Handling imbalance and few-sample size in ML based Onion disease classification

Abhijeet Manoj Pal; Rajbabu Velmurugan

REVIEW 4 major objections 5 minor 15 references

Handling imbalance and few-sample size in ML based Onion disease classification

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A compact CNN with attention and CutMix classifies eight onion disease/pest classes at 96.90% accuracy and 0.96 macro F1.

desk verdict Competent applied classifier paper with a real field dataset and a useful ablation, but the headline 96.90% is selected from seven configurations on one split, so it needs a statistical rework before the number can be trusted. read the letter →

arxiv 2509.05341 v1 pith:QBBNTVRT submitted 2025-09-01 cs.CV cs.LG

Abhijeet Manoj Pal , Rajbabu Velmurugan This is my paper · ORCID

classification cs.CVcs.LG

keywords oniondiseaseclassificationmulti-classplantclassimbalanceCutMixaugmentationCBAMattentionweightedcross-entropylossDenseNet-121

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that a single image classifier can tell apart seven onion diseases and pests plus healthy plants, instead of only saying healthy or unhealthy. The proposed model is a pre-trained DenseNet-121 augmented with a convolutional block attention module (CBAM), weighted cross-entropy loss, and CutMix augmentation; the authors report 96.90% overall accuracy and a 0.96 macro F1 on a field dataset of 5,330 images. Their message is that the imbalance in real onion-field data can be handled at the loss and augmentation level, without resampling the dataset or locating disease regions first. If correct, this gives an annotation-free, lightweight classifier that is practical for targeted onion pest management.

What carries the argument

The load-bearing mechanism is the joint effect of CBAM and CutMix on a DenseNet-121 backbone trained with weighted cross-entropy. CBAM is an attention module that first reweights feature channels and then highlights important spatial regions, letting the network focus on the informative parts of a leaf without explicit localization. CutMix generates a new training image by cutting a patch from one image, pasting it into another, and setting the target label as the area-weighted mixture of the two labels. Weighted cross-entropy assigns each class a weight equal to max_count divided by the class count. Together, these components let an 8-million-parameter pre-trained network absorb a skewed re

What would settle it

Retrain the best configuration on the same dataset across five different random 80/20 splits and record the mean and spread of macro F1. If the average F1 falls below about 0.90, or the spread overlaps the plain DenseNet-121 baseline, then the attention-and-CutMix combination is not the real driver of the reported performance.

Watch

Extended reading notes

Core claim

The central claim is that the combination of DenseNet-121, CBAM attention, weighted cross-entropy, and CutMix beats other configurations on the same dataset, reaching 96.90% accuracy and 0.96 macro F1 across eight classes. The paper further claims that weighted cross-entropy outperforms both focal loss and an imbalanced-dataset sampler, and that CutMix is the single most effective augmentation for class imbalance. It also merges the Anthracnose and Twister labels into one class on pathological evidence, redefining the task from nine to eight classes and reporting per-class accuracies from 0.87 to 1.00.

Load-bearing premise

The central assumption is that the single holdout split of this 5,330-image dataset reflects how the model will perform on new onion-field images; if that split is unrepresentative, the reported 96.90% accuracy and 0.96 F1 will not hold up in practice.

Editorial extensions

If this is right

The same model can be used at inference on full leaf images, with no object-detection step and no region annotations.
Class imbalance is handled inside the loss and the augmentation rather than by undersampling or oversampling the data.
The eight-class output is more actionable than a binary healthy/unhealthy decision because each disease or pest needs a different treatment.
The merging of Anthracnose and Twister on external evidence simplifies the label space and improves class-wise accuracy.
With roughly 8 million parameters, the winning architecture is light enough to consider for IoT or field-deployable devices.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The CBAM-plus-CutMix-plus-weighted-CE recipe is a plausible template for other imbalanced plant-disease datasets, but the paper does not demonstrate that transfer.
The reported numbers are point estimates from one holdout split; repeated-seed evaluation would show how stable the 96.90/0.96 figures are.
Testing the model on images from other farms, seasons, or camera types would clarify whether the accuracy is tied to this dataset's collection conditions.
Because the dataset was collected with DSLR and smartphone cameras, a natural extension is to benchmark the model on lower-resolution or drone imagery.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Hi,

Here's my read of the onion disease paper. It's a solid applied contribution: a real multi-class field dataset (5330 images, eight classes after merging Anthracnose/Twister with a defensible external citation), a systematic comparison of DenseNet-121 with CBAM, weighted CE, and several augmentation pipelines, and honest reporting of class-wise metrics and confusion matrices. Nothing here is new as a method — every component is standard — but that's not the point. The value is the dataset and the ablation for a domain that mostly has binary healthy/unhealthy classifiers.

The soft spot is the evaluation protocol, and it's the load-bearing one. Section IV says they did an 80/20 split and five-fold CV for tuning, but Table II reports one point estimate per configuration. If those numbers all come from the same held-out 20%, then choosing the best row (D121+CBAM+WCE+CutMix at 96.90%) is selecting the max over seven models on one split. That's optimistic bias. The minority classes make it worse: Basal Rot has 140 images total, so about 28 test images; 0.87 accuracy is 24 correct. A different split could move macro-F1 by several points. No code, data, seed, or variance estimates, so the exact split and augmentation order cannot be independently checked.

The comparison to Raj et al. is also apples-to-oranges — they do detection with bounding boxes on five classes, you do classification on eight — so claiming 'better results' is not a fair direct comparison. They do acknowledge the task difference, so this is minor.

None of this kills the paper. The central claim is plausible, the method is competent, and the authors clearly know the imbalance literature. The fix is straightforward: report mean±std over multiple splits (or at least a fixed seed with validation-based model selection) and release the code and data. I'd send this to peer review with that requirement. It's not a breakthrough, but it's a useful, honest piece of applied work for the plant-disease-imaging community.

Referee Report

4 major / 5 minor

Summary. The paper addresses multi-class onion disease and pest classification from field images, motivated by the need to move beyond binary healthy/unhealthy detection. The authors compare ResNet-50 and DenseNet-121 with different augmentation pipelines (normal, imbalanced sampler, Albumentations, CutMix) and loss functions (weighted cross-entropy, focal loss), and add CBAM attention to DenseNet-121. After merging Anthracnose and Twister into one class, they report that DenseNet-121 with CBAM, weighted cross-entropy loss, and CutMix achieves 96.90% overall accuracy and 0.96 macro F1 on an 80/20 test split, and they claim this is better than prior approaches on the same dataset.

Significance. If the reported results are reliable, the proposed lightweight model could be practically valuable for IoT-based onion disease classification, avoiding the annotation overhead of detection-based pipelines. The systematic comparison of augmentation and loss strategies for class imbalance is also useful to the community. However, the central empirical claim rests entirely on a single split with no variance estimates, and the selection of the best configuration from the same split introduces optimism bias. The comparison with prior work is informal, and reproducibility is hampered by missing hyperparameters and no code/data release. These issues limit the current significance, though they are addressable with additional experiments and reporting.

major comments (4)

[Section IV, Table II] The headline 96.90% accuracy is the best of seven configurations reported on a single 80/20 split. Section IV states that five-fold cross-validation was used for hyperparameter tuning and robustness, but Table II gives no fold-wise statistics, no mean±std, and no information about how validation was used for model selection. If the same test split was used to compare configurations and then report the best one, the reported number is an optimistically biased maximum over seven models. This is especially concerning for minority classes: Basal Rot has ~140 total images, so ~28 test images; its reported 0.87 accuracy corresponds to about 24 correct predictions. A different split could move macro-F1 by several points. Please report results across folds or repeated runs, provide seeds, and clarify the model-selection procedure.
[Section III.B, III.D, IV] Reproducibility is insufficient to support the central empirical claim. The paper does not report learning rate, batch size, number of epochs, optimizer, CBAM insertion locations, CutMix parameters, or focal loss alpha/gamma. No code or data release is indicated. Without these details, the exact experiment cannot be independently replicated. Please provide a complete training protocol and, ideally, release code and split indices.
[Section V and Abstract] The abstract claims the model 'gives better results than other approaches using the same datasets,' but the only quantitative comparison in the text is informal against Raj et al. [3], which is the authors' own prior work, uses a different task (five classes, detection-based), and is not evaluated under the same test protocol. No head-to-head comparison on the same split, with the same metrics, is provided. Please either add a direct quantitative comparison or temper the claim to avoid unsupported superiority.
[Section IV, Tables I and II] The class merge of Anthracnose and Twister changes the classification task from 9 to 8 classes, but Table I reports 9-class results while Table II reports 8-class results. This makes the headline accuracy not directly comparable with the earlier 9-class experiments. Since the merge is justified by an external domain source, it is acceptable, but the paper should report the final configuration's 9-class accuracy as well, or clearly state that the 8-class task is the primary one and discuss the impact of the merge on the reported improvements.

minor comments (5)

[Section I] In the paragraph discussing Raj et al., the sentence 'Also, they focus on only five image classes. In image augmentation.' is incomplete and should be rewritten.
[Section IV] Typo: 'DeseNet-121' should be 'DenseNet-121' in the bullet list of models.
[Section III.B] Typo: 'guassian' should be 'Gaussian'; 'Sammple' in Figure 2 caption should be 'Sample'.
[Section IV and VI] Typos: 'Pipleline' should be 'Pipeline', 'Blocal Attention Module' should be 'Block Attention Module', and 'it's practical applicability' should be 'its practical applicability'.
[References] Reference [8] has typos: 'Internation Conference on Computer Visison' should be 'International Conference on Computer Vision'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the 96.90% accuracy is a held-out test measurement, and no equation reduces the result to its inputs; the only self-citation (Raj et al.) is a non-load-bearing baseline comparison.

full rationale

The central claim is an empirical evaluation on a held-out 20% test split (Section IV), not a quantity derived from the model's definitions. The class weights in Eq. (4) are inverse-frequency weights used in the weighted cross-entropy loss; the reported overall accuracy (Eq. 7) and macro F1 (Eq. 10) are computed from test predictions, so they are not equivalent by construction to any fitted target. The architecture combines externally published components (DenseNet-121 [10], CBAM [12], CutMix [8]) with standard loss functions; none of these is redefined in terms of the reported accuracy. The only self-citation, Raj et al. [3] (coauthor Rajbabu Velmurugan), is used as a comparison baseline in Sections II and V, but it does not supply a premise from which the 96.90% figure is derived, so it is not load-bearing under the rules. The main validity concerns—a single 80/20 split without confidence intervals, and the possibility that the best of seven Table II configurations was selected on the same test split—are statistical-evidence limitations rather than circular reasoning; they concern generalizability, not an equation reducing the result to its inputs. No circular step can be quoted from the paper.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central result depends on the assumption that merging Anthracnose and Twister is scientifically valid, on the transferability of ImageNet features, and on the representativeness of the single dataset. There are no invented entities.

free parameters (2)

Training hyperparameters (learning rate, batch size, epochs, optimizer)
Not reported in the paper; the authors say tuning was done but do not list values.
Focal loss hyperparameters (alpha, gamma)
Used in one configuration; the values are not given, and the best model uses weighted cross-entropy instead.

assumptions (3)

domain assumption Anthracnose and Twister are the same disease and can be merged into one class
Section IV cites a Pathogens review (reference [14]) to justify merging the classes, which changes the task from nine to eight classes and materially affects the accuracy figures.
domain assumption ImageNet pre-trained features transfer to onion leaf images
The paper uses ImageNet-initialized DenseNet-121 and fine-tunes; no domain-specific pre-training or feature analysis is provided.
domain assumption The test split is representative of real-world field images
The dataset is a single collection from ICAR, and the paper does not describe collection protocol or variety coverage.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Handling imbalance and few-sample size in ML based Onion disease classification." pith.science (2026). https://pith.science/paper/QBBNTVRT

@misc{pith2026250905341,
  author       = {Pith},
  title        = {Pith review of: Handling imbalance and few-sample size in ML based Onion disease classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBBNTVRT}},
  note         = {Machine review of arXiv:2509.05341}
}

read the original abstract

Accurate classification of pests and diseases plays a vital role in precision agriculture, enabling efficient identification, targeted interventions, and preventing their further spread. However, current methods primarily focus on binary classification, which limits their practical applications, especially in scenarios where accurately identifying the specific type of disease or pest is essential. We propose a robust deep learning based model for multi-class classification of onion crop diseases and pests. We enhance a pre-trained Convolutional Neural Network (CNN) model by integrating attention based modules and employing comprehensive data augmentation pipeline to mitigate class imbalance. We propose a model which gives 96.90% overall accuracy and 0.96 F1 score on real-world field image dataset. This model gives better results than other approaches using the same datasets.

Figures

Figures reproduced from arXiv: 2509.05341 by the authors.

**Figure 2.** Sammple Images from Each Class B. Data Pre-Processing and Augmentation We employed different data pre-processing pipelines tailored for each model setup: 1) Pipeline A: Normal Augmentation: This pipeline includes resizing images to 224×224, normalization, and basic augmentations such as horizontal flip and rotation. 2) Pipeline B : Imbalanced Sampler: As the classes in the dataset are highly imbalanced, an imbalan… view at source ↗

**Figure 5.** Architecture of Model with CBAM D. Loss Functions and Training Procedure 1) Weighted Cross-Entropy Loss: We used weighted-cross entropy loss (WCE) in order to mitigate the effects of weight imbalance in the dataset. The loss is as given in Equation 3. LWCE = − 1 N X N i=1 wyi log(pyi ) (3) where: • N is the total number of samples in the batch. • yi is the true class label for sample i. • pyi is the predicted probab… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Architecture of Basic Model 2) Models with CBAM: DenseNet-121 model was combined with Convolutional Block Attention Module (CBAM) [12]. The CBAM module uses attention mechanism and enhances the feature representation by focusing on the informative parts of the featu…

**Figure 6.** Figure 6: Dataset Distribution after combining Anthracnose and Twister classes [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

**Figure 8.** Figure 8: Confusion Matrix on test data by Densenet-121 with CBAM, WCE, [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 7.** Figure 7: Confusion Matrix on test data by Densenet-121 with CBAM, WCE, [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages

[3]

YOLO-ODD: An improved YOLOv8s model for onion foliar disease detection,

A. Raj, M. Dawale, S. Wayal, K. Khandagale, I. Bhangare, S. Banerjee, A. Gajarushi, R. Velmurugan, M. S. Baghini, and S. Gawande, “YOLO-ODD: An improved YOLOv8s model for onion foliar disease detection,” in Frontiers in Plant Science ,

work page
[1]

FAO- STAT: Crops and livestock products – Onion production by coun- try,

Food and Agriculture Organization of the United Nations (FAO), “FAO- STAT: Crops and livestock products – Onion production by coun- try,” 2025. [Online]. Available: https://www.fao.org/faostat/en/#rankings/ commodities by country. [Accessed: May 30, 2025]

work page 2025
[2]

Detection of Onion Leaf Disease Using Hybridized Feature Extraction and Feature Selection Approach,

G. O. Gbadebo, J. K. Alhassan, and O. A. Ojerinde, “Detection of Onion Leaf Disease Using Hybridized Feature Extraction and Feature Selection Approach,” in 5th Information Technology for Education and Development (ITED), 2022 , doi: 10.1109/ITED56637.2022.10051500

work page arXiv 2022
[4]

Comparative analysis of different plant leaf disease classification and detection using CNN,

N. Lokhande, V . Thool, and P. Vikhe, “Comparative analysis of different plant leaf disease classification and detection using CNN,” in Proc. 2024 International Conference on Recent Innova- tion in Smart and Sustainable Technology (ICRISST) , 2024, doi: 10.1109/ICRISST59181.2024.10921975

work page arXiv 2024
[5]

Identification of Maize Leaf Diseases Using Improved Deep Convolutional Neu- ral Networks,

X. Zhang, Y . Qiao, F. Meng, C. Fan, and M. Zhang, “Identification of Maize Leaf Diseases Using Improved Deep Convolutional Neu- ral Networks,” in IEEE Access , vol. 6, pp. 30370–30377, 2018 doi: 10.1109/ACCESS.2018.2844405

work page arXiv 2018
[6]

Imbalanced Dataset Sampler,

M. Yang, J. Borovec, J. Bang, D. H. Cho, W. U. Baig, Z. Ni- trome, T. Machnitzki, and L. Tae-hoon, “Imbalanced Dataset Sampler,” GitHub Repository, 2021. [Online]. Available: https://github.com/ufoym/ imbalanced-dataset-sampler. [Accessed: May 30, 2025]

work page 2021
[7]

Albumentations: Fast and Flexible Image Augmentations,

A. Buslaev and A. Parinov, E. Khvedchenya, V . I. Iglovikov, A. A. Kalinin, “Albumentations: Fast and Flexible Image Augmentations,” arXiv:1809.06839v1, 2018. [Online]. Available: https://arxiv.org/pdf/ 1809.06839. [Accessed: May 30, 2025]

arXiv 2018
[8]

CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,

S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,” in IEEE/CVF Internation Conference on Computer Visison (ICCV), 2019, pp. 6022–6031

work page 2019

Show all 15 references

[9]

Deep Residual Learning for Image Recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016 , pp. 770–778, doi: 10.1109/CVPR.2016.90

2016 doi
[10]

Densely Connected Convolutional Networks,

G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely Connected Convolutional Networks,” in Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), 2017 , pp. 2261–2269, doi: 10.1109/CVPR.2017.243

2017 doi
[11]

ImageNet: A Large-Scale Hierarchical Image Database,

J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei, “ImageNet: A Large-Scale Hierarchical Image Database,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009 , pp. 248–255, doi: 10.1109/CVPR.2009.5206848

2009
[12]

CBAM: Convolu- tional Block Attention Module,

S. Woo, J. Park, J.Y . Lee, and I. S. Kweon, “CBAM: Convolu- tional Block Attention Module,” in arXiv, 2018, [Online]. Available: https://arxiv.org/pdf/1807.06521

2018 arXiv
[13]

Focal Loss for Dense Object Detection,

T. Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal Loss for Dense Object Detection,” in arXiv, 2018, [Online]. Available: https://arxiv.org/pdf/1708.02002

2018 arXiv
[14]

Anthracnose of Onion (Allium cepa L.): A Twister Disease,

R. Dutta, K. Jayalakshmi, S. M. Nadig, D. C. Manjunathagowda, V . S. Gurav, and M. Singh, “Anthracnose of Onion (Allium cepa L.): A Twister Disease,” Pathogens, 2022 Aug 5;11(8):884. doi: 10.3390/pathogens11080884. PMID: 36015005; PMCID: PMC9415854. 6

2022 doi
[2025]

[Online]. Available: [https://www.frontiersin.org/journals/plant- science/articles/10.3389/fpls.2025.1551794/full] 5 Metric D121 + WCE + A D121 + CBAM + WCE + A D121 + CBAM + Focal + A D121 + WCE + C D121 + WCE + D D121 + CBAM + WCE + C+D D121 + CBAM + WCE + D Overall Accuracy...

2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [3]

YOLO-ODD: An improved YOLOv8s model for onion foliar disease detection,

A. Raj, M. Dawale, S. Wayal, K. Khandagale, I. Bhangare, S. Banerjee, A. Gajarushi, R. Velmurugan, M. S. Baghini, and S. Gawande, “YOLO-ODD: An improved YOLOv8s model for onion foliar disease detection,” in Frontiers in Plant Science ,

work page

[2] [1]

FAO- STAT: Crops and livestock products – Onion production by coun- try,

Food and Agriculture Organization of the United Nations (FAO), “FAO- STAT: Crops and livestock products – Onion production by coun- try,” 2025. [Online]. Available: https://www.fao.org/faostat/en/#rankings/ commodities by country. [Accessed: May 30, 2025]

work page 2025

[3] [2]

Detection of Onion Leaf Disease Using Hybridized Feature Extraction and Feature Selection Approach,

G. O. Gbadebo, J. K. Alhassan, and O. A. Ojerinde, “Detection of Onion Leaf Disease Using Hybridized Feature Extraction and Feature Selection Approach,” in 5th Information Technology for Education and Development (ITED), 2022 , doi: 10.1109/ITED56637.2022.10051500

work page arXiv 2022

[4] [4]

Comparative analysis of different plant leaf disease classification and detection using CNN,

N. Lokhande, V . Thool, and P. Vikhe, “Comparative analysis of different plant leaf disease classification and detection using CNN,” in Proc. 2024 International Conference on Recent Innova- tion in Smart and Sustainable Technology (ICRISST) , 2024, doi: 10.1109/ICRISST59181.2024.10921975

work page arXiv 2024

[5] [5]

Identification of Maize Leaf Diseases Using Improved Deep Convolutional Neu- ral Networks,

X. Zhang, Y . Qiao, F. Meng, C. Fan, and M. Zhang, “Identification of Maize Leaf Diseases Using Improved Deep Convolutional Neu- ral Networks,” in IEEE Access , vol. 6, pp. 30370–30377, 2018 doi: 10.1109/ACCESS.2018.2844405

work page arXiv 2018

[6] [6]

Imbalanced Dataset Sampler,

M. Yang, J. Borovec, J. Bang, D. H. Cho, W. U. Baig, Z. Ni- trome, T. Machnitzki, and L. Tae-hoon, “Imbalanced Dataset Sampler,” GitHub Repository, 2021. [Online]. Available: https://github.com/ufoym/ imbalanced-dataset-sampler. [Accessed: May 30, 2025]

work page 2021

[7] [7]

Albumentations: Fast and Flexible Image Augmentations,

A. Buslaev and A. Parinov, E. Khvedchenya, V . I. Iglovikov, A. A. Kalinin, “Albumentations: Fast and Flexible Image Augmentations,” arXiv:1809.06839v1, 2018. [Online]. Available: https://arxiv.org/pdf/ 1809.06839. [Accessed: May 30, 2025]

arXiv 2018

[8] [8]

CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,

S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,” in IEEE/CVF Internation Conference on Computer Visison (ICCV), 2019, pp. 6022–6031

work page 2019