{"id":"4cba4018-1e79-48b7-a815-22ef23312c0a","arxiv_id":"2504.13045","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"EKGNet combines a context-aware mapping network with K dynamically weighted 3D expert convolution kernels in a 3D-DenseNet, reporting overall accuracies around 99.8 to 99.98 percent on Indian Pines, Pavia University, and KSC under the authors' chosen splits.","lead":"This paper proposes EKGNet, a hyperspectral image classifier that generates input-specific 3D convolution kernels from a pool of expert kernels guided by global image context. The authors report near-perfect classification accuracy on three standard hyperspectral datasets, though the comparison setup is incompletely controlled.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own comparison tables contradict its headline: EKGNet is numerically below LGCNet on both Indian Pines (99.84 vs 99.85) and Pavia University (99.98 vs 99.99), and no KSC comparison table is provided.","rationale":"I read the paper in good faith as a proposal of a dynamic-kernel 3D-DenseNet for hyperspectral image classification, with the central claim being empirical superiority over mainstream methods on IN, UP, and KSC. For that claim to hold, the reported experiments would need to show EKGNet ahead of the baselines under a controlled and reproducible protocol. That condition is not met. The strongest failure is not merely an unstated baseline protocol; it is the internal contradiction in the reported numbers. On Indian Pines, Table 8 gives LGCNet OA=99.85 and EKGNet OA=99.84, so EKGNet is lower. On Pavia University, Table 9 gives LGCNet OA=99.99 and EKGNet OA=99.98, again lower. The abstract and conclusion nevertheless assert superiority on both datasets, and on KSC, which has no comparison table at all. A claim that is contradicted by the paper's own evidence cannot support acceptance, and this is independent of any speculation about how the baselines were configured. The reader's weakest_assumption identifies the uncontrolled baseline protocol as the key issue; I agree that this is a serious problem, but I would locate the more decisive problem in the internal table contradiction, so my agreement is partial. I also considered whether the dynamic-kernel mechanism itself could be the weakest point, but the mechanism is a standard input-dependent weighted combination of base kernels, and the paper does not claim a formal proof of superiority; the empirical comparison is the load-bearing evidence. The manuscript has no released code, no formal verification, and several mechanical inconsistencies, including a missing KSC table and a placeholder figure reference; these do not by themselves prove the method is wrong, but they erode confidence in the numerical claims. The verdict should remain REJECT: the central claim is not supported by the supplied evidence, and the only path to a positive verdict would be a controlled, reproducible comparison with matched training ratios, patch sizes, and data splits, plus a complete KSC table. That path is not present in the current preprint, so no change to the reader's verdict is needed.","tokens_in":10695,"tokens_out":4069,"duration_ms":38208,"concrete_test":"First, directly verify the reported OA rows: Table 8 shows LGCNet OA=99.85 vs EKGNet OA=99.84 on Indian Pines, and Table 9 shows LGCNet OA=99.99 vs EKGNet OA=99.98 on Pavia University; this already falsifies the unqualified 'outperforming mainstream' claim. To test whether a corrected comparison could rescue it, rerun LGCNet, DGCNet, SSRN, 3D-CNN, 3D-SE-DenseNet, and EKGNet from released code under the exact EKGNet protocol: 60% training on Indian Pines, 50% training on Pavia University and KSC, 15x15 overlapping patch extraction, identical validation/test splits, and identical input size and epoch budget, and produce the missing KSC comparison table.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that EKGNet 'demonstrates superior performance on IN, UP, and KSC datasets, outperforming mainstream hyperspectral image classification approaches.' The evidence in Section 3.4 does not support this claim and, taken at face value, contradicts it. In Table 8, on Indian Pines, LGCNet reports OA=99.85±0.04 while EKGNet reports OA=99.84; in Table 9, on Pavia University, LGCNet reports OA=99.99±0.00 while EKGNet reports OA=99.98. Thus EKGNet is not superior to LGCNet on either of the two datasets for which a comparison table is given, and the differences are within the reported standard deviations, so even a claim of statistical parity is not established. The abstract and conclusion also claim superiority on KSC, but no KSC comparison table appears anywhere in the manuscript. This is load-bearing because the paper's contribution is an empirical performance claim; the architecture is a combination of known dynamic-convolution and DenseNet ideas, and without a valid superiority result there is no demonstrated reason to prefer EKGNet. The comparison-protocol concern raised by the reader is also real: Section 3.2.1 selects training ratios 6:1:3 on Indian Pines and 5:1:4 on Pavia and KSC, and Section 3.2.2 selects 15x15 neighboring blocks, but the baseline methods' training ratios, patch sizes, and data splits are not reported. However, the more immediate problem is internal: even under the authors' own reported numbers, the headline statement is not true as written. The manuscript also contains unresolved placeholders (e.g., 'Figure ??'), inconsistent input-size statements between Section 3.2.2 and Section 3.4, and a mention of wavelet convolution in Section 1 that never appears in the method; these reinforce the impression that the empirical evidence was not carefully checked, but the table contradiction alone is sufficient to invalidate the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes EKGNet, a hyperspectral image classification network built on a modified 3D-DenseNet. A context-aware mapping network pools global spatial-spectral information and outputs attention weights over K base 3D convolutional kernels; the weighted combination of these kernels forms a per-sample dynamic convolution. The DenseNet backbone is modified with an exponentially increasing growth rate and fully dense connectivity. The method is evaluated on Indian Pines, Pavia University, and KSC, with reported overall accuracies above 99.8% on all three datasets, and the abstract and conclusion claim that EKGNet outperforms mainstream hyperspectral classification approaches.","tokens_in":10953,"tokens_out":5658,"duration_ms":48914,"significance":"The architecture is coherent, and the dynamic kernel aggregation in Section 2.2.2 is clearly specified, which is a strength. The proposed growth-rate rule k=2^{m-1}k0 is simple, and the fully dense connectivity is a reasonable architectural modification. If the empirical results were obtained under controlled conditions, the paper would offer a useful incremental contribution to hyperspectral image classification. However, the significance currently rests entirely on the claimed superiority over baselines, and that claim is contradicted by the reported numbers; without a valid comparison the contribution reduces to an unvalidated combination of known dynamic-convolution and DenseNet components. No code is provided, which further limits reproducibility of the empirical claims.","major_comments":[{"comment":"The headline claim that EKGNet 'demonstrates superior performance' on IN, UP, and KSC is not supported by the paper's own experimental tables. On Indian Pines, Table 8 reports OA=99.84 for EKGNet versus OA=99.85±0.04 for LGCNet; on Pavia University, Table 9 reports OA=99.98 for EKGNet versus OA=99.99±0.00 for LGCNet. EKGNet is therefore numerically below the leading baseline on both datasets for which comparison tables are provided, and no standard deviation is reported for EKGNet, so statistical equivalence is not established either. The abstract and conclusion also claim superiority on KSC, but no KSC comparison table is present anywhere in the manuscript. This is load-bearing because the paper's contribution is an empirical performance claim; the claim must be revised or supported by a KSC table and matched statistical comparisons.","section":"Abstract, Section 3.4 (Tables 8 and 9)"},{"comment":"The comparison with SSRN, 3D-CNN, 3D-SE-DenseNet, SpectralFormer, LGCNet, and DGCNet is not controlled. The authors selected the training ratio (6:1:3 on Indian Pines; 5:1:4 on Pavia University and KSC) and the patch size (15) by sweeping on the same datasets and reporting the resulting OA, but the training ratios, patch sizes, and data splits used for the baselines are never reported. If the baselines were evaluated under the standard smaller training sets, the reported margins in Tables 8 and 9 could reflect training-data volume and patch geometry rather than the proposed dynamic-kernel mechanism. Please rerun all baselines under exactly the same split, patch extraction, and evaluation protocol, or clearly state the protocol used for each baseline.","section":"Sections 3.2.1 and 3.2.2"},{"comment":"The experimental protocol is internally inconsistent and missing details needed for reproducibility. Section 3.2.1 states that KSC uses a 5:1:4 ratio with 11x11 neighboring blocks, while Section 3.2.2 states that size 15 was selected for all datasets; Section 3.4 additionally reports input sizes of 17x17 for all three datasets. The manuscript must state the exact patch size used in the final experiments and explain the relation between the 15x15 blocks selected in Tables 4-6 and the 17x17 inputs used for evaluation. Also, the patch extraction mechanism is not described: if neighboring blocks are extracted by sliding a window with stride 1, training and test blocks share spatial context, which can inflate accuracy; please specify the stride and confirm that no training pixel appears in a test block.","section":"Section 3.2.2, Section 3.4"}],"minor_comments":[{"comment":"The context-aware mapping network is said to be implemented in the 'keys module', but this module is never defined; the number of residual blocks N, the hidden channel widths, the initial temperature, and the annealing schedule for tau are not specified, which prevents reproduction.","section":"Section 2.2.1"},{"comment":"There is an unresolved cross-reference: the text says 'Figure ?? shows the loss and accuracy changes during training and validation', but no figure number is given; Figure 8 is subsequently titled as training/validation curves.","section":"Section 3.4"},{"comment":"Contribution 3 states that EKGNet improves representation capability 'through wavelet convolution with expanded receptive fields', but no wavelet convolution appears in the architecture description in Section 2; this claim appears to be inherited from another paper and should be removed or supported.","section":"Introduction, contribution 3"},{"comment":"There are spelling and naming inconsistencies, including 'EGKNet' in the conclusion, 'LGCNET' in Tables 8 and 9 versus 'LGCNet' in the text, and 'DGCNET' versus 'DGCNet'; these should be unified.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is that the paper's own tables contradict its headline; even a sympathetic reader cannot verify superiority without a KSC table and matched baseline protocols. I would support acceptance only after those empirical fixes. The manuscript also leans heavily on the authors' own prior arXiv preprints; please ensure novelty relative to references [7], [9], [10], and [11] is explicitly stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline claim of this paper doesn't survive contact with its own tables. EKGNet is a dynamic-convolution 3D-DenseNet where a small mapping network turns global average pooling into per-example weights that mix K base kernels. That's a known dynamic-convolution trick, and the DenseNet modifications (exponential growth, full dense connectivity) are minor. What is new here is the combination, and the description is competent. If the empirical section were controlled, this might be a modest engineering contribution.\n\nIt isn't controlled. The abstract says EKGNet 'demonstrates superior performance' on IN, UP, and KSC. Table 8 shows LGCNet at 99.85 OA on Indian Pines while EKGNet gets 99.84. Table 9 shows LGCNet at 99.99 on Pavia University while EKGNet gets 99.98. So on the two datasets with comparison tables, EKGNet is numerically below LGCNet and within error bars. No KSC table appears at all. That is a load-bearing internal contradiction.\n\nThere are additional problems. The training ratio and patch size are chosen by sweeping on the same datasets that then produce the headline numbers; baseline conditions (train ratio, patch size, splits) are not reported. Overlapping windows likely leak spatial context. The manuscript has a 'Figure ??' placeholder, says the Indian Pines input is padded to 155×155 and block size 15 is selected, then later says the input size is 17×17×200; and contribution claim 3 mentions wavelet convolution with expanded receptive fields, but the method never uses wavelets. No code is provided.\n\nNone of this is fraud. The architecture is coherent and the writing is understandable. But a paper whose central empirical claim is contradicted by its own comparison tables, with missing baseline protocols, does not meet the bar for serious review. A revised version with matched training budgets, per-class KSC results, and code would be worth a second look.\n\nRecommendation: desk reject now; authors should rebuild the comparison and report honestly.","headline":"EKGNet is an incremental dynamic-convolution extension of 3D-DenseNet whose headline superiority claim is contradicted by its own comparison tables; the paper needs a controlled re-evaluation, not referee time as-is.","tokens_in":11637,"tokens_out":3817,"would_cite":false,"duration_ms":33302,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EKGNet assembles a fresh 3D kernel per input, reaching 99.84% accuracy on Indian Pines.","keywords":["hyperspectral image classification","3D convolution","dynamic convolution","expert system","contextual mapping","DenseNet"],"falsifier":"Rerun SSRN, 3D-CNN, 3D-SE-DenseNet, SpectralFormer, LGCNet, and DGCNet exactly under EKGNet's protocol (6:1:3 or 5:1:4 train/validation/test split, 15×15 neighbouring patches, same epochs and optimizer) and compare OA on Indian Pines, Pavia University, and KSC; if any baseline matches or exceeds EKGNet's 99.84/99.98/99.95, the reported superiority is a protocol artifact rather than a dynamic-kernel effect.","tokens_in":10376,"feed_emoji":"🛰️","tokens_out":6768,"duration_ms":60546,"temperature":0.7,"pith_summary":"This paper claims that hyperspectral image classification can be pushed past 99.8% overall accuracy by replacing static 3D convolution kernels with per-input dynamic kernels. EKGNet couples a context-aware mapping network—which condenses each input patch into a global summary and turns it into K attention weights—with an \"expert\" bank of K base convolution kernels whose weighted sum forms the kernel actually applied. The dynamic kernel is embedded in a 3D-DenseNet whose growth rate doubles at each stage and whose layers are fully densely connected. The authors report overall accuracy of 99.84% on Indian Pines, 99.98% on Pavia University, and 99.95% on KSC, beating SSRN, 3D-CNN, 3D-SE-DenseNet, SpectralFormer, LGCNet, and DGCNet. A careful reader would care because dynamic convolution offers a parameter-effective way to adapt to sparse, spectrally redundant data without deepening or widening the network.","feed_headline":"Dynamic expert kernels lift hyperspectral accuracy to 99.84%","feed_subtitle":"Per-input context builds each 3D kernel, beating SSRN, 3D-CNN, LGCNet, and DGCNet on three benchmarks.","key_machinery":"The load-bearing object is the dynamic expert convolution system: a set of K learnable base 3D kernels treated as experts, aggregated by input-dependent attention weights into one kernel $W_{dyn}=\\sum_{k=1}^K \\alpha_k W_k$. The weights come from the context-aware mapping network, which pools the input to a global context vector and maps it through residual 3D-convolution blocks to K softmax scores with a learnable temperature. This mechanism does the work of adapting the receptive field and spectral emphasis per sample; the modified 3D-DenseNet—with exponentially increasing growth rate $k=2^{m-1}k_0$ and full connectivity between blocks via average-pooled skip connections—carries feature reuse and depth efficiency.","core_discovery":"The central claim is that a tightly coupled pair—a context-aware mapping network and an expert convolution system—yields per-sample 3D kernels that improve hyperspectral classification. The mapping network $f_{map}$ takes the global average pooling $g = \\mathrm{AvgPool3d}(X)$ of a hyperspectral patch and returns attention weights $\\alpha = \\mathrm{Softmax}(f_{map}(g)/\\tau)$, where $\\tau$ is a learnable temperature. The expert system holds K base kernels $\\{W_1,\\dots,W_K\\}$ that are combined as $W_{dyn} = \\sum_{k=1}^K \\alpha_k W_k$, so each input patch effectively gets its own convolutional kernel. Because each sample receives a customized kernel, the model can emphasize fine spectral details in sparse regions and structural patterns in complex spatial layouts, while the 3D-DenseNet backbone reuses features through dense connections and an exponentially increasing growth rate. The paper's stated conclusion is that this design achieves leading OA, AA, and Kappa on Indian Pines, Pavia University, and KSC under the chosen data splits and patch size, outperforming the compared mainstream classifiers.","pith_inferences":["Editor's inference: if the per-input kernel assembly is truly the active ingredient, the same mapping-plus-expert module should transfer to 2D CNN and Transformer hyperspectral classifiers, a test the paper does not run.","Editor's inference: the reported experiments use generous training fractions (60% on Indian Pines, 50% on Pavia University and KSC) and 15×15 overlapping patches, so the practical margin under small-sample regimes—say 10–20 labelled pixels per class—remains open; evaluating EKGNet under standard small-sample splits would clarify where the gain survives.","Editor's inference: the temperature $\\tau$ decays during training, but the paper does not ablate this schedule; holding $\\tau$ fixed or annealing it differently would isolate how much of the gain comes from sharpening attention versus the mapping network's feature extraction."],"forward_implications":["Because the kernel is assembled per sample, EKGNet can adapt its spatial-spectral emphasis to different ground-object distributions without adding network depth or width.","The exponential growth-rate schedule concentrates parameters in later dense blocks, improving computational efficiency while preserving accuracy.","Fully dense connectivity between the input and all subsequent layers across blocks increases feature reuse beyond the original DenseNet design.","On Indian Pines, Pavia University, and KSC, the reported overall accuracies (99.84%, 99.98%, 99.95%) exceed the compared baselines, including LGCNet and DGCNet.","Dynamic combination of base kernels can replace stacking many static kernels, reducing the parameter redundancy that commonly appears in 3D-CNN hyperspectral models."],"supporting_citations":[{"why":"Supplies the DenseNet architecture and the observation that deeper layers rely on high-level features, which motivates EKGNet's exponentially increasing growth rate and full connectivity.","marker":"[5]"},{"why":"Provides the 3D-DenseNet backbone that EKGNet modifies with dynamic kernels and revised growth/connectivity.","marker":"[22]"},{"why":"Defines 3D-SE-DenseNet, a dense 3D-CNN baseline against which EKGNet reports higher OA, AA, and Kappa.","marker":"[14]"},{"why":"Introduces DGCNet's dynamic grouped convolution, the closest dynamic-kernel baseline EKGNet extends and outperforms.","marker":"[7]"},{"why":"Introduces LGCNet's learnable grouped convolution, a comparison baseline in the Indian Pines and Pavia University experiments.","marker":"[11]"},{"why":"Defines SpectralFormer, the Transformer baseline that EKGNet compares against on Indian Pines and Pavia University.","marker":"[4]"},{"why":"Presents DFAN, a deep feature aggregation network cited as a widely used method whose dense connections motivate EKGNet's redundancy-aware design.","marker":"[23]"},{"why":"Presents MSDN, a multi-scale dense network in the same family that EKGNet's improved DenseNet builds on.","marker":"[21]"}],"fun_headline_variants":["Context mapping builds per-sample 3D kernels for sharper HSI results","Expert kernels generated on the fly beat static 3D-CNNs in HSI","Per-patch dynamic kernels from contextual mapping lift HSI accuracy","Adaptive expert convolution: new best on three hyperspectral benchmarks","Input-driven kernel combinations push hyperspectral accuracy higher"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That EKGNet beats the baselines depends on an unstated premise: those baselines were given the same large training sets (50–60% of labels) and the same 15×15 overlapping patches, but the paper does not document that protocol for them.","fun_headline_variants_meta":{"raw":{"variants":["Context mapping builds per-sample 3D kernels for sharper HSI results","Expert kernels generated on the fly beat static 3D-CNNs in HSI","Per-patch dynamic kernels from contextual mapping lift HSI accuracy","Adaptive expert convolution: new best on three hyperspectral benchmarks","Input-driven kernel combinations push hyperspectral accuracy higher"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1573,"prompt_tokens":1031,"completion_tokens":542,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":452}},"tokens_in":647,"tokens_out":542,"duration_ms":5871,"temperature":1.0,"reasoning_tokens":452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:15:46.389697+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun SSRN, 3D-CNN, 3D-SE-DenseNet, SpectralFormer, LGCNet, and DGCNet exactly under EKGNet's protocol (6:1:3 or 5:1:4 train/validation/test split, 15×15 neighbouring patches, same epochs and optimizer) and compare OA on Indian Pines, Pavia University, and KSC; if any baseline matches or exceeds EKGNet's 99.84/99.98/99.95, the reported superiority is a protocol artifact rather than a dynamic-kernel effect.","supporting_citations":[{"cited_title":"Spatial-spectral hyperspectral classification based on learnable 3d group convolution","cited_arxiv_id":null,"evidence_quote":"Introduces LGCNet's learnable grouped convolution, a comparison baseline in the Indian Pines and Pavia University experiments."}],"review_version":1}