Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes an approach based on YOLOv12n for tomato leaf disease detection and assessment. Here are some considerations.

1. The authors should highlight recent relevant works in the same field, such as https://doi.org/10.1016/j.atech.2025.101324, in the discussion on the state of the art.
2. Please enhance the resolution of Figures 2 and 3.
3. Please underline that SPDConv and MPDIoU were not introduced by the authors.
4. Please reduce the verbosity of Section 2 by highlighting relevant concepts and removing unnecessary phrasing.
5. From Table 1, the model appears to achieve 600+ FPS. Is this correct? If so, why didn't the authors consider the use of small or medium models, at least? It is also important to evaluate newer models (e.g., YOLO26) that are specifically targeted at edge devices.
6. Please report the results at the second decimal place. Furthermore, it is important to note that mAP should also be evaluated at 0.5:0.95, depending on resolution and target sizes.
7. From Table 3, C3RFEM appears to provide better results in terms of precision and recall, but lower mAP 0.5. Please provide further evaluation of this result, as it is counterintuitive.
8. To provide proper visualization, consider evaluating activation maps.
9. Overall, all the provided images are barely readable and provide low quality and resolution. Please improve them for the final version of the manuscript.

Overall, the study is interesting, but there are some inconsistencies in the evaluation that require further explanation and experiments to be properly framed.

Author Response

We sincerely appreciate the reviewer’s recognition of our work and the constructive suggestions. We have carefully addressed all comments as follows:

Comment 1: The authors should highlight recent relevant works in the same field, such as https://doi.org/10.1016/j.atech.2025.101324, in the discussion on the state of the art.

Response 1: We sincerely thank the reviewer for this valuable suggestion. We agree that incorporating recent relevant studies would strengthen the discussion of the state of the art. As suggested, we have cited the recommended study and added a concise description in the Introduction. The added text highlights the use of incremental learning and domain adaptation in YOLOv11-based tomato phenotyping to improve detection stability and efficiency, thereby providing more up-to-date context for our study.

Comment 2: Please enhance the resolution of Figures 2 and 3.

Response 2: We sincerely thank the reviewer for this helpful suggestion. The high-resolution source files for all figures have been collected in a compressed file and uploaded together with the revised submission for editorial and production use.

Comment 3: Please underline that SPDConv and MPDIoU were not introduced by the authors.

Response 3: We sincerely thank the reviewer for this helpful comment. We agree that the original wording could have caused ambiguity regarding SPDConv and MPDIoU. Accordingly, we revised the relevant descriptions in Sections 2.2, 2.3, and 2.5 to explicitly identify SPDConv and MPDIoU as existing methods adopted in this study. Specifically, the term “existing” was added to their descriptions, and “novel” was removed from the description of SPDConv. These revisions clarify that SPDConv and MPDIoU were applied within DFR-YOLOv12n rather than proposed in the present study, while preserving the original technical meaning.

Comment 4: Please reduce the verbosity of Section 2 by highlighting relevant concepts and removing unnecessary phrasing.

Response 4: We sincerely thank the reviewer for this helpful suggestion. We carefully revised Section 2 to improve its conciseness and focus. Specifically, the descriptions of dataset construction and preprocessing in Section 2.1 were streamlined, and the overall architecture description in Section 2.2 was substantially condensed. In Sections 2.3–2.5, repetitive background descriptions and redundant performance-oriented statements were removed or revised, while the remaining text focuses on the core mechanism, implementation position, and intended role of each module. In addition, the standalone Section 2.6 was removed and only one concluding sentence was left. These revisions reduce unnecessary repetition and make the methodological description more direct and easier to follow.

Comment 5: From Table 1, the model appears to achieve 600+ FPS. Is this correct? If so, why didn't the authors consider the use of small or medium models, at least? It is also important to evaluate newer models (e.g., YOLO26) that are specifically targeted at edge devices.

Response 5: We sincerely thank the reviewer for raising this important concern. We confirm that the FPS values exceeding 600 in Table 1 are the actual results measured on a Tencent Cloud server equipped with an NVIDIA GeForce RTX 3080 Ti. The reported FPS represents the model throughput obtained during batch validation at an input resolution of 640 × 640 and a batch size of 16.

We also agree that comparisons with small-, medium-, and newer edge-oriented models are necessary. Accordingly, Table 5 has been expanded to include YOLOv12s, YOLOv12m, and YOLO26n. YOLOv12s and YOLOv12m achieved mAP@0.5:0.95 values of 71.7% and 72.4%, respectively, but required 9.28 M and 20.20 M parameters and 21.70 G and 68.10 G FLOPs. In comparison, DFR-YOLOv12n achieved the highest mAP@0.5 of 88.6% with only 2.25 M parameters, 5.27 G FLOPs, and a model size of 4.71 MB. Compared with the edge-oriented YOLO26n, DFR-YOLOv12n improved mAP@0.5 and mAP@0.5:0.95 by 2.2 and 0.8 percentage points, respectively, while maintaining slightly lower computational and storage requirements. These additional comparisons and the corresponding discussion have been incorporated into Section 4.5. They further clarify that the nano-scale architecture was selected to balance detection accuracy and deployment-oriented model complexity.

Comment 6: Please report the results at the second decimal place. Furthermore, it is important to note that mAP should also be evaluated at 0.5:0.95, depending on resolution and target sizes.

Response 6: We thank the reviewer for this valuable suggestion. We agree that mAP@0.5:0.95 provides a more stringent assessment across multiple IoU thresholds and is particularly informative for evaluating targets with different sizes and localization requirements. Accordingly, we have added the definition and calculation of mAP@0.5:0.95 to the evaluation metrics section. We have also added this metric to Tables 1–5 and revised the corresponding analyses in Sections 4.1–4.5.

Regarding decimal precision, the original evaluation logs report Precision, Recall, mAP@0.5, and mAP@0.5:0.95 to three decimal places on the normalized 0–1 scale. After conversion to percentages, this corresponds to one decimal place, for example, 0.705 corresponds to 70.5%. We therefore retained one decimal place for the percentage-based accuracy metrics to remain consistent with the precision of the original outputs and to avoid implying unsupported numerical precision by appending an additional zero. Nevertheless, we have standardized the numerical formatting throughout the revised tables, and FLOPs, parameter counts, model sizes, and FPS values are reported to two decimal places where supported by the original measurements.

Comment 7: From Table 3, C3RFEM appears to provide better results in terms of precision and recall, but lower mAP 0.5. Please provide further evaluation of this result, as it is counterintuitive.

Response 7: We thank the reviewer for identifying this apparently counterintuitive result. We re-examined the results in Table 3 and clarified their interpretation in Section 4.3. C3RFEM achieved the highest Precision of 89.1%, but its Recall was 76.2%, which was lower than both Model 1 at 77.5% and the A2C2f_DEConv configuration at 79.1%. Therefore, C3RFEM did not simultaneously improve both Precision and Recall.

More importantly, the reported Precision and Recall correspond to a particular operating point determined by the confidence threshold, whereas AP summarizes the precision–recall curve across confidence thresholds and is subsequently averaged across disease categories. Consequently, a higher Precision at one operating point does not necessarily produce a higher mAP when Recall or performance consistency across confidence thresholds and classes is lower. This explains why C3RFEM achieved a Precision of 89.1% but an mAP@0.5 of only 85.7%.

To provide a more comprehensive evaluation, we have added mAP@0.5:0.95 to Table 3 and revised the accompanying analysis. Compared with C3RFEM, A2C2f_DEConv achieved higher Recall and mAP@0.5 values of 79.1% and 87.0%, respectively. It also reduced FLOPs from 5.48 G to 5.27 G and model size from 4.77 MB to 4.71 MB, while increasing FPS from 438.46 to 647.21. Therefore, the selection of A2C2f_DEConv was based on its combined performance in Recall, mAP, computational complexity, model size, and inference speed rather than on Precision alone.

Comment 8: To provide proper visualization, consider evaluating activation maps.

Response 8: We sincerely thank the reviewer for this constructive suggestion. We agree that activation-map analysis could provide additional post hoc interpretability regarding the image regions contributing to the model predictions. After careful consideration, however, we did not add this analysis in the present revision. The primary purpose of Section 4.6 is to evaluate the observable detection behavior of the proposed model rather than to infer its internal causal mechanisms. Figure 6 already provides matched visual comparisons across representative natural scenes involving small lesions, densely distributed targets, complex backgrounds, and visually ambiguous lesion boundaries. False detections and missed detections are also explicitly marked in the figure. These qualitative observations are further supported by the quantitative ablation and comparative results presented in Tables 1–5. We have revised the corresponding discussion to avoid attributing the observed improvements directly to unverified internal attention mechanisms and now restrict our interpretation to the evaluated detection outcomes. We therefore consider the current evidence adequate for the scope of the claims made in the revised manuscript. Nevertheless, if the editor considers activation-map analysis essential for the assessment of this manuscript, we would be pleased to provide Grad-CAM visualizations in a further revision.

Comment 9: Overall, all the provided images are barely readable and provide low quality and resolution. Please improve them for the final version of the manuscript.

Response 9: We sincerely thank the reviewer for this important comment. We have improved the readability and resolution of all figures in the revised manuscript. To ensure that the figures can be assessed and used at publication quality, the high-resolution source files for all figures have also been provided in a compressed file uploaded with the revised submission.

Reviewer 2 Report

Comments and Suggestions for Authors

1.A repeated shortcoming I usually and always observed in such works is always around dataset split ratios and distribution. You have almost 13k images which was built from multiple public sources and online images then manually screened and re-annotated into 8 classes with an 8:1:1 split. However, you didnt provide the source wise image counts, class imbalance, bounding-box annotation rules, lesion level vs leaf level labeling criteria, inter-annotator agreement, or near-duplicate removal procedure. Since public tomato disease datasets often contain visually similar images, random splitting always cause source leakage and overestimation in performance. So plz add source wise and class wise dataset statistics, representative annotation examples for ambiguous cases. Mention annotation quality control and a source disjoint or near-duplicate-filtered test split with per-class AP, confusion matrix and error analysis for visually similar diseases like early blight, late blight, bacterial spot, and Septoria. This is impt. before claiming robustness in such natural environments.

2.Analysing ur results, I observed that ablation shows only modest gains over YOLOv12n with mAP@0.5 improving from 86.5% to 88.6%, recall from 76.9% to 81.7% and model size from 5.22 MB to 4.71 MB. However the regression loss comparison is not convincing enough as DIoU gives the highest mAP@0.5:0.95 of 71.2% while the proposed MPDIoU model reports 70.7% and yet the text still emphasizes superior localization. So plz justify why MPDIoU is preferred if stricter localization accuracy was lower. Another thing is that ur edge experiment reports Jetson Nano deployment using ONNX and TensorRT and a GUI but it only shows a single real time example with 5.77 FPS and 0.105s inference. Why wasnt there any baseline Jetson comparisons, memory use and such measurements. Kindly report end-to-end latency for YOLOv12n and DFRYOLOv12n on the same Jetson Nano settings while including FP32 or FP16 or maybe INT8 mode with input resolution, batch size, GPU memory and associated metrics. How can we judge ur lightweight and edge deployment conclusions without these??

3.Recent state of the art works around agricultural object detection studies have shown that YOLO family models perform well in controlled conditions but may vary substantially under real field variability, where occlusion, illumination changes, spectral differences, and false-detection behaviour strongly affect deployment reliability. One such strong work by Rana et al., 2026 validated this fact that although the crop target might differ, a multi-domain YOLOv8–YOLOv12 benchmark showed that field reliability depended not only on mAP but also on robustness to spectral variability, occlusion, lighting heterogeneity and FP - FN behaviour. Source: https://doi.org/10.1016/j.rineng.2025.108833. This would help you better contextualise the present DFRYOLOv12n work focused on real-world deployment and model reliability.

Author Response

We appreciate the reviewer’s recognition of our manuscript and the constructive suggestions. We have carefully addressed all comments as follows:

Comment 1: A repeated shortcoming I usually and always observed in such works is always around dataset split ratios and distribution. You have almost 13k images which was built from multiple public sources and online images then manually screened and re-annotated into 8 classes with an 8:1:1 split. However, you didnt provide the source wise image counts, class imbalance, bounding-box annotation rules, lesion level vs leaf level labeling criteria, inter-annotator agreement, or near-duplicate removal procedure. Since public tomato disease datasets often contain visually similar images, random splitting always cause source leakage and overestimation in performance. So plz add source wise and class wise dataset statistics, representative annotation examples for ambiguous cases. Mention annotation quality control and a source disjoint or near-duplicate-filtered test split with per-class AP, confusion matrix and error analysis for visually similar diseases like early blight, late blight, bacterial spot, and Septoria. This is impt. before claiming robustness in such natural environments.

Response 1: 

We sincerely thank the reviewer for this comprehensive and constructive comment. In response, we revised Sections 2.1.1 and 2.1.2 to improve the transparency of dataset construction and preprocessing. We now explicitly identify the four public data sources and clarify the image-screening, category-harmonization, and duplicate- and near-duplicate-removal procedures. The final curated dataset contains 12,755 images across eight categories, and representative images are provided in Figure 1.We have also clarified that the complete visible leaf was defined as the bounding-box target. In addition, the annotations were checked for missing or incorrect labels and for bounding boxes extending beyond image boundaries. The dataset was partitioned using stratified random sampling at an 8:1:1 ratio, yielding 10,204 training images, 1,276 validation images, and 1,275 test images. Online data augmentation was applied exclusively to the training set.We respectfully note that the dataset was constructed as a unified curated dataset after cross-source screening, category harmonization, re-annotation, and duplicate removal, rather than as a comparative evaluation of the individual original sources. Therefore, we report the characteristics of the final curated dataset rather than separate source-wise image counts. Duplicate and near-duplicate samples were removed before dataset partitioning to reduce the potential risk of information leakage.

Regarding the reviewer’s final suggestion, we agree that per-class AP, confusion matrices, and detailed error analysis would provide an additional class-level perspective on the model performance. After careful consideration, we did not add these analyses in the present revision because the primary objective of this study is to evaluate overall detection performance and lightweight deployment characteristics, rather than to establish a fine-grained differential diagnostic framework for visually similar diseases. The revised manuscript reports Precision, Recall, mAP@0.5, and mAP@0.5:0.95 on the same test set, together with ablation experiments and comparisons with mainstream detection models. In addition, Figure 6 presents representative results for Early Blight, Late Blight, Bacterial Spot, and Septoria, with false detections and missed detections explicitly marked. To avoid overstatement, we have restricted our robustness claims to the aggregate test-set results and the selected qualitative examples under the evaluated conditions. We no longer interpret these results as conclusive evidence of class-specific robustness. We therefore consider the current evidence adequate for the revised scope of the manuscript. Nevertheless, if the editor considers class-level evaluation essential, we would be pleased to add per-class AP, a normalized confusion matrix, and a targeted error analysis in a further revision.

 

Comment 2: Analysing ur results, I observed that ablation shows only modest gains over YOLOv12n with mAP@0.5 improving from 86.5% to 88.6%, recall from 76.9% to 81.7% and model size from 5.22 MB to 4.71 MB. However the regression loss comparison is not convincing enough as DIoU gives the highest mAP@0.5:0.95 of 71.2% while the proposed MPDIoU model reports 70.7% and yet the text still emphasizes superior localization. So plz justify why MPDIoU is preferred if stricter localization accuracy was lower. Another thing is that ur edge experiment reports Jetson Nano deployment using ONNX and TensorRT and a GUI but it only shows a single real time example with 5.77 FPS and 0.105s inference. Why wasnt there any baseline Jetson comparisons, memory use and such measurements. Kindly report end-to-end latency for YOLOv12n and DFRYOLOv12n on the same Jetson Nano settings while including FP32 or FP16 or maybe INT8 mode with input resolution, batch size, GPU memory and associated metrics. How can we judge ur lightweight and edge deployment conclusions without these??

Response 2: First,we sincerely thank the reviewer for identifying this important inconsistency. We agree that the original manuscript did not adequately explain why MPDIoU was retained despite the higher mAP@0.5:0.95 achieved by DIoU, and that our previous description overstated the localization advantage of MPDIoU. Our selection was based on the specific objective and annotation strategy of this study. The complete visible leaf was used as the detection target, and the primary objective was reliable disease-leaf detection and classification rather than precise lesion-boundary delineation. Accordingly, mAP@0.5 was used as the primary criterion for selecting the bounding-box regression loss, while mAP@0.5:0.95 was used as a complementary metric under stricter IoU thresholds. MPDIoU achieved the highest mAP@0.5 of 88.6%, exceeding DIoU by 0.6 percentage points, whereas DIoU achieved a higher mAP@0.5:0.95 of 71.2% compared with 70.7% for MPDIoU. We therefore retained MPDIoU because it performed better under the primary evaluation criterion of this study.

Second, we added a quantitative edge-deployment benchmark and a new comparison table (Table 6). The deployment environment is now described in Section 3.2. Both YOLOv12n and DFR-YOLOv12n were converted to FP16 TensorRT engines and evaluated on the same Jetson Nano B01 platform using an input size of 640 × 640 and a batch size of 1. Each model was warmed up for 30 runs and evaluated over 300 runs using identical preprocessing and post-processing procedures.

As shown in Table 6, DFR-YOLOv12n reduced the mean end-to-end latency from 139.061 ms to 124.599 ms and the P95 end-to-end latency from 152.849 ms to 130.132 ms. The end-to-end throughput increased from 7.191 FPS to 8.054 FPS, while peak shared memory usage decreased from 3165 MB to 3122 MB. In addition, compared with YOLOv12n, DFR-YOLOv12n reduced FLOPs, parameter count, and model size by 9.6%, 10.4%, and 9.8%, respectively. Here, end-to-end latency includes image preprocessing, data transfer, TensorRT inference, and CPU-based non-maximum suppression, but excludes camera acquisition and GUI rendering.

 

 

Comment 3: Recent state of the art works around agricultural object detection studies have shown that YOLO family models perform well in controlled conditions but may vary substantially under real field variability, where occlusion, illumination changes, spectral differences, and false-detection behaviour strongly affect deployment reliability. One such strong work by Rana et al., 2026 validated this fact that although the crop target might differ, a multi-domain YOLOv8–YOLOv12 benchmark showed that field reliability depended not only on mAP but also on robustness to spectral variability, occlusion, lighting heterogeneity and FP - FN behaviour. Source: https://doi.org/10.1016/j.rineng.2025.108833. This would help you better contextualise the present DFRYOLOv12n work focused on real-world deployment and model reliability.

Response 3: We sincerely thank the reviewer for this valuable suggestion. As recommended, we have cited the study by Rana et al. and added a concise discussion in the Introduction. The revised text emphasizes that field reliability depends not only on mAP but also on robustness to occlusion, illumination variation, and false detections. This addition better contextualizes DFR-YOLOv12n in terms of real-world deployment and model reliability.

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript (horticulturae-4418962) demonstrates that DFR‑YOLOv12n is a lightweight tomato leaf disease detection model that enhances fine‑grained lesion recognition and boundary clarity while reducing model size for edge deployment. It improves accuracy compared with YOLOv12n and operates efficiently on the Jetson Nano, making it suitable for real‑world agricultural applications.

The manuscript is interesting and well written. The sections with introduction, materials and methods, results, and discussion are appropriate and aligned with the study’s aims. I offer a few minor suggestions to the authors: the figure captions could be made more self‑explanatory and accessible to readers. They should be expanded to clearly state the total sample size and the proportions of images used for training and validation.

I also suggest that the conclusions should not be presented as bullet points. Even if the section has an introductory nature, the final considerations and conclusions should be written as a cohesive summary that integrates the main findings and highlights the manuscript’s key contributions.

Regarding the abstract, I recommend reducing the word count, while maintaining the essential content to meet the journal’s editorial standards.

Author Response

We sincerely appreciate the reviewer’s recognition of our work and the constructive suggestions. We have carefully addressed all comments as follows:

Comment 1: They should be expanded to clearly state the total sample size and the proportions of images used for training and validation.

Response 1: Thank you for this helpful suggestion. To ensure that the dataset information is clear and readily accessible, we have explicitly stated in Sections 2.1.1 and 2.1.2 that the dataset contains 12,755 images and was divided into training, validation, and test sets at an 8:1:1 ratio, corresponding to 10,204, 1,276, and 1,275 images, respectively. Because these data describe the overall dataset construction and partitioning procedure rather than the content of a specific figure, they are presented in the relevant dataset subsections instead of being repeated in the caption of Figure 1. This arrangement keeps the figure caption concise while providing the complete sampling information in its appropriate methodological context.

 

Comment 2: The conclusions should not be presented as bullet points. Even if the section has an introductory nature, the final considerations and conclusions should be written as a cohesive summary that integrates the main findings and highlights the manuscript’s key contributions.

Response 2: Thank you for this constructive suggestion. We agree that connected prose provides a clearer and more cohesive synthesis of the study. Accordingly, we removed the numbered points and rewrote Section 6 as a continuous narrative that integrates the model improvements, quantitative results, edge-deployment findings, study implications, and future directions.

 

Comment 3: Regarding the abstract, I recommend reducing the word count, while maintaining the essential content to meet the journal’s editorial standards.

Response 3: Thank you for this helpful suggestion. In response, we carefully condensed the Abstract from 313 to approximately 277 words by removing redundant descriptions while retaining the research background, main methodological improvements, key quantitative results, and edge-deployment conclusions. The revised Abstract is more concise while preserving the essential content and principal findings of the study.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors successfully fixed all the issues highlighted during the previous round of reviews. Therefore, the paper can be considered for publication.

Reviewer 2 Report

Comments and Suggestions for Authors

Your revision has improved substantially particularly through the clearer definition of the leaf-level detection task and also the more balanced interpretation of MPDIoU, and the addition of a fair Jetson Nano benchmark against the YOLOv12n baseline. The new latency, throughput, and memory results strengthen the practical deployment claims, while the moderated discussion and added limitations improve the overall scientific transparency of the study. It can now be accepted.