{"id":"56ac08c1-92d6-45ec-87a6-7c2eb288740b","arxiv_id":"2509.00677","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A cross-state Mamba architecture that fuses HSI and LiDAR features via token and state exchange reports state-of-the-art accuracy on MUUFL and Houston2018.","lead":"This paper proposes CSFMamba, a Mamba-based neural network that fuses hyperspectral and LiDAR remote sensing data by exchanging classification tokens and state parameters between the two modalities. The authors report slightly higher classification accuracy and lower parameter/FLOP counts than Transformer and prior Mamba baselines on two public datasets, but provide no code, error bars, or train/test split details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority claim rests on uncontrolled baseline comparison and single-run numbers; reported margins (~0.5–1%) could vanish under a matched protocol.","rationale":"The reader's weakest assumption correctly identifies the lack of a controlled comparison protocol as the principal threat to the central claim. My analysis confirms that the reported gains are small relative to plausible cross-paper variation: on MUUFL the improvement over MACN is 0.47% OA, and on Houston2018 the improvement over MSFMamba is 1.00% OA. These margins could easily be reversed by a different split or sampling strategy, especially since HSI/LiDAR classification accuracy is highly sensitive to the number of training samples per class. The paper also omits any error bars or repeated-run statistics, so the result may not even be stable. I considered alternative concerns—such as the untested 'cross-state' fusion design or the FLOPs calculation—but they are secondary to the fact that the headline empirical comparison is not currently interpretable. The verdict CONDITIONAL is therefore appropriate, and no change to the reader's verdict is needed. A controlled reproduction with fixed splits and multiple seeds would settle the concern.","tokens_in":6307,"tokens_out":3878,"duration_ms":48864,"concrete_test":"Re-implement CSFMamba and at least the strongest baselines (MACN on MUUFL; MSFMamba and MACN on Houston2018) under one fixed protocol: use the same training samples per class (e.g., 30 for MUUFL, 200 for Houston2018), the same test set, and the same patch size. Run each method 5 times with different random seeds and report mean±std OA/AA/Kappa. If CSFMamba's mean accuracy does not exceed the best baseline by more than the pooled standard deviation on both datasets, the claimed superiority is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CSFMamba outperforms all baselines in OA/AA/Kappa on MUUFL and Houston2018. Tables I and II import baseline numbers from prior papers (EndNet, CoupledCNN, FGCN, HCT, MACN, MSFMamba), but Section III.A does not specify the train/validation split ratio, the number of training samples per class, or the patch sampling strategy. Different HSI papers commonly use very different splits (e.g., 200 labeled samples per class vs. 30 samples per class), which can shift OA by several points. The accuracy margins over the strongest baselines are small: +0.47% over MACN on MUUFL and +1.00% over MSFMamba on Houston2018. Without a controlled reimplementation under a single protocol, these differences could be due to split mismatch rather than architecture. Additionally, no standard deviations or repeated runs are reported, so even under a fixed protocol the margins could be within run-to-run noise. Table III's FLOPs/parameter comparison is similarly unanchored: the input size and calculation method for baselines are not given, making the burden advantage unverifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CSFMamba, a multimodal fusion network for HSI/LiDAR remote sensing image classification. The architecture combines CNN-based local feature extraction with a simplified Mamba encoder, and introduces a Cross State Fusion Mamba block in which CLS tokens and the SSM parameter matrices B and C are exchanged between modalities to achieve cross-modal state-space fusion. Experiments on MUUFL and Houston2018 report state-of-the-art overall accuracy, average accuracy, and kappa, while using 0.219M parameters and 12.29M FLOPs. The paper also includes an ablation study and a brief burden analysis.","tokens_in":6594,"tokens_out":3215,"duration_ms":43949,"significance":"If the reported results are reproducible, the cross-state fusion idea is a genuinely interesting adaptation of SSM/selective state-space models to multimodal fusion, and the linear complexity claim is attractive for large remote sensing scenes. The paper explicitly targets a known limitation of plain Mamba—its inability to directly perform feature fusion—and proposes a concrete architectural remedy. However, the evidence supporting the central superiority claim is currently weakened by a poorly specified experimental protocol, single-run results, and unverified baseline comparisons.","major_comments":[{"comment":"The experimental protocol is underspecified. Section III.A gives dataset dimensions and hyperparameters but omits the train/validation/test split ratio, the number of labeled samples per class, and the patch sampling strategy. Table I and Table II report single-run OA/AA/Kappa with no standard deviations or significance tests, and all baseline numbers are imported from prior publications. The claimed margins are small: +0.47 percentage points over MACN on MUUFL and +1.00 over MSFMamba on Houston2018. Under a different split or with run-to-run variation, these margins could vanish. The authors should specify the exact data split, report results over multiple random seeds (mean±std), and either reimplement baselines under the same protocol or provide a matched-protocol comparison. This is load-bearing for the paper's main claim of superiority.","section":"Section III.A and Tables I-II"},{"comment":"The text says there are 200 iterations in total with batch size 256. Taken literally, this is only about 51,200 training samples, which is a very small fraction of the Houston2018 image (1202×4768 pixels). This seems inconsistent with the high reported accuracy and suggests the term 'iterations' may mean 'epochs' or that a much smaller training set is used than the reader would expect. The authors must clarify the training schedule and dataset partition. This detail directly affects whether the comparisons with baselines trained under different schedules are meaningful.","section":"Section III.A, 'training process has 200 iterations'"},{"comment":"The burden analysis is not anchored to a concrete protocol. FLOPs and parameter counts depend on input tensor size, patch size, sequence length, and implementation details. The table reports values for CoupledCNN, FGCN, HCT, MACN, and Ours, but does not state the input size or the FLOPs calculation method for each model. The claim that 'the training and testing time is shorter' is also unsupported by any timing measurement. To make the complexity advantage verifiable, the authors should specify the exact input configuration and, ideally, provide measured runtime on the same hardware.","section":"Table III"},{"comment":"The cross-state fusion mechanism needs a precise formulation. In Eq.(7), B and C are generated from X^cross_L,t while the state update uses X^cross_H,t. The dimensions and ordering of these operations are not defined: it is unclear whether Eq.(7) follows Eq.(6) directly, whether the same procedure is applied symmetrically for the LiDAR branch, and how the two modalities' states are interleaved across the L layers. Additionally, B is used in Eq.(1) as an output dimension and in Eqs.(3)-(7) as an SSM parameter; this overloading makes the derivation hard to follow. Please clarify the shapes and provide a precise pseudocode or diagram of the fusion-SSM.","section":"Section II.D, Eq.(7)"}],"minor_comments":[{"comment":"Equation (4) uses (∆A)^{-1}(exp(∆A)-I), which is the standard zero-order hold discretization but can be numerically unstable for small ∆A. In practice this is often computed with a scaled matrix exponential; please clarify whether a numerically stable implementation is used.","section":"Equations (3)-(4)"},{"comment":"The symbol B is used for both the linear-layer output dimension and the SSM input matrix. Use a different symbol (e.g., d_model) for the dimension to avoid ambiguity.","section":"Eq.(1) and Eq.(5)"},{"comment":"The unit 'FLOPs(M)' is ambiguous; it should be written as 'MFLOPs' or '×10^6 FLOPs' for consistency with common usage.","section":"Table III"},{"comment":"The paper says CSFMamba 'simplifies the structure of Mamba by removing Conv1D and a projection function.' This design choice is not justified or ablated. It would be useful to include an ablation variant that retains these components, since removing them may affect the reported performance.","section":"Section II.C"},{"comment":"No code or data availability statement is provided. Given the protocol sensitivity of the results, releasing code and exact split indices would substantially strengthen reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the cross-state fusion idea is worth considering, but the empirical evidence in its current form is not sufficient to support the strong claims. The missing split details, single-run results, and unverified baseline numbers are the main blockers. If the authors can provide a matched-protocol comparison, multiple runs with variance, and clarify the training schedule, the paper could become acceptable. The novelty relative to MSFMamba [18] should also be more crisply positioned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: CSFMamba is a reasonable next step in applying Mamba to multimodal remote sensing. The cross-state fusion idea—exchanging CLS tokens and feeding the other modality's tokens into the SSM's parameter generation (Eq. 7)—is not in the cited baselines, and the ablation suggests it carries the architecture. If I work in this area, I'd want to test that block myself.\n\nThe paper does a few things well. The combination of CNN for local features and a simplified Mamba for long-range context is sensible. The claimed resource footprint (0.219M params, 12.29M FLOPs) is noticeably smaller than the quoted CNN/Transformer baselines, which matters for deployment. The ablation table is useful evidence that the cross-state fusion block contributes, even if the numbers are single-run.\n\nThe soft spot is the experimental reporting, and it's not minor. Section III.A gives dataset sizes and a few hyperparameters but omits the train/validation/test split, the number of training samples per class, and the patch sampling strategy. The baseline numbers in Tables I and II are imported from earlier papers; we don't know if they were produced under the same protocol. The accuracy margins over the strongest baselines are 0.47% on MUUFL and 1.00% on Houston2018. Without matched protocols and some measure of run-to-run variance, those differences could easily be protocol mismatch or noise. The FLOPs/parameters comparison in Table III is also unanchored: input size and measurement methodology for baselines are not given.\n\nThat said, I don't think the paper is trying to hide anything. The architecture is described clearly, the math is standard, and the ablation is honest. The central problem is that the superiority claim is stated too strongly for the evidence. A revision that adds split details, per-class sample counts, multiple runs with standard deviations, and ideally a re-implementation of at least one baseline under the same protocol would make this a solid contribution.\n\nWho is this for? Researchers working on efficient multimodal remote sensing classification, especially those interested in Mamba-based architectures. It's a step-forward engineering result, not a conceptual breakthrough.\n\nFor peer review: yes, I'd send it out. The cross-state fusion idea deserves referee time, and the reviewers can push for the missing experimental details. Without those details, I wouldn't accept it as-is; but it's a legitimate paper that with revision could be citable.\n\nReading group: maybe, if your group works on state-space models for vision. Otherwise, it's not a must-read.","headline":"A workmanlike Mamba variant for HSI/LiDAR fusion with a genuinely new cross-state mechanism, but the accuracy claims rest on under-specified experimental protocols and small margins.","tokens_in":7107,"tokens_out":2794,"would_cite":true,"duration_ms":31824,"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":"The paper claims Mamba's state-space operator can be turned into a multimodal fusion mechanism by cross-feeding each sensor's state parameters, and that the network beats CNN, Transformer, and direct Mamba baselines on HSI-LiDAR classificat","keywords":["multimodal remote sensing classification","hyperspectral image","LiDAR","state space model","Mamba","cross-state fusion","land cover classification","linear complexity"],"falsifier":"Run all compared baselines and CSFMamba on a single fixed train/validation split of MUUFL and Houston2018 with identical patch size, augmentation, and sample counts; then run CSFMamba with the cross-state B/C exchange replaced by simple concatenation. If the accuracy margins over MSFMamba and MACN vanish under matched protocols, or if the cross-state-to-concatenation gap disappears, the central superiority claim fails.","tokens_in":6168,"feed_emoji":"🛰️","tokens_out":7872,"duration_ms":84254,"temperature":0.7,"pith_summary":"CSFMamba tries to show that the Mamba state-space operator, originally a single-stream sequence encoder, can also be the place where two remote sensing modalities are fused. It proposes a cross-state block in which hyperspectral and LiDAR branches exchange their CLS tokens and, more aggressively, one branch's input-dependent state parameters condition the other branch's hidden-state update. Paired with CNN feature extraction and an information-preprocessing stage, the network reports the highest accuracy among the compared CNN, Transformer, and direct Mamba methods on MUUFL and Houston2018, with the smallest parameter and FLOP budget. The practical payoff claimed is a lighter, faster multimodal classifier that still captures long-range spatial-spectral context.","feed_headline":"Cross-state Mamba fusion beats CNN and Transformer baselines","feed_subtitle":"HSI-LiDAR: 91.13% OA on MUUFL and 93.38% on Houston2018, at 0.219M parameters.","key_machinery":"The load-bearing object is the Cross State Fusion Mamba (CSFM) block. It takes the CLS token learned in each unimodal Mamba encoder, splices it into the other modality's token set, and feeds the combined tokens through a Mamba block in which the input-dependent SSM matrices B and C are generated from the other modality. This mixes the sequence state from one sensor with the contextual parameters of the other, converting the hidden state into a cross-modal fusion variable. The block is iterated over L layers and is paired with a simplified Mamba core that removes the Conv1D and one projection to cut the computational burden.","core_discovery":"The central claim is that Mamba's selective state space is not only a cheaper replacement for self-attention but a usable fusion operator. Each modality is tokenized with a CLS token that absorbs the modality's global abstract features through a Mamba encoder. Then, in the Cross State Fusion Mamba block, the CLS token of one modality is spliced into the other modality's sequence, and the combined token sequence is processed by a Mamba layer whose B and C parameters are generated from the opposite modality, so the hidden state of HSI is updated under parameters supplied by LiDAR, and vice versa. The authors argue that this intercrossing of internal state spaces resolves Mamba's inability to f","pith_inferences":["The cross-state trick is not obviously limited to two sensors: a chain of CLS-token handoffs among HSI, LiDAR, SAR, or other spectral data is a natural extension, though the paper does not test it.","One direct experiment to isolate the mechanism would be to keep token splicing but disable the cross-generation of B and C, forcing fusion to happen only through the CLS token; the accuracy gap would show how much of the gain comes from state-parameter exchange rather than token pooling.","A matched train/validation split and per-class sample counts for every baseline would let a reader separate architecture gains from protocol effects, since the paper specifies only patch size, batch size, optimizer, and iteration count."],"forward_implications":["On the two reported datasets, Mamba-based cross-state fusion reaches higher OA, AA, and kappa than CNN, Transformer, and the direct multimodal Mamba baseline, with 0.219M parameters and 12.29M FLOPs.","Ablations show each design element contributes: removing the preprocessing stage drops OA from 91.13% to 89.68%, removing the CNN feature extraction drops it to 88.61%, and removing the cross-state block drops it to 82.94%.","Because the SSM formulation is linear in sequence length, the fusion layer's cost scales linearly with image patches rather than quadratically, which is the paper's main reason for expecting scalability to larger scenes.","The classification output averages the CLS tokens of both modalities, so the final decision is explicitly forced to use fused information rather than a single modality's features."],"supporting_citations":[{"why":"Supplies the selective state-space (Mamba) block that CSFMamba strips down and reuses as the base encoder.","marker":"[14]"},{"why":"Gives the SSM discretization equations (A, B, C, Delta) that the cross-state update in Eq. 7 inherits.","marker":"[15]"},{"why":"The direct multimodal Mamba baseline (MSFMamba) that CSFMamba is compared against on Houston2018 and aims to improve by adding cross-state fusion.","marker":"[18]"},{"why":"The Transformer-based HCT baseline whose quadratic attention is the computational burden the paper argues Mamba avoids.","marker":"[12]"},{"why":"The MACN baseline combining self-attention and convolution; it is the strongest non-Mamba comparison and motivates the unified fusion design.","marker":"[13]"},{"why":"Coupled CNNs establish the two-branch feature extraction pattern and provide a CNN baseline on both datasets.","marker":"[8]"}],"fun_headline_variants":["Mamba fusion beats transformers at 0.219M params","Cross-state Mamba fuses HSI and LiDAR efficiently","Mamba's hidden state becomes a fusion operator","CSFMamba: cross-state fusion for HSI-LiDAR","Mamba fusion: 91% OA on MUUFL, 93% on Houston"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The comparison's load-bearing premise is that every baseline was trained and validated under the same protocol; the paper reports patch size, batch size, optimizer, and iteration count but not the train/validation split ratio or per-class sample counts, so reported gains could in principle be protocol effects rather than architecture effects.","fun_headline_variants_meta":{"raw":{"variants":["Mamba fusion beats transformers at 0.219M params","Cross-state Mamba fuses HSI and LiDAR efficiently","Mamba's hidden state becomes a fusion operator","CSFMamba: cross-state fusion for HSI-LiDAR","Mamba fusion: 91% OA on MUUFL, 93% on Houston"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000431,"raw_usage":{"total_tokens":2058,"prompt_tokens":789,"completion_tokens":1269,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1189}},"tokens_in":533,"tokens_out":1269,"duration_ms":11655,"temperature":1.0,"reasoning_tokens":1189,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:19:38.452952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run all compared baselines and CSFMamba on a single fixed train/validation split of MUUFL and Houston2018 with identical patch size, augmentation, and sample counts; then run CSFMamba with the cross-state B/C exchange replaced by simple concatenation. If the accuracy margins over MSFMamba and MACN vanish under matched protocols, or if the cross-state-to-concatenation gap disappears, the central superiority claim fails.","supporting_citations":[{"cited_title":"MSFMamba: Multi-Scale Feature Fusion State Space Model for Multi-Source Remote Sensing Image Classification","cited_arxiv_id":"2408.14255","evidence_quote":"The direct multimodal Mamba baseline (MSFMamba) that CSFMamba is compared against on Houston2018 and aims to improve by adding cross-state fusion."},{"cited_title":"Joint classification of hyperspectral and lidar data using a hierarchical cnn and transformer,","cited_arxiv_id":null,"evidence_quote":"The Transformer-based HCT baseline whose quadratic attention is the computational burden the paper argues Mamba avoids."},{"cited_title":"Mixing self-attention and convolution: A unified framework for multi- source remote sensing data classification,","cited_arxiv_id":null,"evidence_quote":"The MACN baseline combining self-attention and convolution; it is the strongest non-Mamba comparison and motivates the unified fusion design."},{"cited_title":"Classification of hyperspectral and lidar data using coupled cnns,","cited_arxiv_id":null,"evidence_quote":"Coupled CNNs establish the two-branch feature extraction pattern and provide a CNN baseline on both datasets."}],"review_version":1}