A Lightweight Real-Time Tomato Leaf Disease Detection System for Edge-Based Smart Agriculture

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

To address the problem of imbalance between detection quality, model complexity, and deployment feasibility, this study proposes HGS-YOLO, a system-oriented deployable lightweight adaptation of YOLOv11 for leaf-level tomato disease detection. Beyond model architecture, the study establishes a complete engineering pipeline that includes training, optimization, post-training quantization, and hardware deployment with BPU acceleration on a D-Robotics RDK X5 handheld platform. Experimental results demonstrate that HGS-YOLO offers a favorable accuracy-efficiency-deployability trade-off and provides a feasible technical route for precision agriculture, rapid on-site disease screening, and low-cost intelligent plant health monitoring. And the introduction provides sufficient background, the references are appropriate, the research design is appropriate, the description of the methods can be improved, the description of the results is appropriate, some figures can be improved, the conclusions are supported by the results. The specific problems to be solved are as follows.

(1) Line 161-164: It is recommended to supplement the description of the detection accuracy of the model, such as MAP50, Precision, Recall and F1.

(2) Line 267 and 320: Formula 7 is inconsistent with Formula 2, there is no F₅ in Formula 2, and whether Up(P₄) should be corrected to Up(F₄), please check. It is suggested to further optimize the description of the feature fusion process in combination with Figure 3.

(3) Line 330: The description of “CIoU loss function” should be corrected to “IoU”. Please check.

(4) The subscripts of some variables are not set properly. Such as “Bt” in line 332, “Wg and Hg ” in line 337, “Wgt and Hgt” in line 344.

(5) Line 340: The description of “IoU loss function” should be corrected to “L_IoU”. Please check.

(6) Line 364: Please check whether the description of “Equation (12)” should be corrected to “Equation (15)”.

(7) Line 387-389: It is recommended to explain the basis for setting the values of the confidence threshold and IoU threshold.

(8) Line 498: In Figure 8, it is necessary to clearly define which axis represents the true label and the predicted label respectively.

(9) Line 518-522: This paragraph should be placed at the end of Section 3.2.1. Please check.

(10) The text in Figure 9-11 needs to be enlarged for clear display.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The abstract reports that PTQ makes mAP50 decrease by 93.6 to 93.0. This limitation of accuracy degradation assertion is not sufficiently contextualized - a loss of 0.6 percentage points on a fairly easy, domain-restricted dataset is possibly more important in actual field situations. Do not frame it as a categorically limited drop without field validation evidence to back that framing.

The introduction based its literature review on a shallow review which is dominated by the YOLO family papers. The authors cite foundation-model-assisted UAV forest monitoring (refs 20-21) and drug synergy scoring (refs 18-19), which they do not properly justify as being relevant to this work - this is padding their citations. More importantly, no discussion of the existing lightweight edge-deployment systems about plant disease detection is discussed except the YOLOv8n and YOLOX papers mentioned. Competitors which apply MobileNet-SSD or EfficientDet-Lite, or transformer-lite versions on RPi/Jetson-class platforms are not reviewed. The authors ought to considerably broaden the landscape of related-work and place HGS-YOLO more accurately in it.

Line 31-32: the statistic global tomato production was more than 186 million tonnes is referred to as FAOSTAT 2024 (ref 1), although the reference record on page 24 does not take the correct form of an MDPI record and looks unfinished. Please check and re-format all references to Sensors style.

The dataset is stated as being a modified version of PlantVillage which has 4,323 images and 15,135 annotations, although the process of modification is not specified. The article mentions that the low-quality images were eliminated, but there is no mark and standard of what would be considered low-quality images. This is a major reproducibility issue. The authors should include: (a) the total number of images eliminated and the reasons, (b) formal description of the script-generated collage protocol (e.g., the number of source images per collage, strategy of overlap), and (c) the inter-annotator agreement measure (e.g. kappa of Cohen) of the two or three annotators. The definition of annotation quality control as having been determined by a senior researcher is not enough to be scientifically reproducible.

The authors admit that the samples of collage are produced using plain- background leaves and that the separation at the source-image level was imposed. Nevertheless, as the proportion of synthetically constructed part of the dataset is approximately 31 percent of the entire dataset, there is a potential for a realistic overlap of the distribution. Just by not including the source images in validation/test sets, it does not imply that the synthetic collages do not resemble validation images in terms of background, texture or lighting statistics. It should also have a formal analysis (e.g. t-SNE or SSIM similarity check between splits) which should show that the train/val/test splits are actually not in feature space discontinuous.

HGS-YOLO is a new contribution of the paper, yet all three parts, HGNetV2 (Baidu PaddlePaddle), HS-FPN (Yang et al., 2024), and MPDIoU (Ou and Shen, 2024) are direct borrowings of the existing literature with no in-house adjustments. Their combination and implementation on a particular platform is the sole originality. This is valid systems contribution that should be framed in a more cautious way. The paper should have a clear section or table under what is new to clearly define what is borrowed and what is designed by the authors. This should not be what the reviewers and the readers have to puzzle together with footnotes and parenthetical citations.

Training hyperparameters are underspecified. The batch size and number of epochs are reported as 8, 300, and the optimizer is AdamW with = 0.001, however, they do not specify: the learning rate scheduler (cosine decay or step decay or both), the weight decay coefficient, data augmentation (mosaic, mixup, flips, color jitter), and early stopping criteria (e.g., any).. All these have a great impact on reproducibility and have to be disclosed in their entirety.

The CIoU and MPDIoU equations are obtained in section 2.2.3 at length, a visual comparison plot of the gradient landscape of CIoU versus MPDIoU in degenerate cases would have made the mathematical exposition of the same equations much easier. The authors state that CIoU degenerates into standard IoU when the aspect ratio and center of boxes are equal, which is the main driving fact behind the use of MPDIoU, which needs concrete example, not formulae.

Equation (17) formulates the total loss having balancing weights _cls, _box, and _dfl. What are these weights of HGS-YOLO training? Did YOLOv11n defaults or tuning inherit them? If tuned, how?

All the performance comparisons between models (Table 1, Tables 3-5) are given as point estimates without statistical significance testing. In detection benchmarking, a difference of 0.2-0.5 mAP50 percent variations is considered normal run-to-run variance. The authors should report at least: (a) the results obtained in several runs (n 3) with the means and standard deviation, (b) a significance test (e.g., a paired t-test or a Wilcoxon signed-rank test on the per-image AP) between HGS-YOLO and the YOLOv11n baseline. In the absence of this, claims such as the mAP50 of HGS-YOLO is 93.6% are not scientifically defensible due to comparisons.

The deployment test (Table 8) is done with the condition of controlled indoor light only with individually gathered tomato leaf samples. Although it is better than pure benchmark evaluation, it does not warrant the sale of a system that is being sold as precision agriculture and on-site disease screening. Real agricultural fields present problems such as: direct sunlight, which results in overexposure and specular highlights, motion blur due to wind, some of the stems and surrounding leaves, which partially cover the target, variable distances to the target, and the soil, insects, and other background objects. None of these conditions is sufficiently simulated by the brightness-sensitivity analysis based on a synthetic linear scaling factor (Section 3.2.2-3.2.3). The authors must, at least, perform and report outdoor field validation, even on a lesser scale, or renounce it as a straight-out laboratorium-validated prototype and push field validation to explicitly identified upcoming work with a definable roadmap.

Table 1 compares HGS-YOLO to YOLOv5n, YOLOv8n, YOLOv11n and 2 ablation versions. This is a comparative set that is narrow. It is not comparable to: (1) recent agricultural-specific lightweight detectors (e.g. the YOLOv8n-based approach by Liu et al. 2024, referenced as [48], which was explicitly optimised to run on the same RDK X5 platform), (2) any transformer-lite or MobileViT-based baseline, or (3) any model actually deployed to the same R Inferring the speed on a GPU workstation (Table 1, which does not specify FPS on the edge device) and only reporting edge performance on HGS-YOLO is methodologically unfair. All the competing models are supposed to be profiled on the common hardware.

Per-class AP50 is welcome in Table 2. Nevertheless, the discussion only reports the best ( Yellow Leaf Curl Virus, 97.0) and the worst ( Spider Mites, 90.7) performing classes, without stating the reasons. The most visually difficult class is Spider Mites which leave tiny scattered leaf surface scars. At this level, a research article should include a short qualitative or quantitative study of why this class is performing poorly and what could be done about that.

Figure 8 shows a normalized confusion matrix, but the textual discussion (lines 494-498) is dismissive: "the model performed well... there were a few misclassifications." It needs to be analyzed in a systematic manner. What are the most commonly mixed pairs of classes? Is the off-diagonal mass concentrated in disease pairs with similar visual symptoms (e.g. Early Blight vs. Late Blight) or does it occur by chance? The critical analysis is necessary in the interpretation of the system clinical safety because the appearance of the Late Blight as Early Blight may result in completely incorrect medication choices.

The paper indicates 9.8 +/-0.4 W average power of the system deployed with INT8 at 40 FPS. As it is being reported, it is never mentioned in relation to the DJI TB48s battery capacity or anticipated continuous working range. A farmer would be quite entitled to enquire: how long does this device take a single charge? It would greatly enhance the practical usefulness of this work to Sensors readers in giving an estimate of how long it can operate (even as a rough estimate).

Table 9 reports end-to-end FPS of 40.0 +/- 0.6 for INT8 PTQ but pure inference FPS of 53.2 +/- 0.9. The difference suggests that it loses approximately 13 FPS to preprocessing, NMS, and rendering. What is the distribution of this pipeline budget? The decomposition of latency elements (capture -> preprocess -> inference -> NMS -> render) would be valuable to the systems practitioner.

Recommendations for Revision

1. Perform at least a restricted field test with actual agricultural lighting and background and provide quantitative data. Even 50-100 field pictures with meticulous annotation would support the claims of the paper greatly. In the event that this cannot be possible, retitle and reword the paper as a clear laboratory-scale prototype validation and delete all of the language suggesting field readiness.
2. Report all benchmark comparisons as mean +/- std across 3 independent training runs and report a statistical significance test between HGS-YOLO and the YOLOv11n baseline.
3. Infer profile and report speed, memory and power consumption of all competing models (Table 1) not only on a GPU workstation, but on the same RDK X5 hardware. The poor proxy of deployment latency on BPU architectures is hardware-agnostic FLOPs.
4. Please submit the complete training configuration file (YAML or similar) as an additional resource to make it reproducible. This is as common in the YOLO ecosystem as it is in Sensors reviewers.
5. Extend the discussion on the confusion matrix to find the top pairs of classes in the confusion matrix and comment on their clinical/agronomic importance.
6. The battery endurance is added as an estimate of the power draw measured.
7. Provide the inter-annotator agreement measures of the dataset annotation process.
8. paraphrase the conclusion part to include original synthesis instead of paraphrasing the abstract.
9. Edit or delete the references to [18] and [19] (drug synergy and software defect prediction) - their applicability to this paper is not proven, and they appear as citation padding.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. The introduction includes some citations from broader AI domains such as biomedical synergy scoring and software defect prediction. These references may be technically valid, but they feel peripheral to the paper’s central topic. The introduction would be stronger if it focused more tightly on plant disease detection, lightweight detectors, agricultural edge AI, and embedded deployment studies.

2. The gap statement is useful, but the manuscript should better distinguish between: algorithmic novelty, which is limited here, and
practical deployment contribution, which is the real strength.Currently, the article occasionally sounds more novel than the evidence supports.

3. The authors should more explicitly explain why YOLOv11n was chosen as the base model, especially since the final gains over the baseline are mostly in compactness and deployment rather than accuracy. That would strengthen the rationale for the architecture selection.

4. Although the dataset is described as modified for object detection, it is still dominated by plain-background leaves and synthetic collages. Only about 23% of the images represent more complex greenhouse or field-like backgrounds. This makes it difficult to infer how well the detector will perform in real agricultural scenes.

5. The manuscript states that a separately collected deployment-test set was annotated and used for controlled indoor evaluation, but I did not see a clear sample count for that set in the extracted text. That needs to be stated explicitly in the manuscript. Without it, readers cannot judge the strength of the deployment validation.

6. Synthetic collages can help create multi-object scenes, but they are not a substitute for real field complexity. The authors should clarify how much of the final performance may depend on collage-based augmentation and whether models were separately evaluated on plain-background, realistic-background, and collage subsets.

7. The paper reports single benchmark values for mAP50, recall, parameters, FLOPs, and model size. For an incremental study, it would be much stronger to report results across multiple training runs or at least provide variability for the benchmark metrics, not only for deployment FPS and latency.

8. Artificial brightness scaling from 0.2 to 1.8 is not equivalent to real illumination change in agricultural settings. It simulates exposure variation, but not shadows, specular highlights, sensor noise, motion blur, or partial occlusion. The authors should tone down claims of real-world robustness and clarify that this is only a controlled synthetic robustness probe.

9. The methods would benefit from clearer implementation details for reproducibility. The manuscript gives training hardware, optimizer, batch size, epochs, and image size, but it would still help to specify: augmentation settings, confidence/NMS settings used during validation versus deployment, whether pretrained weights were used, calibration-set size and selection for PTQ, whether layer fusion or graph optimization was applied before deployment. Some of this is implied, but not all of it is sufficiently explicit.

10. The mathematical presentation is more descriptive than necessary in places. Some equations summarize standard operations of the modules and PTQ mapping, but the manuscript should make clear which equations are: standard background, exact implemented formulas, or conceptualized summaries of the module behavior. This is especially important because the paper is not introducing a fundamentally new loss or new backbone.

11. The performance gains are practical, not scientific breakthroughs. HGS-YOLO does not outperform the YOLOv11n baseline in accuracy; instead, it preserves most of the performance while reducing complexity. That is fine, but the discussion should consistently reflect that trade-off rather than imply a net accuracy advance.

12. The deployment test remains too controlled. The indoor deployment results are useful, but controlled indoor lighting with separately collected leaf samples is still far from open-field or greenhouse operation under natural variability. Therefore, the statement that the system is ready for practical smart agriculture should be softened unless supported by more realistic tests.

13. The confusion-matrix discussion is too general. The paper notes that most categories were detected well and misclassifications were limited, but the discussion should dig deeper into which classes are confused and why. Since Spider Mites had the lowest AP50 and Yellow Leaf Curl Virus the highest, a class-level error analysis would make the results much more informative.

14. No comparison is given against stronger recent edge-deployment baselines on the same hardware. The manuscript compares with classic YOLO variants and cites prior Jetson Nano studies, but hardware comparisons across devices are not directly comparable. A stronger result section would clarify that the advantage is shown within the authors’ own platform and workflow, not necessarily across all edge devices.

15. The manuscript needs clearer separation between benchmark and deployment claims. Benchmark mAP50 of 93.6% on a modified PlantVillage-based set should not be blended too closely with practical deployment claims, because the data domains differ and the deployment validation is smaller and more controlled.

16. The manuscript should consistently use the same naming format for variants such as HSFPN versus HS-FPN to avoid confusion.

17. Please clarify whether the equation block for PTQ is the actual implementation used by the toolchain or a simplified explanatory formulation.

18. Consider adding a comparison between benchmark subset types: plain-background single leaves, realistic-background individual plant images, collage images. This would directly show where the model is actually strong and where it remains weak.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Thanks for updating according to the comments

Author Response

Thank you for your time and the update.

Reviewer 3 Report

Comments and Suggestions for Authors

The reviewer revised the manuscript. No more comments for the current version.

Comments on the Quality of English Language

Fine

Author Response

Thank you for your time and the update.