{"id":"8bdd8a79-01a3-4a8f-bc3b-286a3bf939cc","arxiv_id":"2505.00805","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey of deep learning for hyperspectral wheat analysis, but it contains errors, off-topic papers, and an unreconciled overlap with the authors' own 2024 review.","lead":"This paper is a survey that tries to organize deep learning methods applied to hyperspectral images of wheat crops. It lists datasets, groups models by learning style, and covers classification, disease detection, nutrient estimation, and yield prediction.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'first wheat-specific review' claim is undercut by the paper's own corpus: many retained methods are generic HSI benchmark studies with no wheat experiment.","rationale":"Stress-testing the paper's own stated goal, the weakest link is not the survey's novelty as such but the unverified boundary of its corpus. For a firstness claim to be meaningful, the retained papers must actually be wheat-specific. The paper's own Figure 5 and Tables 6–7 show that many cited methods are evaluated on standard remote-sensing benchmarks, not wheat; Table 6's 'wheat crop class acc' is just the per-class accuracy for the wheat class in Indian Pines. The Section 2 protocol is simultaneously irreproducible (268 vs. 294 discarded; 193 vs. 173 retained; Google Scholar vs. four databases), and the paper does not clearly differentiate its contribution from the authors' own 2024 review of DL for agricultural HSI. These are internal inconsistencies rather than disagreements with field consensus. My read therefore reinforces rather than changes the reader's REJECT verdict, and a reference-level audit would cleanly settle whether the scope problem is as extensive as the taxonomy suggests.","tokens_in":38112,"tokens_out":9235,"duration_ms":92306,"concrete_test":"Conduct a reference-level audit: for each of the 51 taxonomy entries in Figure 5 and each row of Tables 6–7, retrieve the cited paper and determine whether it contains any wheat-specific experiment (Triticum aestivum kernels, spikes, canopy, or field trials) or only generic HSI benchmarks such as Indian Pines, Pavia University, Salinas, or Botswana. If more than about 15% of the audited entries are non-wheat generic HSI studies, then Section 2's inclusion criteria were not enforced and the paper's central claim of being the first wheat-specific review is unsupported. The same audit will also reveal whether the tables' reported 'wheat crop class acc' values correspond to actual wheat studies or merely to the wheat class inside Indian Pines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The survey's central claim is firstness: 'this is the first review to focus specifically on the intersection of HSI data and DL techniques for wheat crop analysis.' That claim depends on the Section 2 screening having produced a corpus of wheat-specific studies. The paper's own content contradicts this. Figure 5's 51-entry taxonomy and Tables 6–7 include many generic HSI-classification papers whose datasets are Indian Pines, Pavia, Salinas, Botswana, or KSC, not wheat; examples include refs. [66, 67, 68, 70, 72, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 95, 96, 98, 99, 101, 102, 103, 104, 105, 107, 108, 111, 112, 113]. Tables 6–7 even report 'wheat crop class acc' as the per-class accuracy for wheat inside Indian Pines, which is not a wheat-crop analysis. Section 2's counts are also irreproducible: the text says 268 articles discarded and 193 retained, while Figure 2 says 294 excluded and 173 selected, and Figure 2 labels the source as Google Scholar rather than the four databases named in the text. The paper also cites its own prior broad review of DL for agricultural HSI [147] without explaining how the present survey differs. If the screening did not enforce wheat specificity, the claimed first wheat-specific review is unsupported and the survey is a general HSI-DL review with a wheat veneer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of deep learning (DL) methods applied to hyperspectral imaging (HSI) for wheat crop analysis. It claims to be the first review focusing specifically on the intersection of HSI and DL for wheat crops, and it organizes the literature into supervised, semi-supervised, and unsupervised learning paradigms. The paper reviews hyperspectral sensors and platforms, public and private wheat datasets, DL architectures (CNNs, RNNs, Transformers, Mamba models, DBNs, SAEs, GANs, diffusion models), and applications such as wheat classification, disease detection, nutrient estimation, and yield estimation. It also discusses challenges and future directions and points to a GitHub repository for tracking state-of-the-art papers.","tokens_in":38373,"tokens_out":5370,"duration_ms":49798,"significance":"If the survey were accurate and comprehensive, it would be a valuable reference for the precision-agriculture and remote-sensing communities, bringing together datasets, a broad taxonomy of DL architectures, and application areas in a single place. The manuscript deserves credit for attempting a structured taxonomy, for compiling sensor and dataset tables, and for providing a public GitHub repository. However, the central claim of being the 'first review' on wheat-specific HSI+DL is not supported by the actual content: the methodology section gives irreproducible screening counts, and many retained papers are generic HSI classification studies evaluated on non-wheat benchmark scenes. These problems undermine the reliability of the survey as a reference and directly affect its core claim.","major_comments":[{"comment":"The screening counts are inconsistent: Section 2 states that 268 articles were discarded and 193 retained, whereas Figure 2 reports 294 excluded and 173 selected, and Figure 2 labels the source as 'Google Scholar' rather than the four databases named in the text. Because the size and provenance of the corpus are central to the survey's comprehensiveness claim, this discrepancy makes the methodology irreproducible and needs to be resolved with a single consistent PRISMA-style accounting.","section":"2 / Figure 2"},{"comment":"Many entries in the 51-item taxonomy correspond to generic HSI classification papers evaluated on benchmark scenes such as Indian Pines, Pavia, Salinas, Botswana, and KSC, with no wheat-crop experiment; examples include references [66, 67, 68, 70, 72, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 95, 96, 98, 99, 101, 102, 103, 104, 105, 107, 108, 111, 112, 113]. The 'wheat crop class acc' column in Tables 6-7 reports per-class accuracy for the wheat class inside Indian Pines, which is a land-cover benchmark, not a wheat-crop analysis. This directly contradicts the inclusion criteria stated in Section 2 and undermines the claim that the survey is wheat-specific.","section":"5.1-5.3 / Tables 6-7"},{"comment":"Table 3 is presented as a summary of public wheat-crop analysis datasets, but it includes Indian Pines and GHISA, which are general agricultural datasets where wheat is only one class among many. The text says the table includes 'six specifically related to wheat crop analysis,' yet only four entries (Soilborne Wheat Mosaic Virus, Wheat HyperSpectral, Early Detection of Crown Rot, DRUM) are wheat-focused. This mismatch further weakens the wheat-specific scope and makes the dataset inventory misleading.","section":"4.1 / Table 3"},{"comment":"Table 6 contains numerical errors that undermine its reliability as a reference, such as 'OA:IP=96.2.6%' for reference [86] and several rows reporting wheat per-class accuracy of 100% (e.g., references [68] and [79]). Combined with the inclusion of non-wheat benchmark papers, this suggests the summary tables were assembled without careful verification of the original studies.","section":"5.4 / Table 6"},{"comment":"The introduction claims that this is the first review focused specifically on HSI and DL for wheat crop analysis, but reference [147] is a 2024 review by the same research group on deep learning for hyperspectral image analysis in agriculture. The manuscript does not explain how the present survey differs from that prior review, so the novelty and firstness claims are not properly supported.","section":"1 / reference [147]"}],"minor_comments":[{"comment":"The phrase 'extract and analysis complex structures' should be corrected to 'extract and analyze complex structures.'","section":"Abstract"},{"comment":"The caption of Figure 11 reads 'Number of Published Articles by Year on DL with Hyperspectral Data in Wheat Crops,' but the figure itself is a pie chart showing the percentage distribution of application categories; these do not match.","section":"Figure 11"},{"comment":"There are numerous typographical and formatting errors, including 'have have' in Section 5.2.2, 'Univerity' in the affiliations, inconsistent 'UA V' versus 'UAV' spacing, and 'lLSTMs' in Section 7.","section":"Throughout"},{"comment":"The sensor name is written inconsistently as 'A VIRIS' and 'AVIRIS'; the standard spelling is AVIRIS.","section":"Tables 1-2 / text"},{"comment":"The abbreviation 'VNI' is defined as 'Very Near Infrared'; the standard abbreviation is VNIR.","section":"Table 3"},{"comment":"The GitHub repository URL appears only in the abstract; it would be helpful to include it in the conclusion or a dedicated data-availability statement.","section":"Abstract / Repository"}],"recommendation":"reject","confidential_remarks":"The paper has fundamental scope-integrity problems: the screening numbers are internally inconsistent, and a large portion of the retained corpus consists of generic HSI classification studies that are not about wheat crops. The central firstness claim is therefore unsupported, and the survey cannot currently serve as a reliable wheat-specific reference. The authors would need to redo the screening, remove non-wheat papers, and substantially revise the taxonomy and summary tables, which in my view goes beyond a normal revision. I also note that the self-citation overlap with reference [147] should be explicitly disclosed if a resubmission is considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe best part of this survey is not the taxonomy—that's the standard supervised/semi-supervised/unsupervised split you've seen in other HSI reviews—but the wheat-specific tables. Tables 4 and 5 compile private Fusarium and yellow rust datasets with sensors, wavelengths, and key bands, and Tables 8–11 summarize wheat classification, disease detection, nutrient estimation, and yield estimation papers with concrete results. A practitioner looking for a starting point on wheat HSI would find those genuinely useful, and the GitHub tracking repo is a nice touch.\n\nThe problem is the survey's load-bearing claim of firstness. The screening is irreproducible. The text says 268 articles were discarded and 193 retained; Figure 2 says 294 excluded and 173 selected, and the figure labels the source as Google Scholar while the text names four databases. Those are not cosmetic errors; they make the methodology impossible to check.\n\nWorse, the corpus isn't wheat-specific. The taxonomy and Tables 6–7 are full of generic HSI papers evaluated on Indian Pines, Pavia, Salinas, Botswana—papers like HybridSN, SSRN, and Mamba models that have no wheat experiment. The tables even report 'wheat crop class acc' as the per-class accuracy of the wheat class inside Indian Pines. That is a standard benchmark class, not wheat crop analysis. So the 'first wheat-specific review' claim is undercut by the paper's own content.\n\nThe self-citation doesn't help: reference [147] is the authors' own 2024 review on DL for HSI in agriculture, cited in passing without explaining how the present survey differs. That needs to be addressed directly.\n\nThere are also smaller errors—typos like '96.2.6%' and 'AO:IP=98.3%'—which, in a reference work, undermine trust.\n\nSo the verdict is reject in current form. But it's not worthless. The wheat-specific dataset collection and the application summaries are worth salvaging. If the authors redo the screening with a genuinely wheat-specific inclusion criterion, reconcile the counts, and clearly differentiate from their prior review, this could become a useful survey for the precision-agriculture community.\n\nI'd send it to a serious peer review rather than desk-reject, because a referee can push for the needed rework and the underlying material has value. But I wouldn't cite it in its current state.","headline":"A useful wheat-specific dataset collection buried under an irreproducible screening process and a corpus of generic HSI papers.","tokens_in":38970,"tokens_out":4272,"would_cite":false,"duration_ms":42061,"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 survey claims to be the first comprehensive review of deep learning methods on hyperspectral imaging for wheat crop analysis, organizing 193 studies into a structured taxonomy.","keywords":["hyperspectral imaging","deep learning","wheat crops","disease detection","yield estimation","nutrient estimation","literature survey","precision agriculture"],"falsifier":"Re-run the stated database searches (ScienceDirect, IEEE Xplore, SpringerLink, MDPI; 2015–2024; 'Hyperspectral Imaging' OR 'HSI' AND 'Deep Learning' AND 'Wheat Crops') and count how many of the 193 retained articles actually apply their method to wheat data. A decisive test: check the Section 5 method entries—for example the Mamba papers [80]–[85] and transformer papers [77]–[79]—and see whether wheat appears anywhere beyond a single class in the Indian Pines benchmark; if most never address wheat, the survey's wheat-specific claim would collapse.","tokens_in":37857,"feed_emoji":"🌾","tokens_out":5618,"duration_ms":50247,"temperature":0.7,"pith_summary":"This review paper sets out to establish that deep learning applied to hyperspectral imaging (HSI) for wheat crops has grown into a distinct research area that deserves its own comprehensive survey. It claims to be the first review focused specifically on that intersection, covering work published from 2015 to 2024 and synthesizing 193 retained articles. The paper organizes the field into a taxonomy of learning paradigms—supervised, semi-supervised, and unsupervised—and maps each to model families such as CNNs, RNNs, Transformers, Mamba networks, DBNs, SAEs, GANs, and diffusion models. It then ties those methods to four application areas: crop classification, disease detection and monitoring, nutrient estimation, and yield estimation. A sympathetic reader would care because a reliable map of methods, datasets, and open problems could guide practitioners choosing models and expose where public data are lacking.","feed_headline":"First survey maps deep learning for hyperspectral wheat analysis","feed_subtitle":"Organizes 193 studies into a taxonomy spanning classification, disease detection, nutrient and yield estimation.","key_machinery":"The organizing device is a three-level taxonomy built from learning paradigm, model family, and application. Supervised learning covers CNNs, RNNs, DBNs, SAEs, Transformers, and Mamba; semi-supervised learning covers pseudo-labeling and generative adversarial networks; unsupervised learning covers DBNs, SAEs, and diffusion models. The taxonomy is cross-referenced to four wheat applications—classification, disease detection, nutrient estimation, yield estimation—and anchored by a dataset inventory that includes Indian Pines, GHISA, DRUM, and several Fusarium and yellow-rust collections. The taxonomy does the paper's main work: it turns a scattered literature into a structured map, and it is the basis for the claim that this intersection is a coherent field with identifiable trends and gaps.","core_discovery":"On its own terms, the paper's central claim is that it provides the first comprehensive review of deep learning methods applied to hyperspectral imaging data for wheat crop analysis. To support that claim, it assembles and categorizes 193 papers from four literature databases, groups the methods by learning paradigm and architecture, and summarizes the benchmark datasets—six public and a set of private collections focused mainly on Fusarium head blight and yellow rust. The survey finds that classification dominates the literature at 68% of the retained work, with disease detection at 13.1%, nutrient estimation at 11.8%, and yield estimation at 7.2%, and it argues that the field's main bottlenecks are scarce labeled data and the computational cost of high-dimensional HSI processing.","pith_inferences":["The survey treats general HSI benchmarks such as Indian Pines as wheat-relevant because they contain a wheat class; a stricter reading of 'wheat crop analysis' would exclude many of the cited architecture papers, so the claimed 193-article corpus likely overcounts wheat-specific studies.","If the field really is at the maturity the survey describes, a natural next step would be a shared benchmark with standardized splits and evaluation metrics for wheat-specific tasks, allowing reported accuracy numbers to be compared fairly across architectures.","The maintained companion repository could evolve into a living index; its usefulness is testable by whether it continues to be updated with post-2024 papers."],"forward_implications":["Practitioners get a single reference map of deep-learning models and datasets for hyperspectral wheat analysis, organized by learning paradigm and application.","The survey's statistics make the field's imbalance explicit: classification accounts for 68% of published work, while yield estimation is under-served at 7.2%.","Because most wheat-disease datasets are private, the paper's call for large, open, region-diverse datasets is a concrete agenda item for the community.","The review identifies Mamba and transformer architectures as the current state of the art and points to lightweight models and self-supervised learning as priority research directions."],"supporting_citations":[{"why":"An extensive review of hyperspectral image classification and prediction that the paper positions itself against to establish the gap it claims to fill.","marker":"[10]"},{"why":"A review of deep learning used in hyperspectral image analysis for agriculture, used to show that existing surveys neglect wheat-specific challenges.","marker":"[11]"},{"why":"A systematic review of hyperspectral imaging with machine and deep learning for agricultural applications, another general survey the paper distinguishes itself from.","marker":"[12]"},{"why":"A review of plant disease diagnosis using deep learning on aerial hyperspectral images, domain-adjacent to wheat disease detection and used to justify the need for a wheat-focused review.","marker":"[13]"},{"why":"The Indian Pines dataset, the benchmark whose wheat class supplies the wheat-crop label for many of the classification results reported in the survey.","marker":"[42]"},{"why":"Bauriegel et al.'s Fusarium head blight hyperspectral dataset, a core private dataset for the most-studied wheat disease in the survey.","marker":"[48]"},{"why":"Zhang et al.'s UAV-based yellow rust hyperspectral dataset, a core private dataset for yellow rust detection in wheat.","marker":"[52]"}],"fun_headline_variants":["First review maps 193 deep learning wheat HSI studies","Deep learning wheat analysis: first hyperspectral survey","Wheat HSI deep learning: taxonomy of 193 papers","Survey covers deep learning for wheat hyperspectral imaging","First comprehensive DL survey on wheat HSI analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the screening process genuinely captured wheat-specific hyperspectral deep-learning studies and that the 193 retained papers were summarized faithfully; if many included papers only use a generic HSI benchmark containing a wheat class, or if the per-paper summaries misrepresent their methods, the survey's map is distorted.","fun_headline_variants_meta":{"raw":{"variants":["First review maps 193 deep learning wheat HSI studies","Deep learning wheat analysis: first hyperspectral survey","Wheat HSI deep learning: taxonomy of 193 papers","Survey covers deep learning for wheat hyperspectral imaging","First comprehensive DL survey on wheat HSI analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000342,"raw_usage":{"total_tokens":1863,"prompt_tokens":908,"completion_tokens":955,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":879}},"tokens_in":524,"tokens_out":955,"duration_ms":10051,"temperature":1.0,"reasoning_tokens":879,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:33:17.297504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the stated database searches (ScienceDirect, IEEE Xplore, SpringerLink, MDPI; 2015–2024; 'Hyperspectral Imaging' OR 'HSI' AND 'Deep Learning' AND 'Wheat Crops') and count how many of the 193 retained articles actually apply their method to wheat data. A decisive test: check the Section 5 method entries—for example the Mamba papers [80]–[85] and transformer papers [77]–[79]—and see whether wheat appears anywhere beyond a single class in the Indian Pines benchmark; if most never address wheat, the survey's wheat-specific claim would collapse.","supporting_citations":[],"review_version":1}