{"id":"6595de20-7465-4d8c-a1f1-41545a06c29a","arxiv_id":"2501.04944","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A pixel-level Mamba model with spatial and spectral branches reports state-of-the-art hyperspectral image classification on four datasets.","lead":"MambaHSI applies the Mamba state-space model to classify hyperspectral images at the full-image level, using separate spatial and spectral branches plus a learned fusion. The authors report higher accuracy than CNN, transformer, and graph-based baselines on four public datasets, with linear complexity in pixel count.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The spatial Mamba block's causal 1D raster scan with no positional encoding makes the claimed pixel-level long-range spatial modeling untested; the accuracy gains may not be caused by spatial reasoning.","rationale":"The paper is a solid empirical study: four public datasets, ten repeats, ablations, complexity measurements, and code. The strongest claim, however, is not just 'this model works' but 'this model works because Mamba captures long-range spatial dependencies at pixel level with linear complexity.' That causal explanation depends on SpaMB actually exploiting 2D spatial structure. The architecture as described in Eq. 7 flattens the image to a single raster sequence and applies a standard causal Mamba. Mamba has no built-in positional encoding, and the causal scan means each output token can only attend to its prefix. Without coordinates or a designed scan, the model has no way to represent up/down/left/right relations; it only sees an arbitrary 1D order. This is an internal plausibility gap, not a disagreement with community consensus. The empirical gains are credible but under-specified; the reader's conditions (add other Mamba baselines, significance tests, encoder depth) are all reasonable. My proposed test of permuting pixel order is the cleanest way to falsify or validate the spatial-modeling assumption. If the model is order-insensitive, the paper's central narrative collapses even though the accuracy numbers may stand. I therefore keep the CONDITIONAL verdict, adding this spatial-order verification as an explicit condition, so no change to the reader's verdict is needed.","tokens_in":24100,"tokens_out":7211,"duration_ms":71512,"concrete_test":"Run the PaviaU and Houston experiments under three SpaMB orderings: (1) the original raster scan, (2) a fixed random permutation of pixel positions, and (3) a bidirectional snake or Hilbert 2D scan, keeping all other settings fixed. If OA under the random permutation does not drop by more than the reported run-to-run standard deviation (about 1%), the spatial branch is effectively order-insensitive and the 'long-range spatial modeling' explanation is vacuous. If random permutation drops accuracy but a 2D scan improves on raster, the design is order-dependent but the current choice is suboptimal. Additionally, report a paired significance test (e.g., McNemar or paired t-test over the ten runs) between MambaHSI and the best competitor on each dataset.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the claim that SpaMB (Eq. 7) models long-range interactions of the whole image at pixel level, and that this is why MambaHSI achieves SOTA. But Eq. 7 flattens the 2D image into a 1D sequence of H*W tokens and feeds it to a standard causal Mamba with no positional encoding and no 2D scan strategy. In a causal scan, the representation at token t can only see tokens before t in raster order; no pixel can see its right or lower neighbors. With no coordinate information, the model cannot know the 2D location of any pixel, so the raster order is arbitrary. The reported accuracy gains are therefore not attributable to 'long-range spatial dependencies' from this architecture alone. It is possible the gains come from the spectral branch, the fusion module, or a global bag-of-pixels effect rather than spatial modeling. The paper's central narrative, that Mamba's linear-complexity long-range modeling makes it a next-generation backbone, rests on an unvalidated assumption about what SpaMB actually encodes. Missing Mamba baselines and significance tests further weaken the SOTA claim, but the spatial-sequence assumption is the deepest unresolved issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MambaHSI, an image-level hyperspectral image classification architecture built entirely on state space models. Each pixel is embedded individually; a spatial Mamba block processes all pixels as a rasterized 1D sequence, a spectral Mamba block processes grouped spectral vectors, and a spatial-spectral fusion module combines the two branches with learned scalar weights. The authors claim that this is the first image-level SSM-based HSI classifier, that it models long-range interactions of the whole image at pixel level with linear complexity, and that it outperforms ML-, GCN-, CNN-, and Transformer-based baselines on Pavia University, Houston, HanChuan, and HongHu. The experimental section provides ten-run comparisons, ablations, hyperparameter studies, and a scaling experiment comparing Mamba with self-attention.","tokens_in":24345,"tokens_out":7570,"duration_ms":73540,"significance":"If the empirical results hold, the paper is a useful contribution: an image-level, linear-complexity backbone for HSI classification with public code and evaluation on four datasets. The complexity scaling experiment in Table VIII is a clean demonstration of linear versus quadratic behavior, and the ablation and hyperparameter studies are thorough. The significance is currently limited by the lack of validation of the spatial long-range mechanism, the absence of any Mamba-based baseline in the comparison, and the lack of statistical significance tests, so the 'next-generation backbone' claim is not yet established.","major_comments":[{"comment":"The spatial Mamba block flattens the entire 2D image into a raster-scan 1D sequence and applies a standard causal Mamba without positional encoding. Because the scan is causal, token t cannot see any later token, so no pixel receives information from its right or lower neighbors; without coordinate information, the 2D arrangement is only an arbitrary ordering. The paper's recurring claim that SpaMB models 'long-range interaction of the whole image at the pixel level' is therefore not established by the architecture. The reported gains could plausibly come from the spectral branch, the fusion module, or a global bag-of-pixels effect. Please provide a control experiment, for example randomizing the pixel permutation, using bidirectional or multi-directional scans, or replacing SpaMB with a non-spatial sequence model of the same capacity, and show that the spatial ordering specifically is responsible for the accuracy gain.","section":"Section IV-C, Eq. (7), and Section V-B"},{"comment":"No statistical significance tests are reported. On PaviaU, MambaHSI's OA is 95.74±0.90 versus 94.97±1.43 for CLOLN, a difference of 0.77 percentage points, which is within one standard deviation of both means. The claim that the model 'outperforms all ... and achieves SOTA on all datasets' is too strong for this dataset. Please report paired tests (for example, McNemar's test or pairwise comparisons across the ten trials) and adjust the wording accordingly, or justify that the differences are meaningful.","section":"Section V-B, Tables I and V"},{"comment":"The related work cites several Mamba-based HSI classifiers, including SpectralMamba [57] and 3DSS-Mamba [58], but none of these are included in the experimental comparison. Because the paper's central contribution is a Mamba architecture for HSI, the absence of any Mamba baseline makes the 'superiority' claim incomplete. Please add at least one patch-level Mamba method evaluated under the same training/validation split, or explain with runtime and accuracy evidence why the image-level comparison omits them; in either case, the SOTA statement should be qualified to 'among the compared methods'.","section":"Section II.B and Section V.A comparison list"}],"minor_comments":[{"comment":"The notation A and B is reused for both the continuous and discrete parameters; please disambiguate these symbols.","section":"Section III, Eq. (2)"},{"comment":"There are several grammatical slips, including 'we divide a spectral features into G group' and 'the semantic vector'; please correct these throughout.","section":"Section IV.D and other places"},{"comment":"The table header and row labels do not clearly indicate which row corresponds to the Mamba encoder block and which to the self-attention variant; please add explicit row labels.","section":"Table VIII"},{"comment":"The 'w SSI' OA on PaviaU is 80.67±2.07, while the SpeMB row in Table V reports 78.34±4.00 for what appears to be the same configuration; if these are different configurations or different numbers of runs, please clarify in the text.","section":"Section V-F, Table IX"},{"comment":"The patch-size ablation plots show lines without error bars even though the text says multiple runs were performed; please add error bars or explain why they are omitted.","section":"Section V-E, Fig. 11"},{"comment":"There is a typo in the header, 'Paramters', which should be 'Parameters'.","section":"Table VI"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is solid and the code is available, but the main narrative overclaims the spatial mechanism and the SOTA status. The missing Mamba baselines and significance tests are fixable, and the spatial-mechanism concern can be addressed with control experiments. I do not see grounds for rejection, but the paper needs another round of experiments and careful rewriting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a competent, well-executed empirical paper: four datasets, ten trials, ablations, complexity analysis, code. Second, the central narrative—that the spatial Mamba block provides pixel-level long-range spatial modeling—is not actually tested by the experiments as designed.\n\nThe novelty is real but modest: it's the first image-level (whole-image) Mamba for HSI classification, using pixel embeddings instead of patches, a group-wise spectral Mamba, and two learned scalar fusion weights. The results show consistent OA gains over the best non-Mamba baselines they compare with, especially on HanChuan and HongHu.\n\nThree soft spots. The most serious is the spatial Mamba block: it flattens the whole image into a 1D raster scan and runs a standard causal Mamba with no positional encoding. In that setup, a pixel cannot see its right or lower neighbors, and the model has no way to know where it is in the 2D grid. So the accuracy gains are not demonstrably attributable to spatial long-range reasoning; they could come from the spectral branch, the fusion weights, or simply a global bag-of-pixels effect. The paper's title and abstract hinge on the spatial claim, so this needs to be addressed with scan-direction ablations or positional information.\n\nSecond, the comparison omits the prior Mamba-based HSI methods cited in the related work (SpectralMamba, 3DSS-Mamba, etc.). Since the claim is that SSM is the next-generation backbone, they need to beat the strongest same-family baselines.\n\nThird, no significance tests. On PaviaU the gain over CLOLN (95.74 vs 94.97) is within one standard deviation, so \"outperforms\" is not established by chance. A paired test across their ten trials would settle it.\n\nMinor: encoder depth isn't specified in the setup, and the SSFM ablation is thin, though the component ablations overall are useful.\n\nThe paper deserves a serious referee. The architecture is sensible, the experiments are reproducible, and the claims are mostly clear. It needs a revision adding Mamba baselines, significance testing, and a spatial-scan ablation—not a desk reject.","headline":"Useful empirical Mamba-for-HSI paper; the spatial long-range claim is untested because of a causal raster scan with no positional encoding, and the Mamba baselines are missing—still worth reviewing after experiments.","tokens_in":24865,"tokens_out":2940,"would_cite":true,"duration_ms":28367,"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":"MambaHSI claims to be the first image-level, Mamba-based hyperspectral classifier, reaching state-of-the-art accuracy on four benchmarks with linear computational complexity.","keywords":["hyperspectral image classification","state space models","Mamba","spatial-spectral fusion","image-level classification","linear computational complexity","long-range dependencies"],"falsifier":"Train the spatial-only branch (or the full model) with two different fixed pixel orderings, for example normal raster order versus a random but fixed permutation of all pixels, and compare per-pixel accuracy on the same training and test splits. If accuracy does not change, the spatial branch is not actually using spatial adjacency; if accuracy collapses, the reported gains depend on the specific arbitrary ordering rather than on a learned spatial representation.","tokens_in":23872,"feed_emoji":"🛰️","tokens_out":4878,"duration_ms":41886,"temperature":0.7,"pith_summary":"The paper proposes MambaHSI, a hyperspectral image classifier built entirely from state-space layers, and argues that this design can model long-range spatial and spectral dependencies while scaling linearly with the number of pixels. It claims this is the first image-level, Mamba-based HSI classification model, taking the whole image as input rather than patches, and reports that it outperforms the ML, GCN, CNN, and Transformer baselines on the Pavia University, Houston, HanChuan, and HongHu datasets. The motivating point is that patch-based Transformers pay quadratic attention cost and lose pixel-level detail, while CNNs cannot reach far enough spatially. If the claim holds, whole-scene hyperspectral classification becomes feasible at pixel resolution with modest compute.","feed_headline":"MambaHSI classifies whole hyperspectral images in linear time","feed_subtitle":"A spatial-spectral state-space model labels every pixel at once and beats patch-based baselines on four benchmarks.","key_machinery":"The load-bearing object is the Mamba layer, a selective state-space sequence model that processes a one-dimensional sequence through a hidden state with input-dependent gating, in time linear in sequence length. MambaHSI applies it twice: the spatial Mamba block flattens the image into a raster scan of pixels to capture spatial context, and the spectral Mamba block partitions the spectral dimension into $G$ groups and treats the groups as a sequence to capture spectral continuity. The fusion module computes $H_{fus} = H_i + w_{spa}H_{spa}^o + w_{spe}H_{spe}^o$ with learned scalar weights $w_{spa}$ and $w_{spe}$, and residual connections are added throughout. The whole encoder is built only from these blocks, which is what makes the claimed linear complexity and end-to-end whole-image training possible.","core_discovery":"The central claim is that a pure state-space (Mamba) backbone, with separate spatial and spectral branches and an adaptive fusion step, gives the best hyperspectral classification accuracy on four diverse benchmarks. The spatial branch flattens the entire image into a one-dimensional sequence of pixel embeddings and runs a Mamba layer over it; the spectral branch splits each pixel's spectrum into groups, orders the groups, and runs a Mamba layer across them; the fusion module learns two scalar weights to combine the branches. The paper reports that this arrangement surpasses all compared ML-, GCN-, CNN-, and Transformer-based methods on all four datasets, with a single forward pass over the whole image and computational complexity that scales linearly in pixel count.","pith_inferences":["Beyond the paper, a natural next test is whether the raster-scan ordering actually matters: a multi-directional or bidirectional scan variant might improve accuracy further, since the paper does not ablate the scan direction.","The learned fusion weights $w_{spa}$ and $w_{spe}$ are not reported; inspecting them per dataset could reveal whether the spatial branch dominates in every scene or whether the balance shifts with ground-object scale.","The same two-branch state-space design could transfer to other dense prediction tasks, such as multispectral land-cover mapping or semantic segmentation, where linear scaling with image size is a practical advantage.","A stronger control experiment would shuffle pixels with a fixed random permutation at training time; the paper does not run this test, so the attribution of gains to 'long-range spatial modeling' remains unverified."],"forward_implications":["Whole-scene hyperspectral classification can be performed in a single forward pass with complexity $O(L)$ in the number of pixels, making very large images tractable.","Pixel-level tokens, rather than patches, improve classification accuracy; the paper shows that increasing patch size degrades performance on the Pavia University dataset.","Modeling the spectrum as an ordered sequence of groups helps classification; setting the group count to one, which removes spectral ordering, drops overall accuracy by roughly 5.6 points on Pavia University.","Adaptive fusion of spatial and spectral branches with learned scalar weights outperforms simple summation, especially on the HongHu dataset, where it adds about 5.5 points of overall accuracy."],"supporting_citations":[{"why":"Supplies the Mamba layer used as the basic unit and the linear-complexity, hardware-aware implementation claim.","marker":"[24]"},{"why":"Establishes the S4 structured state-space foundation that Mamba extends and that the paper's preliminaries build on.","marker":"[22]"},{"why":"Defines the SpectralFormer Transformer baseline that the paper compares against and whose patch-based approach it contrasts.","marker":"[16]"},{"why":"Is the GSC-ViT Transformer baseline that the paper claims to outperform on all datasets.","marker":"[12]"},{"why":"Introduces image-level hyperspectral classification via fully convolutional networks, the precedent for whole-image input that MambaHSI follows.","marker":"[3]"},{"why":"Provides the WHU-Hi HanChuan and HongHu datasets used in evaluation.","marker":"[65]"},{"why":"Is the vision Mamba variant that the paper explicitly says it does not use, to show the basic Mamba layer suffices.","marker":"[19]"}],"fun_headline_variants":["MambaHSI: linear-time classification for entire hyperspectral scenes","State-space model labels every pixel of full HSI in one pass","First image-level Mamba for hyperspectral classification","MambaHSI beats all baselines on four HSI datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's spatial power rests on the assumption that reading the whole image as one flat line of pixels, in fixed raster order with no positional encoding and no scan-direction design, still lets the Mamba layer learn genuinely spatial long-range relationships.","fun_headline_variants_meta":{"raw":{"variants":["MambaHSI: linear-time classification for entire hyperspectral scenes","State-space model labels every pixel of full HSI in one pass","First image-level Mamba for hyperspectral classification","MambaHSI beats all baselines on four HSI datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001132,"raw_usage":{"total_tokens":4717,"prompt_tokens":970,"completion_tokens":3747,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":3676}},"tokens_in":586,"tokens_out":3747,"duration_ms":25855,"temperature":1.0,"reasoning_tokens":3676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:21:33.058752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the spatial-only branch (or the full model) with two different fixed pixel orderings, for example normal raster order versus a random but fixed permutation of all pixels, and compare per-pixel accuracy on the same training and test splits. If accuracy does not change, the spatial branch is not actually using spatial adjacency; if accuracy collapses, the reported gains depend on the specific arbitrary ordering rather than on a learned spatial representation.","supporting_citations":[{"cited_title":"Spectralformer: Rethinking hyperspectral image classification with transformers,","cited_arxiv_id":null,"evidence_quote":"Defines the SpectralFormer Transformer baseline that the paper compares against and whose patch-based approach it contrasts."},{"cited_title":"Hyperspectral image classification using groupwise separable convolutional vision transformer network,","cited_arxiv_id":null,"evidence_quote":"Is the GSC-ViT Transformer baseline that the paper claims to outperform on all datasets."},{"cited_title":"Beyond the patchwise classification: Spectral-spatial fully convolutional networks for hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Introduces image-level hyperspectral classification via fully convolutional networks, the precedent for whole-image input that MambaHSI follows."}],"review_version":1}