{"id":"0a36786d-6cc5-44b0-be16-bdc3057a5958","arxiv_id":"2412.00302","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"HSLiNet, a CNN-based HSI-LiDAR fusion model with forward and backward spectral convolutions and a spatial block, reports 96.68% OA on Houston 2013 and claims state-of-the-art efficiency.","lead":"A neural network called HSLiNet fuses hyperspectral and LiDAR data using bidirectional 1D convolutions over spectral bands plus a spatial branch, reporting 96.68% accuracy on the Houston 2013 benchmark. The result is not yet verifiable because the authors provide no code, no error bars, and an unfinished manuscript with inconsistent naming and citations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on an unstated train/test protocol: with no split, patch size, or per-class counts, HSLiNet's 2.49-point OA gain over S2ENet cannot be distinguished from a protocol difference. A re-run under a locked common protocol is needed.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the central claim is an empirical superiority claim, and its truth depends entirely on a fair, identical evaluation protocol. The manuscript does not specify the split, patch size, per-class training counts, or epochs used for the comparison in Table I, so the reported margins cannot be checked or reproduced. This is not a mere stylistic omission; without a fixed protocol, the 96.68% OA figure has no well-defined meaning relative to the baselines. The internal numerical inconsistency between Table III's text and the table itself (0.9736 vs. 0.9668 OA) and the mislabeled FusAtNet citation further weaken trust in the reported numbers. I agree with the reader's REJECT verdict: the architecture may be a plausible extension of dual-CNN and Mamba-style processing, but the evidence presented does not support the state-of-the-art claim. No ad hominem is intended; the issue is verifiability of the experimental comparison, not author intent.","tokens_in":7211,"tokens_out":3952,"duration_ms":36089,"concrete_test":"Conduct a controlled reproduction: fix one standard Houston2013 split with explicit per-class training sample counts, fix identical patch size, training epochs (e.g., 100), optimizer, and batch size for all methods; train HSLiNet and at least S2ENet and MDLCross using author-provided code (or request it from the authors), and compare per-class accuracies, OA, AA, and Kappa. If HSLiNet's OA no longer exceeds S2ENet's by a margin robust to random seeds, or if the 2.49-point gap disappears, the SOTA claim is a protocol artifact. Additionally, verify Table I's FusAtNet column against the original FusAtNet publication, since the citation in the table points to the S2ENet reference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim ('HSLiNet demonstrated the highest OA at 96.68%, AA at 97.32%, Kappa at 96.39%, significantly outperforming other methods,' Section III-C) is a comparative result. For it to be true, HSLiNet and the six methods in Table I must have been trained and evaluated under an identical Houston2013 protocol. The manuscript never states the training/test split, per-class training sample counts, patch size, or number of epochs used for the comparison. Section III-A says only that counts are 'based on standard training and test data quantity requirements'; Section III-B gives batch size and learning rate but not the patch dimension or the split used for Table I. Since baseline per-class accuracies shift substantially with training budget (e.g., TwoBranch C10 at 65.44 vs. MDLCross at 91.41), a 2.49-point OA gap between HSLiNet and S2ENet is within the range of protocol variation. This concern is reinforced by internal inconsistencies: Table III's text reports Model 1 OA=0.9736/AA=0.9786/Kappa=0.9713 while Tables II and III list 0.9668/0.9722/0.9639; Table I labels FusAtNet as [14], which is the S2ENet reference; and the model name changes between HSLiNet, FusedBiRNet, Hi-Vim, and SS non-linear Model. These features make it impossible to verify from the manuscript alone whether the comparison was fair.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HSLiNets, a dual non-linear feature learning network that fuses hyperspectral image (HSI) and LiDAR data using bidirectional (forward-backward) convolutional pathways and a spatial analysis block. The authors claim state-of-the-art classification accuracy on the Houston 2013 dataset (OA 96.68%, AA 97.32%, Kappa 96.39%) while avoiding the computational cost of self-attention and transformer architectures. The paper also presents an ablation study and a limited efficiency discussion. However, the experimental protocol is critically under-specified, the reported results contain internal contradictions, and the baseline comparisons are difficult to verify due to inconsistent citations and missing details.","tokens_in":7612,"tokens_out":3921,"duration_ms":35767,"significance":"If the performance and efficiency claims were fully validated, the architecture could offer a practical alternative to attention-based fusion models for HSI-LiDAR classification in resource-constrained settings. The idea of learning bidirectional spectral dependencies with a delta-modulated non-linearity is interesting, and the paper explicitly states that attention mechanisms are not required. However, the current manuscript does not substantiate these claims: the missing experimental protocol, the inconsistent numbers between tables, and the unclear baseline definitions mean that the central result cannot be confirmed from the text alone. The contribution as presented is therefore not yet at the standard expected for a journal publication.","major_comments":[{"comment":"The paper never specifies the exact training/test split, patch size, per-class sample counts, or number of epochs used for the main comparison. Section III-A only says the counts follow 'standard training and test data quantity requirements,' and Section III-B gives batch size and learning rate but omits the patch dimension and the training duration for Table I. Without a fully specified protocol, the claimed 2.49-point OA improvement over S2ENet cannot be distinguished from a protocol difference, especially since baseline per-class accuracies vary widely across methods. This is a load-bearing issue for the central SOTA claim.","section":"Section III-A and III-B; Table I"},{"comment":"The ablation results are internally inconsistent: Table II lists Model 1 with OA=0.9668, AA=0.9722, Kappa=0.9639, while the text in Section III-D and the Table III caption report OA=0.9736, AA=0.9786, Kappa=0.9713 for the same Model 1. These numbers cannot both be correct, and the discrepancy directly affects the paper's headline result. The authors must resolve this inconsistency and provide a single, verified set of metrics.","section":"Section III-D, Tables II and III"},{"comment":"The baseline citations are inconsistent and therefore the comparison set is unclear. In the text, FusAtNet is cited as [7], but in Table I the FusAtNet column is labeled [14], which is the S2ENet reference; TwoBranch is also labeled [7] in the table, yet [7] is a different paper (a multisource CNN classification paper). This makes it impossible to know which specific implementations were compared and whether the baselines are the intended state-of-the-art models.","section":"Section III-C and Reference list"},{"comment":"The transformation parameters A and B, and the delta parameter, are not sufficiently defined. The text describes A and B as capturing forward and backward spectral dependencies, but it never states their shape, initialization, whether they are learned, or how they are applied in Eqs. (3)-(4) (elementwise multiplication is implied but not stated). The 'delta-expanded' variable is also unexplained. This lack of detail prevents reproduction of the core mechanism.","section":"Section II-A and II-C, Eqs. (3)-(4)"},{"comment":"The efficiency claim is unsupported. The paper states that testing time is 'always 0.48 seconds' but provides no comparison with any baseline, no parameter counts, no FLOPs, and no memory measurements. The claim of 'linear memory usage' in the conclusion is not demonstrated. Without quantitative comparisons on runtime or complexity, the contribution of reduced computational demand cannot be evaluated.","section":"Section III-C and Fig. 3"}],"minor_comments":[{"comment":"The model name is inconsistent: it is called HSLiNet, FusedBiRNet, 'SS non-linear Model,' and 'Hi-Vim' in different places. Please use one consistent name throughout.","section":"Throughout"},{"comment":"The dataset description contains an empty citation '[]' in the Houston 2013 paragraph; a proper reference to the GRSS Data Fusion Contest dataset is needed.","section":"Section III-A"},{"comment":"The reference to 'TwoBranch CNN []' has an empty citation; also the sentence 'HSLiNet again led with a score of 97.54, outperforms' is grammatically incomplete.","section":"Section III-C"},{"comment":"The term 'kapa' appears instead of 'Kappa' in the conclusion, and the model is called 'Hi-Vim' only in the conclusion; please unify terminology and fix the typo.","section":"Section III-D and Conclusion"},{"comment":"The figure is referenced but the relationship between epochs and OA is not described in detail; also the statement 'testing time is always 0.48 seconds' should clarify whether this is per epoch, per patch, or for the entire dataset.","section":"Figure 3"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft with multiple inconsistencies: contradictory reported metrics, missing citations, empty references, and an undefined experimental protocol. The central SOTA claim cannot be verified independently, and the internal contradictions raise concerns about the reliability of the reported results. Even a major revision would need to re-run the experiments under a fully specified, shared protocol and correct the numerical inconsistencies; given the current state, I cannot recommend publication. The paper would also benefit from a careful technical rewrite before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is a reasonable incremental contribution, but the paper as written cannot support its headline SOTA claim. The Houston 2013 comparison is missing the exact split, patch size, per-class training counts, and error bars, so the 2.49-point OA gain over S2ENet could easily be protocol variation.\n\nWhat is actually new: a forward/backward 1D convolution over spectral bands, paired with a separate spatial branch, is a legitimate extension of existing dual-CNN and SSM-style fusion models. The ablation in Table II does show that each component (forward, reversed, spatial) contributes to the final accuracy, which is the kind of evidence that matters. The reported 0.48s test time is a concrete efficiency data point, though it is only a single number with no context on hardware or batch size.\n\nThe soft spots are substantial. The manuscript never states the Houston 2013 train/test split, patch dimension, number of epochs for the main comparison, or per-class training sample counts. The reader's stress-test is on point: without a locked protocol, the reported per-class accuracies for baselines (e.g., TwoBranch C10 at 65.44 vs. MDLCross at 91.41) are impossible to interpret. There are also clear internal inconsistencies: FusAtNet is [7] in the text but [14] in Table I; the model is called HSLiNet, FusedBiRNet, Hi-Vim, and SS non-linear Model in different places; Table III's paragraph reports OA=0.9736/AA=0.9786/Kappa=0.9713 while the table itself lists 0.9668/0.9722/0.9639; and the ablation section contains a copy-pasted paragraph about Highway and Railway classes. No code is provided, so none of these issues can be resolved by inspection.\n\nThe core idea may have merit, but the paper is not ready for publication. The comparison is not verifiable, the writing is sloppy enough to undermine trust, and the efficiency claim rests on a single measurement. This is not a case where the math is wrong or the method is incoherent; it is a case where the empirical evidence is not presented in a way that supports the claims.\n\nWho is this for? A researcher working on efficient HSI-LiDAR fusion might find the bidirectional spectral convolution worth trying on their own benchmark, but they would need to reimplement it from scratch. The paper could become a credible workshop or short paper after major revision: release code, use the standard split or the official training set, add error bars, fix the reference and naming issues, and correct Table III.\n\nMy recommendation for an editor: desk reject in current form. It does not deserve referee time until the authors provide the missing protocol details and code. If a revised version addresses those issues, it could be worth a look.","headline":"Plausible incremental architecture, but the central SOTA claim is unverifiable from the manuscript because the training/test protocol is unspecified and the text has multiple internal inconsistencies.","tokens_in":8112,"tokens_out":2355,"would_cite":false,"duration_ms":23627,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual reversed CNN with a spatial block claims top Houston 2013 accuracy, 96.68%, without self-attention.","keywords":["hyperspectral image classification","LiDAR fusion","dual non-linear feature space","bidirectional reversed CNN","Houston 2013 dataset","multi-sensor fusion","efficient deep learning"],"falsifier":"Re-run the Table I benchmark with a fixed patch size and identical per-class training counts on the standard Houston 2013 split; if HSLiNet's overall accuracy does not remain above the strongest baseline's 94.19%, the superiority claim is falsified.","tokens_in":7021,"feed_emoji":"🛰️","tokens_out":10563,"duration_ms":90648,"temperature":0.7,"pith_summary":"The paper tries to show that fusing hyperspectral and LiDAR data does not require a large attention-based backbone. It proposes HSLiNet, which forms a dual non-linear feature space from two reversed CNN scans over fused patches plus a spatial branch, and reports that this light architecture reaches 96.68% overall accuracy, 97.32% average accuracy, and a Kappa coefficient of 96.39% on the Houston 2013 dataset, beating six published fusion baselines. The significance, if the result holds, is that high-accuracy multimodal remote sensing classification can be achieved with low memory and a 0.48-second test time, which would make the method practical in GPU-limited or real-time settings.","feed_headline":"No-attention CNN fusion reaches 96.68% on Houston 2013","feed_subtitle":"Dual forward-backward CNN paths claim top accuracy with low compute, making fusion feasible on smaller GPUs.","key_machinery":"The carrying mechanism is the Fused BiNetworks Block (B-Net) paired with the LiDAR Spatial Feature Processing Block (S-Block). On a fused patch, one path computes $x_{\\text{forward}} = f_{\\text{activation}}(\\operatorname{Conv1d}(x_{\\text{proj}}))$ while the reversed path computes $x_{\\text{backward}} = f_{\\text{activation}}(\\operatorname{Conv1d}(z_{\\text{proj}}, \\text{reversed}))$; each direction then applies a delta-modulated tanh state update with learnable matrices $A$ and $B$, and the two resulting states are reduced and summed into $h_{\\text{combined}}$. The S-Block independently processes the same patch through $\\operatorname{ReLU}(\\operatorname{BatchNorm}(\\operatorname{Conv2d}(\\cdot)))$ to produce $h_{l_{\\text{processed}}}$, and the fusion is $y_{\\text{fusion}} = \\operatorname{concat}(h_{\\text{combined}}, h_{l_{\\text{processed}}})$, followed by a linear classifier. The double sweep in opposite spectral directions is the module that replaces self-attention, letting the model read spectral dependencies both forward and backward without a global attention matrix.","core_discovery":"The authors claim that a deliberately small network can outperform attention-based fusion models when hyperspectral and LiDAR patches are fused early and then passed through two reversed 1D convolutional pathways. The forward path uses a transform matrix $A$ to accumulate spectral dependencies in normal band order, while the backward path uses a transform $B$ in reverse order; after the delta-modulated tanh update $h = \\tanh(x + A\\,\\Delta_{\\text{expanded}})$ and its backward counterpart, the two reduced hidden states are summed. A separate spatial block applies $\\operatorname{ReLU}(\\operatorname{BatchNorm}(\\operatorname{Conv2d}(\\cdot)))$ and its output is concatenated with the bidirectional state to form the vector sent to the classifier. On Houston 2013 the full model reports the best overall accuracy (96.68%), average accuracy (97.32%), and Kappa (96.39%) among the methods in Table I, and the ablation attributes the gain to the joint action of the forward, backward, and spatial components.","pith_inferences":["If the reversed dual-pass structure is what carries the gain, the same block could be transplanted to other paired 1D-2D sensor tasks, such as multispectral plus synthetic-aperture radar or spectral plus depth, without redesigning the fusion module.","Because the delta-modulated update resembles a discretized state transition, sweeping the number of reversed passes and the choice of nonlinearity would show whether the gain saturates quickly, which would clarify whether the mechanism behaves more like a state-space model than like plain convolution.","A direct transfer test on a second fusion benchmark with different land-cover classes would show whether the margin over the strongest baseline is specific to Houston 2013 or generalizes."],"forward_implications":["Attention is not a necessary ingredient: the model's 96.68% overall accuracy is the best in Table I without any self-attention module, so reversed convolutional passes can substitute for attention in HSI-LiDAR fusion.","Fusion is what earns the accuracy: the same architecture reaches 96.68% overall accuracy with both modalities, 96.13% with HSI alone, and 27.38% with LiDAR alone, so LiDAR contributes almost nothing by itself but lifts the fused result.","Every module pays: removing both directional passes drops overall accuracy by 3.26 points, removing the spatial block drops it by 0.83 points, and removing only the backward path costs 0.53 points.","The method is cheap: testing takes 0.48 seconds and training time grows only with the number of epochs, so the reported accuracy comes with a runtime profile that fits resource-constrained settings."],"supporting_citations":[{"why":"It supplies the cross-attention FusAtNet baseline that HSLiNet is compared against in the Houston 2013 experiments.","marker":"[7]"},{"why":"It supplies the encoder-decoder EndNet fusion baseline that appears in Table I.","marker":"[5]"},{"why":"It supplies the MDL-Middle and MDL-Cross intermediate- and cross-attention fusion baselines used in the comparison.","marker":"[13]"},{"why":"It supplies the S2ENet baseline, the strongest previous method whose 94.19% overall accuracy HSLiNet's 96.68% is measured against.","marker":"[14]"}],"fun_headline_variants":["Dual reversed CNNs hit 96.68% on Houston 2013","Efficient HSI-LiDAR fusion: reversed CNN paths top 96.68%","No-attention dual CNN fusion wins Houston 2013","96.68% accuracy with low-compute HSI+LiDAR fusion","Reversed 1D CNN pathways outperform Transformers on HSI+LiDAR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim rests on all six comparison models being trained and evaluated under exactly the same Houston 2013 data split, patch size, and training budget, and the paper does not state those numbers.","fun_headline_variants_meta":{"raw":{"variants":["Dual reversed CNNs hit 96.68% on Houston 2013","Efficient HSI-LiDAR fusion: reversed CNN paths top 96.68%","No-attention dual CNN fusion wins Houston 2013","96.68% accuracy with low-compute HSI+LiDAR fusion","Reversed 1D CNN pathways outperform Transformers on HSI+LiDAR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001015,"raw_usage":{"total_tokens":4268,"prompt_tokens":908,"completion_tokens":3360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":3258}},"tokens_in":524,"tokens_out":3360,"duration_ms":22539,"temperature":1.0,"reasoning_tokens":3258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:31:40.998678+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table I benchmark with a fixed patch size and identical per-class training counts on the standard Houston 2013 split; if HSLiNet's overall accuracy does not remain above the strongest baseline's 94.19%, the superiority claim is falsified.","supporting_citations":[{"cited_title":"Multisource remote sensing data classification based on convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"It supplies the cross-attention FusAtNet baseline that HSLiNet is compared against in the Houston 2013 experiments."},{"cited_title":"Deep encoder– decoder networks for classification of hyperspectral and lidar data,","cited_arxiv_id":null,"evidence_quote":"It supplies the encoder-decoder EndNet fusion baseline that appears in Table I."},{"cited_title":"More diverse means better: Multimodal deep learning meets remote- sensing imagery classification,","cited_arxiv_id":null,"evidence_quote":"It supplies the MDL-Middle and MDL-Cross intermediate- and cross-attention fusion baselines used in the comparison."},{"cited_title":"S 2enet: Spatial–spectral cross-modal en- hancement network for classification of hyperspectral and lidar data,","cited_arxiv_id":null,"evidence_quote":"It supplies the S2ENet baseline, the strongest previous method whose 94.19% overall accuracy HSLiNet's 96.68% is measured against."}],"review_version":1}