the text ('Softmax(·) performs weight normalization') but the displayed formula F_hf^fused = Σ_k C(F_hf^(k)) ⊙ F_hf^(k) contains no normalization; please make the equation match the description. Similarly, the terminology switches between 'HF' and 'HL' subbands.","section":"Proposed Method, Multi-directional subbands fusion"},{"comment":"The conclusion claims superiority in 'computational efficiency,' but no FLOPs, parameter counts, or runtime comparisons appear anywhere in the manuscript. Either add an efficiency table or remove the claim.","section":"Conclusion"},{"comment":"The polarimetric input X_ori is never fully specified: the number of channels C_ori and whether the input is the scattering vector, coherency/covariance matrix elements, or a decomposition feature set should be stated in the main text rather than deferred to the supplementary.","section":"Proposed Method / Experimental Setup"},{"comment":"Typos and figure issues: 'muti-directional'; Fig. 6 is referenced as 'Fig. 6(b)' though no panels are described; 'data sets' (Table 3 caption) vs 'datasets'; inconsistent capitalization of 'Direction-adaptive Mamba' vs 'DA-Mamba' throughout.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The strongest baseline, NGDiffSM (Shi et al. 2025), is the authors' own prior work, so the key comparison is against their own pipeline under presumably matched conditions — this is fine, but it makes the absence of multi-run statistics for exactly that comparison more conspicuous. The manuscript is formatted for AAAI 2027 and repeatedly defers essential material (dataset descriptions, baseline implementation details, San Francisco/Flevoland results) to a supplementary that was not available to this reviewer; the editor may wish to verify its contents. The fit is more natural for a remote-sensing venue (TGRS/JSTARS) than a general AI venue, but that is an editorial decision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The usable takeaway is a concrete recipe: NSCT into 8 HF directions, edge-boost with rotated Sobel, serialize each subband by projection onto its angle, run shared Mamba, fuse with SSF, keep a parallel spatial 2D-Mamba on the raw PolSAR cube, then gate the three streams. That assembly is new even though every piece (NSCT/contourlet PolSAR nets, multi-dir scan, S2Mamba-style spatial SSM, gated fusion) already exists in the cited literature. They implement it cleanly, give the usual hyper-parameters, promise code, and run the standard four public sets against six recent baselines including their own NGDiffSM.\n\nWhat works: the direction-aware serialization is a sensible, low-overhead way to respect NSCT anisotropy instead of forcing a fixed 2-D raster; the edge-guided HF residual is simple and ablations show it helps on most sets; qualitative maps look cleaner at boundaries, which is the stated motivation. Sensitivity tables on K, d_state and label ratio are present and sensible (K=8, d_state=32).\n\nSoft spots are real but proportional. Decisive OA edges over NGDiffSM are +0.18 % (Xi’an) and +0.38 % (Oberpfaffenhofen) on a single 5 %/1 % random split with no multi-seed means or error bars. The paper’s own sweeps already show 1–1.7 % swings from K or d_state, so those margins sit inside ordinary run variance. Worse for the dual-domain story: Table 3 shows the full five-module model is not best on San Francisco or Oberpfaffenhofen; ablated variants win. The text only says “remains competitive,” yet the abstract and contribution 2 claim the complete design strengthens discriminability. Baseline protocol match (identical split, patch overlap handling) is also unstated. None of this makes the method useless; it just means the strong SOTA language and clean attribution are not yet supported.\n\nFor anyone already building PolSAR SSMs or frequency-domain classifiers this is worth a look and a re-run once code drops. It is incremental engineering, not a conceptual shift. I would send it to referees: the architecture is clear enough and the experiments are real; they just need multi-seed stats and tempered claims. I would not cite the numbers until those appear, but I might cite the scanning idea.","headline":"Solid engineering combo of oriented Mamba on NSCT bands plus spatial Mamba, but the SOTA and attribution claims rest on sub-0.4% single-split margins that the authors’ own ablations undercut.","tokens_in":13949,"tokens_out":602,"would_cite":false,"duration_ms":13691,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Direction-adaptive Mamba plus spatial-frequency fusion beats prior PolSAR classifiers by following edge orientations instead of fixed scans.","keywords":["PolSAR classification","Mamba","direction-adaptive scanning","NSCT","spatial-frequency fusion","anisotropic scattering","state space model"],"falsifier":"Re-run the exact same training protocol with multiple independent 5 percent draws (or published multi-run means and standard deviations) on Xi’an and Oberpfaffenhofen; if the mean OA of DA-Mamba no longer exceeds NGDiffSM and the other listed baselines, the central superiority claim fails.","tokens_in":13535,"feed_emoji":"📡","tokens_out":887,"duration_ms":19315,"temperature":0.7,"pith_summary":"Polarimetric SAR images encode how surfaces scatter radar energy in different orientations, so land-cover classes often differ by thin, direction-sensitive edges rather than by smooth blobs. Existing Mamba models either stay purely spatial and blur those edges, or use fixed raster scans that ignore the preferred direction of each texture. This paper claims that decomposing the image with a non-subsampled contourlet transform, then letting a shared Mamba scan each high-frequency subband along its own edge orientation, while a second spatial Mamba keeps raw scattering context, and a gate fuses the three streams, yields more accurate class maps. On standard benchmarks the full pipeline reports higher overall accuracy and cleaner boundaries than recent CNN, Transformer, diffusion, and Mamba baselines. A sympathetic reader cares because the same linear-complexity backbone can now respect the anisotropic physics of PolSAR without quadratic attention cost.","feed_headline":"Mamba that follows edges beats fixed-scan PolSAR classifiers","feed_subtitle":"Orientation-aware scans on contourlet subbands plus dual-domain fusion lift accuracy and clean up boundaries","key_machinery":"Direction-adaptive Mamba (DA-Mamba): each NSCT high-frequency subband is serialized by projecting pixels onto that subband’s principal angle, then processed by a shared selective state-space model so long-range context flows along the true edge orientation rather than a fixed raster.","core_discovery":"The authors establish that an edge-aligned, direction-adaptive Mamba operating on NSCT high-frequency subbands, paired with a spatial 2D Mamba on raw polarimetry and a gated dual-domain fusion, produces more discriminative PolSAR features and higher classification accuracy than fixed-scan or single-domain Mamba and than listed state-of-the-art methods.","pith_inferences":["The same orientation-aware serialization could be dropped into other multi-scale decompositions (shearlets, dual-tree wavelets) without redesigning the Mamba core.","Because margins are small and few-shot generalization is already flagged as weak, the practical next test is whether the direction-adaptive prior still helps under 1 percent labels or cross-sensor transfer.","If the gain is truly from matching scan direction to scattering anisotropy, similar adaptive scans should help single-pol SAR and optical edge-heavy segmentation where fixed Mamba scans currently dominate."],"forward_implications":["PolSAR classifiers can keep Mamba’s linear complexity while still modeling anisotropic scattering and weak boundaries.","NSCT multi-directional high-frequency subbands become a natural substrate for orientation-aware state-space scanning rather than only for CNN or wavelet front-ends.","Gated fusion of raw spatial scattering with low- and high-frequency contourlet streams is a reusable pattern for other dual-domain remote-sensing tasks.","Edge-aligned serialization can replace fixed four-direction 2D scans whenever the data have known preferred orientations."],"fun_headline_variants":["Edge-aligned Mamba beats fixed-scan PolSAR classifiers","DA-Mamba fuses NSCT subbands with adaptive scans for PolSAR","Direction-adaptive Mamba sharpens PolSAR boundaries via dual domains","Spatial-frequency Mamba lifts PolSAR accuracy over fixed scans","Edge-following Mamba plus contourlet fusion tops PolSAR methods"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A single random 5 percent labeled split is treated as a stable enough basis to claim superiority even when absolute gains over the strongest baseline are a few tenths of a percent and the full model is not best on every dataset in the authors’ own ablations.","fun_headline_variants_meta":{"raw":{"variants":["Edge-aligned Mamba beats fixed-scan PolSAR classifiers","DA-Mamba fuses NSCT subbands with adaptive scans for PolSAR","Direction-adaptive Mamba sharpens PolSAR boundaries via dual domains","Spatial-frequency Mamba lifts PolSAR accuracy over fixed scans","Edge-following Mamba plus contourlet fusion tops PolSAR methods"]},"model":"grok-4.5","effort":"low","cost_usd":0.003977,"raw_usage":{"total_tokens":1262,"prompt_tokens":790,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":39768000,"prompt_tokens_details":{"text_tokens":790,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":397,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":790,"tokens_out":75,"duration_ms":8687,"temperature":1.0,"reasoning_tokens":397,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T21:33:39.225396+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the exact same training protocol with multiple independent 5 percent draws (or published multi-run means and standard deviations) on Xi’an and Oberpfaffenhofen; if the mean OA of DA-Mamba no longer exceeds NGDiffSM and the other listed baselines, the central superiority claim fails.","supporting_citations":[],"review_version":1}