{"id":"a6d7271a-0991-4b7b-ace0-53053e113bc7","arxiv_id":"2411.18115","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A spatial-spectral transformer combined with active transfer learning and uncertainty-driven sample selection is reported to improve cross-domain hyperspectral image classification with fewer labels.","lead":"This paper combines a spatial-spectral transformer with active learning and transfer learning to classify hyperspectral images using very few labeled samples. It reports accuracy gains over existing methods on six benchmark datasets, but leaves key algorithmic details unspecified and uses a questionable cross-dataset protocol.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-dataset fine-tuning in Section VI is ill-defined: the target datasets have different label sets and the paper never describes how the classification head is replaced or how classes are mapped.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: cross-dataset transfer assumes label compatibility that the paper never establishes. My stress-test confirms this is the single most load-bearing issue because it directly invalidates the 'cross-hyperspectral' claim in the title and abstract. The results in Tables VIII-X are presented as evidence that fine-tuning with 10% target samples yields high accuracy across datasets, but the paper provides no mechanism for transferring the classifier when class sets differ. This is not a matter of external consensus or minor experimental detail; it is an internal inconsistency: the training objective in Eq. (19) cannot be evaluated for target labels without either a new output layer or a semantic class mapping. The dynamic-freezing underspecification is also real but secondary, since it affects reproducibility rather than the logical well-posedness of the transfer experiment. Because the reader already rejected the paper on this basis at moderate confidence, and my analysis strengthens that conclusion, I recommend no change to the verdict. The concrete check I propose—inspecting the released code or attempting to derive the loss—would settle whether the concern lands; if the code contains an undocumented head-resize and class mapping, the paper would need only a revision, but as written the claim is unsupported. I agree with the reader's assessment, so no verdict adjustment is needed.","tokens_in":19382,"tokens_out":5197,"duration_ms":49051,"concrete_test":"Inspect the public repository (github.com/mahmad000/ATL-SST) or contact the authors for the cross-dataset fine-tuning code. Specifically, check whether the model's classification head is reinitialized or resized to the target dataset's class count before fine-tuning, and whether a class-index mapping between source and target is defined. If no such head replacement or mapping exists, the experiments in Tables VIII-X cannot be executed as described; if it does exist, verify that the reported numbers are reproducible. A minimal analytical check is to derive the loss (Eq. 19) for a source-trained head evaluated on target labels: with mismatched output dimensions, the categorical cross-entropy is mathematically undefined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central cross-domain claim rests on Tables VIII-X, where an SST trained on one dataset (SA: 16 classes, PU: 9, UH: 15, HC: 16, HH: 22, QUH: ~24) is said to be 'fine-tuned using 10% samples' from another dataset. The paper never states how the output layer is resized, reinitialized, or how source/target class indices correspond. Equation (19) defines the loss over C classes; if C is the source class count, target labels for a different number of classes are undefined, and if C is the target count, the pretrained source head is incompatible. No class-mapping table, label-alignment strategy, or head-replacement description appears anywhere in Sections III or VI. Without one of these, the reported OA/AA/kappa values in Tables VIII-X cannot be produced by the described architecture, making the paper's headline 'cross-dataset' generalization result unreproducible and internally inconsistent. This is more fundamental than the underspecified dynamic-freezing rule (Sec. III-C), which only affects implementation details, whereas the label mismatch affects the logical validity of the transfer experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SST-ATL, an active transfer learning framework for cross-dataset hyperspectral image classification. The method combines a spatial-spectral transformer backbone with an uncertainty-diversity active learning query strategy, dynamic layer freezing based on MMD, and an uncertainty-calibrated attention mechanism. Experiments are reported on six HSI datasets, including same-dataset active learning curves, cross-dataset fine-tuning tables, comparisons with CNN/Transformer/state-space baselines, and ablations. The manuscript claims consistent gains of +2.5% to +3.8% OA over baselines under constrained annotation budgets.","tokens_in":19653,"tokens_out":5128,"duration_ms":51204,"significance":"If the method is fully specified and reproducible, the result would be practically valuable: label-efficient cross-domain HSI classification is an important problem, and the paper evaluates on multiple standard benchmarks with a public code link. The ablation study and query-strategy comparison are useful steps toward understanding the contribution. However, as written, several load-bearing components are not specified precisely enough to reimplement, and the cross-dataset protocol that supports the main claim is not described at the architectural level. The central empirical claim is therefore not verifiable from the manuscript alone.","major_comments":[{"comment":"The cross-dataset fine-tuning protocol is underspecified with respect to label sets. The source and target datasets have different numbers of classes (SA 16, PU 9, UH 15, HC 16, HH 22, QUH 24), yet the manuscript never states how the output layer Ofinal is replaced or re-initialized, how source and target class indices are aligned, or which C is used in the cross-entropy loss after transfer. Without this information, the reported OA/AA/kappa values in Tables VIII-X cannot be produced by the described architecture, and the paper's headline cross-dataset generalization claim is not supported.","section":"Section VI, Tables VIII-X, Eq. (19)"},{"comment":"Dynamic layer freezing is described only at a high level: 'layers with low MMD variance across domains are frozen.' No MMD threshold, layer-selection criterion, update schedule, or algorithm is given. Since the ablation in Table XIV includes a 'w/o Dynamic Freezing' variant, the implementation must exist, but as written this central component cannot be reimplemented or independently checked.","section":"Section III-C, Table XIV"},{"comment":"The calibrated attention update Acal = Softmax(QK^T/sqrt(d_k)) * (1 + lambda * U) is not well defined. U is introduced in Eq. (13) as the negative maximum predictive probability for a sample, not as a per-token vector, and the paper does not state how U is broadcast against the attention matrix or whether the result is renormalized. Multiplying a row-stochastic attention matrix by a position-dependent factor generally breaks the row-stochastic property, so the mechanism as written is underspecified.","section":"Section III-A, Eq. (9)"},{"comment":"The active-query definition is internally inconsistent. Eq. (14) defines the chosen set as the argmax of Diversity(S) over all subsets of size query size, while Eq. (17) selects the top query-size entries of an argsort over per-pixel diversity; these are different selection rules unless additional structure is assumed. In addition, d_{jk} in Eq. (15) is undefined, Eq. (16) duplicates Eq. (15) with new notation, and the query percentage of 0.02% in Section IV is not reconciled with the AL rounds of 250-1000 labeled samples reported in Table XV.","section":"Section III-B and Section IV"}],"minor_comments":[{"comment":"The backbone is called 'Spectral-Spectral Transformer' in Section III-A but 'Spatial-Spectral Transformer' in the abstract and elsewhere; please make the terminology consistent.","section":"Title and Section III-A"},{"comment":"The header 'Tr Va Te' in Tables XI-XIII is not defined anywhere, and the per-class sample counts do not explain how the comparison methods were trained (e.g., random sampling versus active learning). Several rows report 100% OA, which makes the comparison hard to interpret without information about test-set sizes and variance.","section":"Tables XI-XIII"},{"comment":"The reported FLOPs and parameter counts vary slightly across datasets despite the same backbone architecture, and no unit or calculation formula is provided; please clarify why these values differ.","section":"Section V, Tables II-VII"},{"comment":"U(x_i) = -max p(y|x_i) is a negative confidence score, not an entropy; the text calls it 'entropy-based uncertainty,' which is inaccurate.","section":"Eq. (13)"},{"comment":"The names 'SST-ATL' and 'ATL-SST' are used inconsistently, and Table XI contains the typo 'WaveMamab' for WaveMamba; please unify the nomenclature.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The availability of the source code is a positive signal, but I was not able to verify that the code corresponds to the printed equations. The cross-dataset protocol and dynamic-freezing rule are so underspecified that the paper currently functions more as a system description than a reproducible method. If the authors can supply an explicit fine-tuning/head-replacement protocol, the MMD threshold and layer-selection rule, and a corrected formulation of Eq. (9) and the query selection equations, a revised version could be evaluated on its merits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has two real strengths: the integration is new, and the empirical sweep is broad. The SST backbone with dynamic layer freezing, self-calibrated attention, and a hybrid uncertainty-diversity query is not in the prior literature, including your own earlier 3D-CNN ATL work. The related-work table is useful, and the ablations are honest: each component gives a measurable bump, and the query-strategy comparison shows the hybrid rule is genuinely better than random or single-criterion sampling. That is the kind of systematic evidence a system paper should have.\n\nThe soft spot is the cross-dataset part, and it is not minor. Tables VIII–X are the paper's headline, but they are built on a protocol the paper never defines. Salinas has 16 classes, Pavia University 9, Houston 15, and the others 16, 22, and roughly 24. The paper says the model is fine-tuned with 10% samples from the target dataset, but never states how the output layer is resized, how class indices correspond, or even whether a shared class subset was used. The loss in Eq. 19 sums over C classes; if C is the source count, the target labels are undefined, and if it is the target count, the pretrained head is incompatible. This is not a missing hyperparameter, it is a missing experiment. Without that description, the OA/AA/kappa values in Tables VIII–X cannot be traced to the architecture.\n\nThe other two novelties are underspecified in a similar way, though they hurt less. The dynamic freezing rule never gives the MMD threshold or the layer-freezing algorithm. The self-calibrated attention in Eq. 9 depends on an uncertainty term U that is never defined for tokens, and the non-normalized scaling of softmax outputs is not justified. These are implementation details that might be fixable, but they currently block reproduction.\n\nThere are also smaller issues worth flagging. The abstract claims +2.5% to +3.8% OA over baselines, but the same-dataset comparisons in Tables XI–XIII show gains of roughly +1.3% to +2.1% against SST and HybViT, and the paper never compares against your own 3D-CNN ATL method, which is the natural baseline for this claim. No error bars or statistical tests appear anywhere in the tables.\n\nWho is this for? Someone working on label-efficient HSI classification might read it for the related work and the ablation structure. But as a contribution, the central cross-domain result is unverified and, as written, not reproducible. I would send it out for peer review because the gap is fillable with careful protocol description and re-evaluated experiments, and the framework is interesting enough to justify the referee time. I would not cite it in its current form.","headline":"The cross-dataset claim is the paper's core selling point, but the fine-tuning protocol never explains how models trained on different class sets are adapted, making the main result unreproducible as written.","tokens_in":20155,"tokens_out":2705,"would_cite":false,"duration_ms":26006,"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":"A spatial-spectral transformer combined with active transfer learning claims gains of +2.5% to +3.8% overall accuracy over CNN, transformer, and state-space baselines on six hyperspectral benchmarks.","keywords":["hyperspectral image classification","active learning","active transfer learning","spatial-spectral transformer","cross-domain adaptation","label efficiency","dynamic layer freezing","uncertainty-diversity sampling"],"falsifier":"Run the released code with the reported 1%/49%/50% split and the reported hyperparameters, but replace the dynamic-freezing rule with a fixed rule (freeze all layers except the last) and reinitialize the classifier head for each target dataset's class set; if the +2.5% to +3.8% OA margin over baselines does not survive or the cross-dataset kappa drops below the reported range, then the unstated freezing and label-mapping details, not the framework as described, are carrying the result.","tokens_in":19216,"feed_emoji":"🛰️","tokens_out":9833,"duration_ms":80652,"temperature":0.7,"pith_summary":"This paper is trying to establish that a single framework can make hyperspectral image classification work in the label-scarce, cross-scene setting where most deployments actually happen. The authors propose SST-ATL, which wraps a spatial-spectral transformer in an active transfer-learning loop: a hybrid query criterion selects pixels that are both uncertain and spectrally diverse for annotation, a dynamic freezing rule adapts only the layers most affected by domain shift during fine-tuning, and self-calibrated attention reweights spectral-spatial tokens by uncertainty. They report that the method outperforms CNN-, transformer-, and state-space-model baselines by +2.5% to +3.8% overall accuracy on six benchmarks, and that fine-tuning a model trained on one dataset with only 10% of another dataset's samples keeps kappa at or above roughly 92%. If these numbers hold, the framework would make hyperspectral classification substantially cheaper in practice, since labeling is the dominant cost and scenes differ across sensors, dates, and regions.","feed_headline":"Up to 3.8% accuracy gain from active transfer learning","feed_subtitle":"A spatial-spectral transformer with uncertainty-diversity querying adapts to new scenes using just 10% of labeled samples.","key_machinery":"The load-bearing object is the SST-ATL pipeline: a spatial-spectral transformer encoder (3D patch embedding, sinusoidal position coding, multi-head self-attention, MLP classification head) surrounded by an active transfer loop. The hybrid query function scores each pool sample by uncertainty, $U(x_i)=-\\max p(y|x_i)$, and by diversity, the mean pairwise Euclidean distance between neighboring spectral vectors, then queries the highest-scoring set. Dynamic layer freezing estimates domain shift with maximum mean discrepancy of intermediate features and freezes low-variance layers during fine-tuning, reducing compute while keeping transferable representations. Self-calibrated attention rescales attention weights by token entropy, $\\mathrm{Softmax}(QK^T/\\sqrt{d_k})\\cdot(1+\\lambda U)$, biasing the encoder toward ambiguous spatial-spectral tokens. These three mechanisms are what the ablations remove one at a time to show that each contributes to the final accuracy.","core_discovery":"The central claim is that active transfer learning around an SST backbone solves both halves of the cross-domain problem at once: active learning supplies the most informative target-domain labels, and transfer learning with selective adaptation supplies the source-domain prior. On the UH, PU, and SA benchmarks the paper reports ATL-SST reaching OA of 97.57%, 98.87%, and 99.48%, respectively, ahead of every compared CNN, transformer, and state-space baseline; the same-dataset learning curves show smooth gains as queried samples grow, and the ablation study attributes the largest single contribution to the hybrid uncertainty-diversity query, with secondary contributions from diversity sampling, dynamic freezing, and self-calibrated attention. In the cross-dataset experiments, an SST actively trained on one dataset and fine-tuned with 10% of another's samples keeps kappa at or above about 92% across the tested source-target pairs.","pith_inferences":["Because the datasets carry different class label sets (for example Salinas has 16 classes while Pavia University has 9), the cross-dataset tables must involve an unstated output-layer replacement or label-semantics mapping; that choice could plausibly account for several OA points and should be reported before the transfer gains are taken at face value.","The dynamic-freezing rule is specified only as freezing layers whose MMD variance is low, with no threshold or schedule; until that is pinned down, the ablation 'w/o dynamic freezing' cannot be cleanly separated from ordinary fine-tuning.","The query strategy's early-round advantage suggests the same uncertainty-diversity criterion could transfer to few-shot or self-supervised HSI pipelines, an extension the authors mention only as possible future work.","If MMD-based freezing really identifies domain-invariant layers, then computing MMD on unlabeled target features could serve as a cheap domain-shift probe before any labels are queried; the paper reports no such diagnostic use."],"forward_implications":["A user can start from one actively trained encoder and adapt it to a new hyperspectral scene with only 10% of the scene's labels, keeping $\\kappa$ at or above roughly 92% on the tested pairs.","At the smallest label budgets tested (250 samples), the hybrid query beats random sampling by more than 3% OA on PU, UH, and SA, so the gain is largest where labeling is most expensive.","Ablations attribute the largest accuracy loss to removing active learning (about 2.7 OA points), with smaller but consistent drops from removing diversity sampling, dynamic freezing, or self-calibrated attention.","The model stays near 836K parameters and about 110K FLOPs with roughly stable training times as the labeled set grows, so repeated active-learning rounds do not require proportionally more compute."],"supporting_citations":[{"why":"Attention graph convolutional network used as one of the CNN baselines in the state-of-the-art comparison tables.","marker":"[51]"},{"why":"Hybrid spatial-spectral transformer used as a transformer baseline in Tables XI-XIII.","marker":"[52]"},{"why":"Spatial-spectral transformer baseline whose architecture the proposed SST backbone extends and outperforms.","marker":"[53]"},{"why":"Wavelet-based spatial-spectral transformer used as a baseline for the spectral-redundancy handling comparison.","marker":"[43]"},{"why":"Pyramid-based spatial-spectral transformer baseline in the comparison tables.","marker":"[12]"},{"why":"Spatial-spectral Mamba baseline representing state-space models in the comparison.","marker":"[54]"},{"why":"Wavelet-based Mamba baseline used as the state-space competitor in the comparison tables.","marker":"[55]"},{"why":"First active transfer learning network for HSI; defines the setting SST-ATL is built on.","marker":"[45]"},{"why":"Earlier active-learning-incorporated deep transfer learning pipeline for HSI that motivates the ATL design.","marker":"[46]"},{"why":"Authors' prior 3D-CNN active transfer learning work, the approach SST-ATL extends with a transformer backbone.","marker":"[47]"}],"fun_headline_variants":["Active transfer learning with transformer improves cross-scene HSI accuracy","Uncertainty-diversity querying reduces labels for cross-domain HSI transfer","Dynamic layer freezing and self-calibrated attention sharpen HSI transfer","Only 10% labeled samples needed for active transfer learning in HSI","Transformer active transfer attains 97.57% OA on UH hyperspectral scene"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cross-dataset claim rests on an unspecified transfer protocol: the six datasets have different class label sets, and the paper never states how the output layer is replaced or how the dynamic-freezing rule decides which layers to freeze, so the reported margins could depend on unstated implementation choices.","fun_headline_variants_meta":{"raw":{"variants":["Active transfer learning with transformer improves cross-scene HSI accuracy","Uncertainty-diversity querying reduces labels for cross-domain HSI transfer","Dynamic layer freezing and self-calibrated attention sharpen HSI transfer","Only 10% labeled samples needed for active transfer learning in HSI","Transformer active transfer attains 97.57% OA on UH hyperspectral scene"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001138,"raw_usage":{"total_tokens":4714,"prompt_tokens":924,"completion_tokens":3790,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":3693}},"tokens_in":540,"tokens_out":3790,"duration_ms":25187,"temperature":1.0,"reasoning_tokens":3693,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:29:52.444743+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code with the reported 1%/49%/50% split and the reported hyperparameters, but replace the dynamic-freezing rule with a fixed rule (freeze all layers except the last) and reinitialize the classifier head for each target dataset's class set; if the +2.5% to +3.8% OA margin over baselines does not survive or the cross-dataset kappa drops below the reported range, then the unstated freezing and label-mapping details, not the framework as described, are carrying the result.","supporting_citations":[{"cited_title":"Attention graph convolutional network for disjoint hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Attention graph convolutional network used as one of the CNN baselines in the state-of-the-art comparison tables."},{"cited_title":"Hybrid multi- scale spatial–spectral transformer for hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Hybrid spatial-spectral transformer used as a transformer baseline in Tables XI-XIII."},{"cited_title":"Centerformer: A center spatial–spectral attention transformer network for hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Spatial-spectral transformer baseline whose architecture the proposed SST backbone extends and outperforms."},{"cited_title":"S2mamba: A spa- tial–spectral state space model for hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Spatial-spectral Mamba baseline representing state-space models in the comparison."},{"cited_title":"Wavemamba: Spatial-spectral wavelet mamba for hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Wavelet-based Mamba baseline used as the state-space competitor in the comparison tables."},{"cited_title":"Active transfer learning network: A unified deep joint spectral–spatial feature learning model for hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"First active transfer learning network for HSI; defines the setting SST-ATL is built on."},{"cited_title":"Active-learning- incorporated deep transfer learning for hyperspectral image classifica- tion,","cited_arxiv_id":null,"evidence_quote":"Earlier active-learning-incorporated deep transfer learning pipeline for HSI that motivates the ATL design."}],"review_version":1}