{"id":"153d7ce4-186a-41b4-942d-8f36b94b73d4","arxiv_id":"2508.19307","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A thesis reports roughly 99 percent CNN accuracy on five rice grain varieties and four rice leaf diseases with LIME/SHAP overlays, replicating prior published results on the same public datasets.","lead":"This paper, a master's thesis posted as a preprint, reports about 99 percent accuracy for classifying five rice grain varieties and four rice leaf diseases with convolutional networks, plus LIME and SHAP visual explanations. It is a routine application: earlier cited work already reached 100 percent accuracy on the same grain dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported near-perfect accuracy is untraceable: cited dataset references are method papers and the described preprocessing/model is internally inconsistent, so no reproducible pipeline connects the manuscript to Tables 4.4, 5.3, and 5.4.","rationale":"The reader's verdict is REJECT, and my stress-test identifies the same load-bearing weakness: the reported near-perfect accuracy and XAI transparency depend on being able to connect the text to a real dataset and a defined pipeline. The dataset citations point to method papers rather than datasets, the preprocessing steps are not unambiguously tied to the experimental results, and several model/input-size/hyperparameter statements conflict across sections. These are not stylistic issues; they are missing links in the evidence chain.\n\nI considered whether the XAI aspect itself is the more central problem—the paper claims interpretability as its core contribution but provides only qualitative screenshots and no fidelity or usefulness metric. That is a real weakness, but it is secondary: if the classification numbers cannot be traced to a concrete dataset and pipeline, then the explanations, even if faithful, explain an unverified model. The provenance/reproducibility concern is therefore the most load-bearing.\n\nI am not accusing the authors of fabrication. The reported accuracies are plausible and consistent with prior work, and the confusion matrices reported are internally coherent (e.g., 1,478 of 1,500 test images correctly classified in several grain classes; 147, 149, and 123 correct in disease classes). The issue is that the paper fails as a verifiable research artifact: no code, no dataset identifiers, no architecture checkpoints, and internally inconsistent method descriptions. A direct reproducibility check—re-running the described pipeline on the named datasets—would settle whether the empirical claims actually hold.\n\nBecause this concern reinforces rather than overturns the reader's REJECT verdict, the appropriate recommendation is UNCHANGED.","tokens_in":31658,"tokens_out":4361,"duration_ms":49965,"concrete_test":"Reproduce the two experiments using only the information in the manuscript: (1) download the 'Rice Image Dataset' (75,000 images, five classes) and the 'Rice Leaf Disease Images Dataset' (6,000 images, four classes) from Kaggle; (2) apply the stated 80/10/10 split, Canny edge detection, segmentation, normalization to [0,1], and rotation/flip augmentation; (3) train the exact Table 4.1 CNN for grains and the Section 5.1 CNN for diseases, with batch size 32, Adam, L2 regularization, and image size 50x50; (4) compare reproduced F1-scores and accuracies against Tables 4.4, 5.3, and 5.4. If the cited references [47] and [77] do not resolve to downloadable datasets, or if reproduced metrics differ materially (e.g., more than 1–2 percentage points), the reported results are untraceable. As an additional check, run the disease model at 224x224 as described in Section 3.4.3 to see which input siz","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that the CNN-based framework classifies five rice grain varieties and four rice leaf diseases with near-perfect accuracy and that LIME/SHAP make the decisions transparent. The load-bearing condition is that the data and pipeline described are the data and pipeline actually used. That condition is not met in a verifiable way.\n\nSection 3.3.1 states that the 75K rice grain dataset is the 'Rice Image Dataset' and cites [47]; reference [47] is 'Searching for MobileNetV3,' a network architecture paper, not a dataset. Section 3.4.1 similarly cites [77] for the 6K rice leaf disease dataset; [77] is a deep-feature SVM method paper, not a dataset. These are not minor citation errors: they make the provenance of every reported accuracy number untraceable from the manuscript alone.\n\nThe pipeline itself is also under-specified and internally inconsistent. The text describes Canny edge detection, segmentation, normalization, and rotation/flip augmentation, but never states whether the CNN was trained on raw, edge-detected, segmented, or augmented images. Table 4.1 shows the input size inferred as 50x50x1, while Section 4.2 says 50x50 with 3 RGB channels. For disease classification, Section 5.1 gives a 50x50 image size with 3 channels, but the proposed model description says 224x224 RGB input. Table 5.4 reports ResNet-50, VGG-16, MobileNet-V2, and DenseNet121, while Chapter 6 discusses InceptionResNetV2 and DenseNet201 results that do not appear in any results table. The custom CNN's own accuracy is never stated for disease classification, even though the confusion matrix and ROC curve are attributed to it.\n\nAdditionally, Figure 3.5 is captioned in the list of figures as 'Proposed architecture for crowd counting using self-supervised training,' while the body caption calls it the rice grain classification framework. Section 1.8 also describes the thesis as being about 'crowd scene analysis.' These are mechanical signs that the methodology chapter was","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript, submitted as an arXiv preprint in cs.CV, is an unrevised Master's thesis that proposes a deep-learning pipeline for two crop-analysis tasks: classifying five rice grain varieties (Arborio, Basmati, Ipsala, Jasmine, Karacadag) on a 75,000-image dataset, and classifying four rice leaf diseases (Bacterial Blight, Blast, Brown Spot, Tungro) on a 6,000-image dataset. The authors report near-perfect accuracy: F1-scores of 0.986–0.997 for grains (Table 4.4), 99.82% accuracy for ResNet-50 on disease classification (Table 5.4), and AUC = 1.00 for disease classification (Figure 5.3). They augment the classifiers with LIME and SHAP explanations to make predictions transparent. The central load-bearing claims are the accuracy numbers and the interpretability of the decisions.","tokens_in":31927,"tokens_out":3521,"duration_ms":34877,"significance":"If substantiated, the results would be a competent but not methodologically novel application of standard CNN classifiers and post-hoc XAI tools to two public agricultural datasets. The confusion-matrix counts and class-wise scores are internally plausible: the test-set totals are consistent with an 80/10/10 split of 75,000 and 6,000 images, and the reported F1-scores are in line with published results on the same rice-grain dataset. The explicit use of LIME and SHAP side-by-side is a useful practical comparison. However, the manuscript's central empirical claims are not verifiable from the text alone: the datasets are cited to method papers rather than dataset sources, the described preprocessing is not consistently connected to the reported model inputs, and several models discussed in the narrative (InceptionResNetV2, DenseNet201) do not appear in the results tables. No code, data links, or reproducibility artifacts are provided. These issues are load-bearing because the accuracy claims cannot be traced to a concrete, reproducible pipeline.","major_comments":[{"comment":"The rice grain dataset is claimed to be the 'Rice Image Dataset' on Kaggle, cited as [47], but reference [47] is Howard et al., 'Searching for MobileNetV3' (ICCV 2019), a network-architecture paper, not a dataset. Similarly, Section 3.4.1 cites [77] for the rice leaf disease dataset, but [77] is Sethy et al., a 'Deep feature based rice leaf disease identification using support vector machine' method paper. These are not cosmetic citation errors: the dataset identity (splits, preprocessing, class balance) is the foundation of every accuracy number in Tables 4.4, 5.3, and 5.4. The authors must cite the actual dataset URLs or DOIs and specify the exact versions used; otherwise, the reported numbers cannot be independently checked or reproduced.","section":"Section 3.3.1 / Reference [47]"},{"comment":"The grain-classification pipeline is internally inconsistent about what the CNN actually saw. Section 3.3.3 describes Canny edge detection, segmentation, normalization, and rotation/flip augmentation, but Table 4.1 infers an input size of '50×50×1' while Table 4.2 states 'Image Size 50x50' with '03 (RGB)' channels. More importantly, the text never states whether the model was trained on raw, edge-detected, segmented, normalized, or augmented images. Figure 4.3 shows the 'Original Image' and LIME overlay, not the edge/segmented version. Without this mapping, the preprocessing descriptions are decorative and the provenance of Table 4.4's numbers is untraceable.","section":"Section 4.1 / Table 4.1 vs. Section 4.2 / Table 4.2"},{"comment":"The disease-classification model description is contradictory across chapters. Section 5.1.2/Table 5.1 specify a 50×50 RGB input for the CNN, while Section 3.4.3 describes the proposed CNN as using '224x224 pixels' RGB input. Table 5.4 reports results for ResNet-50, VGG-16, MobileNet-V2, and DenseNet121, but Chapter 6 states that 'InceptionResNetV2 and DenseNet201 have shown that they are better in reaching high levels of accuracies' without those models appearing in any results table. Figure 5.8 is captioned as the 'Custom CNN' SHAP output, but its architecture and relation to the models in Table 5.4 are unspecified. The reader cannot determine which model produced Table 5.3 or the confusion matrix in Figure 5.2.","section":"Section 5.1 / Table 5.1 vs. Section 3.4.3"},{"comment":"The abstract and text claim 'outstanding area under the ROC curve in every class' and AUC = 1.00. Figure 5.3 shows a micro-average ROC curve only; no per-class AUC values, confidence intervals, or statistical comparison are provided. On a near-perfect four-class problem, a micro-average AUC of 1.00 is plausible but says nothing about per-class uncertainty, especially with a small test set (≈150 images per class). The claim as stated is not supported by the reported evidence. The authors should report per-class AUC with confidence intervals or bootstrap estimates, or soften the claim.","section":"Section 5.3 / Figure 5.3"},{"comment":"Figure 3.5, referenced in Section 3.2 as the proposed rice grain classification system, bears the caption 'Proposed architecture for crowd counting using self-supervised training. Stage 1 utilizes the rotation task for self-supervised training, and stage 2 uses distribution matching for density estimation.' This is unrelated to the rice-grain task and appears to be a template artifact. Its presence in the methodology chapter compounds the provenance problem: if the system diagram is not the described system, the reader cannot trust the surrounding experimental description either.","section":"Figure 3.5 caption"}],"minor_comments":[{"comment":"The manuscript contains numerous copy-paste artifacts: Chapter 1 refers to 'crowd scene analysis' in the outline (Section 1.8), Section 2.1 says 'A brief review of the current approaches along with their shortcomings for crowd scene analysis,' and Figure 1.6 is captioned 'Significance of Crowd Scene Analysis' while the surrounding text discusses rice grain classification. These should be corrected.","section":"Throughout"},{"comment":"The class name 'Burst' appears in Table 5.2, while the text and elsewhere use 'Blast.' This is likely a typo but should be fixed because it is a class label.","section":"Table 5.2"},{"comment":"The optimizer is described as 'Adamax' in Section 3.3.4 but 'ADAM' in Table 4.2 and Section 4.2. Please state the exact optimizer and learning rate.","section":"Section 3.3.4 / Table 4.2"},{"comment":"References are internally duplicated (e.g., [73] appears twice with different content, [111] duplicates [110]), and several reference numbers do not match the cited content (e.g., [47], [77] as noted above). A systematic reference audit is needed.","section":"References"},{"comment":"Figures 5.4 and 5.7 both have captions saying 'MobileNet-V2' for the LIME visualization, which is certainly a labeling error; one of them presumably corresponds to a different model or should be removed. This makes the XAI comparison confusing.","section":"Figures 5.4 and 5.7"}],"recommendation":"reject","confidential_remarks":"The manuscript is an unrevised MS thesis with multiple template artifacts, broken dataset citations, and an internally inconsistent experimental description. The core empirical claims are plausible and likely reproduce the published state of the art on the Koklu et al. rice-grain dataset, but as submitted the paper does not describe a verifiable pipeline. This is not a case where a small revision can fix the load-bearing issues: the dataset provenance and the model/preprocessing mapping would need to be reconstructed from scratch, and the reported results tables would need to be re-generated with a clearly specified protocol. I would not consider this manuscript for publication in its current form, even as a workshop paper, because the accuracy claims are untraceable from the manuscript alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is an MS thesis posted to arXiv, and it reads like one: the empirical claims are plausible, the numbers are internally consistent, and the accuracies are right in line with prior work on the same datasets. The grain classification is a reproduction of Koklu et al., the disease classification is comparable to several cited studies, and the confusion matrices roughly add up. I have no reason to think the results are fabricated.\n\nThat said, the paper fails as a research artifact for verifiable, mechanical reasons. The dataset citations are method papers, not datasets: Section 3.3.1 cites [47] for the 75K rice image dataset, but [47] is the MobileNetV3 architecture paper; Section 3.4.1 cites [77] for the disease dataset, but [77] is an SVM-based rice disease paper. Those are not minor citation errors—they make the provenance of every reported accuracy untraceable from the manuscript itself.\n\nThe methodology chapter also contains recycled material from an unrelated crowd-analysis document. The list of figures calls Figure 3.5 a crowd-counting architecture, and Section 1.8 describes the thesis as being about crowd scene analysis. Chapter 6 discusses InceptionResNetV2 and DenseNet201 results that appear nowhere in the experiments. The promised conclusions chapter is missing, and the TOC has unresolved bookmark placeholders.\n\nInterpretability, the stated core contribution, is supported only by qualitative LIME/SHAP screenshots with no fidelity or usefulness metric. AUC 1.00 is reported without confidence intervals, and the best epoch is read off the validation curve.\n\nThe soft spots are not minor—they are load-bearing. Because the dataset provenance and pipeline are under-specified and internally inconsistent, I cannot tell from the paper what images were actually used, what preprocessing was applied, or how the reported numbers were generated. That is a problem no amount of plausible accuracy can fix.\n\nWho is this for? Maybe a master's committee evaluating the student's familiarity with standard tools. As a contribution to the literature, it falls short. I would not send it to peer review in this state. It needs a full rewrite: correct dataset citations, remove the crowd-counting material, add the missing results, and provide a real evaluation of the XAI component.\n\nMy verdict is reject, but not because the topic is uninteresting—because the paper is not yet a coherent research artifact.","headline":"A thesis draft with plausible numbers but untraceable provenance; the crowd-counting leftovers make it impossible to verify what was actually run.","tokens_in":32664,"tokens_out":1887,"would_cite":false,"duration_ms":22185,"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":"A CNN pipeline is reported to classify five rice grain varieties and four rice leaf diseases with near-perfect accuracy, with LIME and SHAP overlays revealing the image regions behind each prediction.","keywords":["rice grain classification","rice leaf disease detection","convolutional neural networks","explainable AI","LIME","SHAP","ResNet-50","transfer learning"],"falsifier":"Re-run the described pipeline on the specific public datasets the text names (after locating the correct dataset sources) and check whether per-class F1 scores for the five grains and ResNet-50's 99.82% disease accuracy reproduce; also verify that the cited references actually contain those datasets. Failure to reproduce the confusion matrices or the reported ROC/AUC values would settle the claim.","tokens_in":31375,"feed_emoji":"🌾","tokens_out":7516,"duration_ms":78539,"temperature":0.7,"pith_summary":"This thesis tries to establish that a deep-learning pipeline can separate five commercially important rice grain varieties (Arborio, Basmati, Ipsala, Jasmine, Karacadag) and four common rice leaf diseases (Bacterial Blight, Blast, Brown Spot, Tungro) with near-perfect accuracy on public image datasets. The reported numbers are per-class F1 scores of 0.986–0.997 for grains and a top disease accuracy of 99.82% from ResNet-50. The authors argue that adding LIME and SHAP explanations makes these automated decisions transparent, showing farmers and quality analysts which visual features drove each classification. If the claim holds, the practical payoff is real: grain grading and early disease diagnosis could move from slow manual inspection to camera-based screening with an audit trail.","feed_headline":"CNN classifies five rice grains and four leaf diseases at 99%+","feed_subtitle":"LIME and SHAP overlays show which grain or leaf features drove each prediction.","key_machinery":"The machinery is a two-stage pipeline. First, a feature extractor: a small custom CNN (two Conv2D blocks, max pooling, a 32-unit dense layer, and a five-unit softmax head) for grains, and pretrained backbones (ResNet-50, VGG16, MobileNet-V2, DenseNet121) for leaf diseases, with Canny edge detection and segmentation applied before training. Second, an explainability stage: LIME constructs local surrogate models to highlight the image regions that pushed a prediction, and SHAP assigns per-pixel additive attributions. The claim is that these two stages together turn a black-box classifier into a transparent crop-analysis tool.","core_discovery":"On the paper's own terms, the central discovery is that a single deep-learning workflow—image preprocessing (grayscale conversion, Canny edge detection, segmentation, normalization, rotation and flipping), a custom CNN for grains, pretrained CNNs for diseases, then LIME and SHAP post-hoc explanations—separates the five grain classes with per-class F1 scores between 0.986 and 0.997, and separates the four disease classes with ResNet-50 reaching 99.82% accuracy. The authors take these results to show that standard, non-specialised CNN architectures are sufficient for this level of crop analysis, and that the LIME/SHAP overlays make the basis of each prediction visible to non-experts.","pith_inferences":["Beyond the paper: the decisive test for the explainability claim is behavioural—whether LIME/SHAP overlays help a human inspector catch wrong classifications faster or more accurately than a bare confidence score.","Beyond the paper: the pipeline should be re-run on field-captured images with varied lighting, angles, and backgrounds; the reported accuracy is on curated public images, and transfer to realistic field conditions is an open question.","Beyond the paper: the manuscript's internal inconsistencies in dataset citations leave the provenance of the accuracy numbers unresolved until the exact datasets are located and the experiments are rerun."],"forward_implications":["If the reported numbers hold, rice grain quality grading could be automated with camera-based CNN inspection, replacing or prioritising manual sorting.","Early leaf-disease diagnosis becomes feasible on still images, giving farmers a low-cost second opinion before the disease spreads.","The LIME/SHAP overlays give quality analysts a concrete audit trail: each automated decision is accompanied by the image regions that drove it.","The model comparison points to ResNet-50 as the strongest of the tested pretrained backbones for this disease dataset, informing future architecture choices."],"supporting_citations":[{"why":"Cited as the source of the 75,000-image rice grain dataset used for training and testing.","marker":"[47]"},{"why":"Supplies the prior CNN/DNN/ANN rice classification baseline that this work's grain results sit next to.","marker":"[60]"},{"why":"Cited as the source of the 6,000-image rice leaf disease dataset used for training and testing.","marker":"[77]"},{"why":"Defines the LIME explanation method used to highlight influential image regions.","marker":"[48]"},{"why":"Defines the SHAP attribution method used for pixel-level explanations.","marker":"[51]"},{"why":"Supplies the machine-learning rice leaf disease baselines (J48, KNN, NB, LR) this work compares against.","marker":"[70]"},{"why":"Provides the RiceNet custom CNN and EfficientNetB0 transfer-learning baselines for disease detection.","marker":"[71]"},{"why":"Provides the CNN-based rice blast recognition baseline that motivates the CNN approach.","marker":"[75]"}],"fun_headline_variants":["Explainable AI identifies five rice types and four leaf diseases at 99%+","LIME and SHAP light up the features behind rice grain and disease AI","One workflow: classify rice grains and diseases with visible reasoning","Deep learning hits 99% on rice grains and diseases, with proof"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the pipeline described in the manuscript is the one actually run on the named public datasets; this premise is currently unverifiable because Section 3.3.1 cites the 75K grain dataset as reference [47], which is a MobileNetV3 architecture paper, and Section 3.4.1 cites the 6K disease dataset as reference [77], which is a rice-disease SVM paper, so the provenance of the reported accuracy numbers is not established by the text itself.","fun_headline_variants_meta":{"raw":{"variants":["Explainable AI identifies five rice types and four leaf diseases at 99%+","LIME and SHAP light up the features behind rice grain and disease AI","One workflow: classify rice grains and diseases with visible reasoning","Deep learning hits 99% on rice grains and diseases, with proof"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000342,"raw_usage":{"total_tokens":1769,"prompt_tokens":846,"completion_tokens":923,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":844}},"tokens_in":590,"tokens_out":923,"duration_ms":9334,"temperature":1.0,"reasoning_tokens":844,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:14:11.592487+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the described pipeline on the specific public datasets the text names (after locating the correct dataset sources) and check whether per-class F1 scores for the five grains and ResNet-50's 99.82% disease accuracy reproduce; also verify that the cited references actually contain those datasets. Failure to reproduce the confusion matrices or the reported ROC/AUC values would settle the claim.","supporting_citations":[{"cited_title":"Searching for mobilenetv3,","cited_arxiv_id":null,"evidence_quote":"Cited as the source of the 75,000-image rice grain dataset used for training and testing."},{"cited_title":"Classification of rice varieties with deep learning methods,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior CNN/DNN/ANN rice classification baseline that this work's grain results sit next to."},{"cited_title":"Deep feature based rice leaf disease identification using support vector machine,","cited_arxiv_id":null,"evidence_quote":"Cited as the source of the 6,000-image rice leaf disease dataset used for training and testing."},{"cited_title":"\" Why should i trust you?","cited_arxiv_id":null,"evidence_quote":"Defines the LIME explanation method used to highlight influential image regions."},{"cited_title":"A unified approach to interpreting model predictions,","cited_arxiv_id":null,"evidence_quote":"Defines the SHAP attribution method used for pixel-level explanations."},{"cited_title":"Rice leaf disease detection using machine learning techniques,","cited_arxiv_id":null,"evidence_quote":"Supplies the machine-learning rice leaf disease baselines (J48, KNN, NB, LR) this work compares against."},{"cited_title":"Automated recognition of rice grain diseases using deep learning,","cited_arxiv_id":null,"evidence_quote":"Provides the RiceNet custom CNN and EfficientNetB0 transfer-learning baselines for disease detection."},{"cited_title":"Rice blast disease recognition using a deep convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"Provides the CNN-based rice blast recognition baseline that motivates the CNN approach."}],"review_version":1}