{"id":"7bea258c-7651-46a1-9219-84558a8f534e","arxiv_id":"2508.18075","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A prototype-learning pipeline with an extra 'unknown' anchor discovers and clusters novel hyperspectral classes under few-shot supervision.","lead":"This paper presents a method for hyperspectral image classification that can both reject pixels from unseen land-cover classes and, after rejection, automatically group those pixels into distinct new classes using only a few labeled examples per known class. It combines an extra 'unknown' class anchor in the classifier with a prototype clustering module, and reports competitive results on four benchmark hyperspectral datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training protocol appears to train on test unknown classes; SOTA claim unsubstantiated without a disjoint split.","rationale":"The most load-bearing assumption is not the single-anchor sufficiency but the training/evaluation protocol. The reader's weakest assumption mentions representativeness of query unknown samples, but the text suggests a stronger issue: the unknown classes themselves are part of training. This is supported by two direct quotes (Section III-B and Section II-D). If the protocol uses the same unknown classes at train and test, the central claim is invalid because the task reduces to supervised classification with a collapsed pseudo-label. The concrete test would settle this by checking the code or running a disjoint-split experiment. Even if the split is disjoint, the paper must disclose it; as written, it is misleading. I therefore recommend the verdict move from CONDITIONAL to UNVERDICTED, since the current evidence cannot verify the claim.","tokens_in":15614,"tokens_out":8038,"duration_ms":92904,"concrete_test":"Use the released code (github.com/KOBEN-ff/OpenFUCD-main) to inspect the data-loading routine: for each dataset, are the class IDs in the training-query unknown set the same as the class IDs marked unknown in the test evaluation? If yes, retrain under a strict split: hold out the designated unknown classes from all training tasks and use an auxiliary HSI dataset (e.g., train on IP as source, test on PU unknown classes) or a disjoint subset as the pseudo-unknown provider. Recompute Table II's Unknown ACC / ALL ACC. If the 5-shot WHU-Hi-HanChuan ALL ACC drops substantially below the 19-point gap over EVML, the reported result depended on leakage.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim is the SOTA result in Table II (e.g., 5-shot WHU-Hi-HanChuan ALL ACC 89.41 vs EVML 70.62). For that claim to be valid, the 'unknown' classes at test time must be genuinely unseen during training. The manuscript undermines this: Section III-B says 'd samples per class for unknown classes are randomly selected to form a task in each episode,' and Section II-D says 'a pseudo-label is used for the samples from all the unknown classes during training.' No disjoint train/test split of the unknown classes is stated anywhere; the dataset descriptions (Section III-A) fix the known/unknown partition once and use those same classes in training and evaluation. If the training and test unknown classes are identical, the model is trained with pseudo-labeled examples of the classes it is later asked to 'discover'. The reported Unknown ACC then reflects recognition of previously seen classes, and the comparison against baselines that were not trained on those examples is unfair. This makes the headline superiority unsubstantiated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a few-shot method for hyperspectral image classification in which the query set may contain pixels from classes not represented in the labeled support set. The method has two main components: (i) a class-anchor classifier with an extra anchor for all unknown classes, trained with pseudo-labeled unknown samples, and (ii) a set of trainable prototypes with contrastive losses and Jaccard/Louvain clustering, used to split rejected unknown samples into discovered classes. Experiments are reported on Indian Pines, Pavia University, Salinas, and WHU-Hi-HanChuan under 1-shot and 5-shot settings, with Known ACC, Unknown ACC, ALL ACC, ablations, and hyperparameter analyses. The code is publicly available.","tokens_in":15843,"tokens_out":9006,"duration_ms":119157,"significance":"Extending open-set hyperspectral classification from simple unknown rejection to unknown-class discovery is a timely and useful goal, and the proposed combination of an open-set logit anchor with prototype-level clustering is reasonable. The release of code is a positive aspect. If the reported results survive a properly specified and fair evaluation, the method would be a useful baseline for few-shot HSI unknown-class discovery. However, the current evidence is not sufficient: the training protocol appears to involve the same unknown classes used in the final evaluation, the comparison with baselines has at least one explicit asymmetry, and no uncertainty quantification is provided. These issues affect the central claim of state-of-the-art performance.","major_comments":[{"comment":"Training/evaluation protocol is under-specified and appears to leak the unknown classes into training. Section III-B states that 'k+d samples per class for known classes and d samples per class for unknown classes are randomly selected to form a task in each episode,' and Section II-D states that 'a pseudo-label is used for the samples from all the unknown classes during training.' Section III-A fixes a single known/unknown class partition and uses the same unknown classes in training and testing; no class-disjoint or pixel-disjoint split is described. Thus the model is trained with coarse labels on the very unknown classes it is later reported to 'discover.' Unknown ACC then measures recognition of previously trained classes, not discovery of novel classes. If the intended setting is novel category discovery, the protocol should be stated explicitly and baselines should be trained under","section":"III-B, II-D, III-A"},{"comment":"The text claims a gain of '11% in terms of known ACC on WHU-Hi-HanChuan' under the 5-shot setup. Table II lists the proposed method's Known ACC as 88.98; the best baseline is SSMLP-RPL at 85.46, a difference of 3.52 percentage points, while EVML is 71.07. No baseline value in the table yields an 11-point gap. This specific claim is unsupported by the reported numbers. The other percentage gains in the same paragraph should be recomputed consistently with Table II.","section":"III-C, Table II"},{"comment":"No uncertainty quantification is provided. The paper reports a single accuracy value per setting. Episodic task sampling, random augmentations, prototype initialization/clustering, and Louvain clustering are stochastic, so small differences (e.g., 1.9% Known ACC on SA, 3% ALL ACC on PU) may be within natural variability. The authors should report mean and standard deviation over multiple random seeds or task sets, and ideally a significance test, before claiming improvements over the best baselines.","section":"III-B, Table II"},{"comment":"The comparison with baselines is asymmetric. For the baselines, k-means is applied with the true number of unknown classes: the paper states that 'we do the comparison under the assumption that the aforementioned methods are aware of the number of unknown classes.' For the proposed method, the number of classes is estimated via prototype clustering (Table IV). This gives the baselines an oracle advantage in the clustering stage. A fair comparison should either provide the true unknown-class count to the proposed method as well, or evaluate all methods under an estimation setting. In addition, the manuscript does not specify whether the baselines were retrained under the same episodic unknown-in-query protocol; if not, the comparison conflates method quality with protocol differences.","section":"III-C"},{"comment":"The metric computation with mismatched class counts is not explained. Table IV reports predicted numbers of classes that often differ from the true numbers (e.g., PU 1-shot: 15 predicted vs 9 true; IP: 17 vs 16). Yet Table II reports Unknown ACC and ALL ACC computed with Hungarian matching. Hungarian assignment requires an explicit rule when the numbers of true and predicted classes differ; the paper does not state how over-segmented or under-segmented clusters are matched and scored. Without this, the reported accuracies are not reproducible.","section":"Table IV, Section III-B"}],"minor_comments":[{"comment":"The summation index in the class-anchor loss is ambiguous: the sum is written over 'j != i' with i a sample index, while d_j appear to be distances to class anchors. The sum should be indexed over anchors, not samples.","section":"Eq. (6)"},{"comment":"The Louvain clustering algorithm is cited as [36], but reference [36] is the MAML paper (Finn et al.). A proper reference for Louvain clustering (e.g., Blondel et al., 2008) is missing.","section":"Ref. [36]"},{"comment":"There are numerous typos and grammatical errors: 'discoverying,' 'unknow,' 'traing,' 'evluation,' 'ground-turth,' 'ground reality map,' and 'MORAN' for 'MORGAN.' The manuscript needs careful proofreading.","section":"Throughout"},{"comment":"The abstract says four benchmark HSI datasets, while Section IV says 'Based on three HSI datasets.' Table IV also says 'across three datasets' but lists four datasets. Please correct the inconsistency.","section":"Abstract / Section IV"},{"comment":"The pre-training set Dpre used in Algorithm 1 is not defined in the experimental setup. Section III-B mentions data augmentation for the pre-training set but does not specify its composition, size, or split, which affects reproducibility.","section":"Algorithm 1 / III-B"},{"comment":"The t-SNE figure caption is confusing: the labels '(a)-right' and '(a)-left' are not clearly matched to the panels, and the text states which classes are unknown without mapping them to the colors in the figure. Please clarify.","section":"Section III-D / Fig. 10"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the protocol ambiguity: the paper oscillates between open-set recognition (truly unseen unknown classes) and novel category discovery (unlabeled classes observed during training). The authors should be required to state the intended setting explicitly, define the train/test split, and rerun the experiments in a class-disjoint fashion if the open-set interpretation is intended. The erroneous 11% claim and the incorrect Louvain reference suggest that the experimental section needs careful verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is reasonable: add an unknown-class anchor to the logit space, then use prototype contrastive learning plus Louvain clustering to split rejected unknown samples into distinct classes. The individual pieces are all borrowed from prior work, but putting them together for few-shot HSI open-set discovery is a legitimate integration, and the paper lays out the losses clearly enough that a reader could reimplement the approach. That is the honest novelty, and it is not nothing.\n\nThe problem is the evaluation. Section III-B says that during training, d samples per class are randomly selected from the unknown classes to form each episode, and Section II-D says a pseudo-label is used for all unknown-class samples during training. The known/unknown partition is fixed once and used for both training and testing, and no disjoint split is stated. So the model is trained on the very classes it is later asked to 'discover.' Under that protocol, the unknown ACC numbers in Table II reflect recognition of previously seen classes, not discovery of unseen ones. The comparison is also asymmetric: baselines are given the true number of unknown classes plus k-means, while the proposed method estimates the number itself, and the baselines were not trained on unlabeled unknown-class samples. That makes the SOTA claim in the abstract and Section III-C unsubstantiated.\n\nThe text also contains numeric claims that do not match the table. It says the method 'exceeds by 11%' on WHU-Hi-HanChuan known ACC under 5-shot, but the largest gap in Table II is roughly 17.9 points versus EVML and about 3.5 points versus SSMLP-RPL. No value there is 11. Similarly, the '16%' improvement on IP unknown ACC does not line up with any adjacent comparison. These are small but symptomatic of sloppy reporting.\n\nOther soft spots: no error bars or multiple seeds; the reference list is scrambled ([36] is cited for Louvain but points to MAML, [37] is used both for softmax non-injectivity and for DFSL); and the conclusion says 'three HSI datasets' when four were used. None of these are fatal to the method, but they add noise.\n\nBottom line: the method deserves a serious referee, but only if the authors fix the evaluation. They need either a genuinely disjoint unknown-class split at test time, or a clear repositioning as a generalized-category-discovery-style setting, with baselines given the same unlabeled data. I would not cite this in its current form, but I would read a revised version.","headline":"The method is a sensible integration of existing building blocks, but the headline SOTA claim is undercut by a training protocol that lets the model see the same unknown classes it is later asked to 'discover.'","tokens_in":16316,"tokens_out":2777,"would_cite":false,"duration_ms":36060,"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":"This paper claims that hyperspectral pixels from classes never seen in training can be not only rejected but sorted into distinct new classes using a class-anchor classifier plus prototype clustering, with only 1 or 5 labeled pixels per kno","keywords":["hyperspectral image classification","few-shot learning","open-set recognition","unknown class discovery","prototype learning","deep clustering","class anchor","remote sensing"],"falsifier":"Train the model with one set of unknown classes in the episodic query sets, then test it with a disjoint set of unknown classes from the same or another HSI scene. If unknown ACC and the discovered cluster count collapse to chance-level or become unstable, the rejection/discovery depends on having seen the test unknowns in training rather than on a generic unknown representation.","tokens_in":15488,"feed_emoji":"🛰️","tokens_out":7290,"duration_ms":81710,"temperature":0.7,"pith_summary":"This paper is trying to establish that open-set few-shot hyperspectral image classification can go one step further than rejecting unknown pixels: the model can also discover what those unknown classes are. The proposed pipeline has two stages. First, a class-anchor classifier expands the logit space with one extra dimension for an 'unknown class' anchor, separating known from unknown pixels. Second, a prototype-learning module initializes many trainable prototypes, clusters them into groups via similarity and Louvain clustering, and assigns rejected unknown pixels to the resulting groups. On four benchmark HSI datasets under 1-shot and 5-shot settings, the paper reports that this two-stage design beats prior open-set HSI methods on most metrics, with the largest reported gain on WHU-Hi-HanChuan at 5-shot (89.41% all-class accuracy). The broader point: with only a few labeled examples per known land-cover class, a model can both classify known classes and turn unlabeled unknown pixels into newly named classes, which matters for mapping new land cover over time.","feed_headline":"New classes in hyperspectral images get discovered, not just rejected","feed_subtitle":"With one to five labeled pixels per known class, prototype clustering sorts rejected pixels into distinct new categories.","key_machinery":"Class anchors and trainable prototype groups. A=(phi*e1, ..., phi*eN) defines N one-hot anchors; sample embeddings are compared to each anchor by Euclidean distance, and the N-th anchor is the rejection region for unknown classes. Discovery uses w trainable prototypes (w greater than true class count), with sample-to-prototype probabilities computed by a softmax over dot products, Jaccard similarity between prototype sample sets, Louvain clustering to form prototype groups, and Hungarian matching to map known prototype groups to real labels. The combination carries the argument: anchors provide the open-set separation, prototype groups provide the unknown-class taxonomy.","core_discovery":"The paper's central claim is that unknown-class discovery in hyperspectral images is achievable in the few-shot regime by combining a class-anchor-based open-set classifier with a dual-level prototype contrastive clustering module. The class anchors are one-hot vectors scaled by a parameter phi; the first N-1 anchors represent known classes and the N-th anchor represents all unknown classes. Training pushes embedded pixels toward their true anchor and away from others via losses Losc and Lca. The discovery stage then uses a large number of trainable prototypes, positive pairs built from weak and strong augmentations, prototype-similarity and prototype-group-similarity losses, a regularizatio","pith_inferences":["The unknown anchor is a single vector; the paper does not analyze when one vector can separate heterogeneous unseen classes. A natural extension is to learn multiple unknown anchors or add per-class unknown uncertainty.","In the experiments, unknown query samples used in training are drawn from the same class pool as test unknowns. Whether the discovery generalizes to truly novel classes that never appear in episodic training remains an open test.","The prototype-count sensitivity suggests a practical calibration rule could be derived: set the number of prototypes relative to expected diversity rather than fixing it per dataset.","Because the method labels unknown classes only relative to prototype groups, a downstream human-interpretable naming step (e.g., spectral signatures per group) would make the discovered classes actionable."],"forward_implications":["Unknown classes can be recovered as distinct named clusters without any labels for them, given only 1 or 5 labeled pixels per known class.","The number of unknown classes is estimated from the prototype-group structure rather than assumed in advance (e.g., 5-shot estimates on IP, PU, SA, and WHU-Hi-HanChuan are close to true class counts).","The ablation study implies each loss term is load-bearing; removing the open-set classification loss or class anchor loss hurts closed/open accuracy and AUROC the most.","Excess prototypes hurt: performance degrades when the number of trainable prototypes is far above the true class count, so the method has a usable operating range.","In applications, the output is a land-cover map where rejected pixels are split into meaningful new classes, not just colored as 'unknown'."],"supporting_citations":[{"why":"Defines the prior multitask-deep-learning approach to few-shot HSI classification with unknown classes; used as a baseline and problem reference.","marker":"[24]"},{"why":"Supplies the outlier calibration network baseline that rejects unknown pixels via calibrated outlier scores.","marker":"[25]"},{"why":"Supplies the reciprocal-points (RPL) baseline with learnable dynamic thresholds for open-set HSI classification.","marker":"[27]"},{"why":"Source of the evaluation protocol and metrics (Known ACC, Unknown ACC, ALL ACC) that the paper adopts.","marker":"[31]"},{"why":"Cited by the paper as the Louvain clustering algorithm used to merge prototypes into groups; the reference-list entry for this marker is MAML, so the bibliographic mapping is inconsistent.","marker":"[36]"},{"why":"Provides the DFSL deep few-shot learning baseline, which the paper adapts with a softmax threshold for unknown rejection.","marker":"[37]"},{"why":"Supplies the Hungarian algorithm used to match prototype groups to known labels and to compute clustering accuracy.","marker":"[38]"},{"why":"Provides the WHU-Hi-HanChuan dataset used in the main evaluation.","marker":"[39]"},{"why":"Provides the MORGAN baseline, a GAN-based meta-learning method for few-shot open-set recognition.","marker":"[43]"},{"why":"Provides the EVML baseline, an extreme-value-theory meta-learning method that rejects unknown samples.","marker":"[44]"}],"fun_headline_variants":["Hyperspectral few-shot: unknown classes get clustered, not just rejected","Discover unknown classes in hyperspectral images with few labels","Prototype learning discovers new classes in few-shot hyperspectral images","Few-shot HSI: reject then cluster unknowns into new classes"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a single unknown-class anchor in logit space can cleanly separate every unseen hyperspectral class from the known classes; if an unseen class is not captured by that one vector, rejected pixels will be contaminated and the discovered clusters will merge or split the true classes.","fun_headline_variants_meta":{"raw":{"variants":["Hyperspectral few-shot: unknown classes get clustered, not just rejected","Discover unknown classes in hyperspectral images with few labels","Prototype learning discovers new classes in few-shot hyperspectral images","Few-shot HSI: reject then cluster unknowns into new classes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1280,"prompt_tokens":750,"completion_tokens":530,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":494,"tokens_out":530,"duration_ms":6455,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:35:33.460879+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model with one set of unknown classes in the episodic query sets, then test it with a disjoint set of unknown classes from the same or another HSI scene. If unknown ACC and the discovered cluster count collapse to chance-level or become unstable, the rejection/discovery depends on having seen the test unknowns in training rather than on a generic unknown representation.","supporting_citations":[{"cited_title":"Few-shot hyperspectral image classi- fication with unknown classes using multitask deep learning,","cited_arxiv_id":null,"evidence_quote":"Defines the prior multitask-deep-learning approach to few-shot HSI classification with unknown classes; used as a baseline and problem reference."},{"cited_title":"Few-shot open-set recognition of hyperspectral images with outlier calibration network,","cited_arxiv_id":null,"evidence_quote":"Supplies the outlier calibration network baseline that rejects unknown pixels via calibrated outlier scores."},{"cited_title":"Spectral-spatial mlp- like network with reciprocal points learning for open-set hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the reciprocal-points (RPL) baseline with learnable dynamic thresholds for open-set HSI classification."},{"cited_title":"Deep few-shot learning for hyperspectral image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the DFSL deep few-shot learning baseline, which the paper adapts with a softmax threshold for unknown rejection."},{"cited_title":"Zhong, X","cited_arxiv_id":null,"evidence_quote":"Provides the WHU-Hi-HanChuan dataset used in the main evaluation."},{"cited_title":"Morgan: Meta-learning- based few-shot open-set recognition via generative adversarial network,","cited_arxiv_id":null,"evidence_quote":"Provides the MORGAN baseline, a GAN-based meta-learning method for few-shot open-set recognition."},{"cited_title":"Extreme value meta-learning for few-shot open-set recognition of hyperspectral im- ages,","cited_arxiv_id":null,"evidence_quote":"Provides the EVML baseline, an extreme-value-theory meta-learning method that rejects unknown samples."}],"review_version":1}