Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This article aims to address the problem of the long-tail distribution of disease categories in plant disease detection and the problem of excessively large model parameters, which makes edge deployment difficult.

The advantages of the article:

The topic is of practical significance. The paper has a clear structure and detailed descriptions of materials and methods. The proposed CALM-Aug strategy enhances the data and achieves the goal of improving the detection performance. Figure 1 shows the CALM-Aug framework pipeline; the diagram is well designed with clear colors, and the text boxes are arranged neatly, which can clearly show the CALM-Aug data-processing pipeline. Table 1 lists performance metrics for models before and after the use of the CALM-Aug framework. Both mAP50 and mAP50-95 have shown significant improvements, demonstrating the effectiveness of the CALM framework on the PlantDoc dataset, and the CALM-Aug framework has been open-sourced. Figure 2 clearly shows the improved YOLO11-ARL model, and detailed explanations are provided for the three improvements. Finally, the model compression using LAFDP is introduced, and the performance before and after compression is compared in detail in Table 4, showing the good performance of the compressed model. In the results and discussion, the authors compare YOLO11-ARL-PD with 14 different models, showing that the overall performance of the model is strong. YOLO11-ARL-PD was also evaluated on three datasets outside the PlantDoc dataset and achieved better results than the original YOLOv11n.

 

Weaknesses of the article:

Does the introduction provide sufficient background and include all relevant references?

The introduction covers most of the background and relevant references, and the introduction is relatively rich. But the discussion of existing Copy-Paste and model-compression methods is insufficient, and there is not enough comparison of the differences between similar Copy-Paste and model-compression methods.

 

Is the research design appropriate?

The overall research idea is rigorous and logical, and it contains a lot of comparison work, including 8:1:1 dataset division and maintaining consistent training conditions during training, which demonstrates the scientific rigor and reasonableness of the experiment. However, there are still some shortcomings in the performance verification of the CALM-Aug framework, which only compares the results with those of the non-augmented method, and does not compare with the experimental results of similar augmentation methods. Now that YOLO26 has been released, we hope that the authors can add comparative experiments and fill in Table 5.

 

Are the methods adequately described?

Section 2.4 describes the model compression using LAFDP in detail, and the LAFDP compression process and underlying idea are introduced in detail; it also gives the validity of LAFDP compression compared with the uncompressed model. However, the paper lacks the comparison with similar pruning and distillation methods, which makes it difficult to better highlight the contribution and innovation of LAFDP. In the first two paragraphs of section 2.4.1, the article describes the LAFDP framework as having a channel importance assessment mechanism at its core, noting that channel importance is determined by three types of information, but the explanation that follows is not clear enough. First, it is unclear how the three indicators mentioned in the first sentence of paragraph 2 are combined, and whether there is a weighting factor. Then there is the second sentence of the second paragraph, which refers to the pruning of each layer at a set ratio, and does not sufficiently tell the reader exactly how the ratio is determined, whether manually or automatically. A clearer description is needed to make it reproducible.

 

Are the results clearly presented?

The results section clearly demonstrates the effectiveness of the CALM-Aug framework proposed in the article on the PlantDoc dataset, fully shows that the LAFDP compression method can reduce the number of parameters while improving model performance, and finally demonstrates that YOLO11-ARL-PD performs effectively on three datasets outside the PlantDoc. However, the parameter count of 2242402 in line 8 of table 4 on page 13 does not match the parameter count of 2.251M before pruning in table 4 on page 14. It may be necessary to explain why the parameter counts in table 4 of page 13 do not match the parameter count of 2.251M and which model before pruning is compared here.

 

Are the conclusions supported by the results?

The existing results can basically support the conclusion that the YOLO11-ARL-PD model can achieve both lightweight deployment and performance improvement.

 

Are all figures and tables clear and well-presented?

The figures and tables are rich and well designed, which can clearly show the data processing pipeline of the CALM-Aug framework, the overall architecture of the YOLO11-ARL-PD model and the architecture of the EMSCP module and other modules. However, there were some editing errors. Both Pages 13 and 14 are annotated as Table 4. The formula on page 19 and the formula on page 4 are both labeled as (1). The eight pictures in the two text boxes in the lower right corner of Figure 1 are not clear enough.

 

Quality of English Language

The quality of English expression was generally good, but in the second paragraph of page 15 “……changes. with a few……” appears to use punctuation incorrectly.

 

Below is a list of other suggestions for improvement in the article. Table 6 of the article was tested on three other leaf disease datasets, but did not cite the sources of these datasets. It is recommended that the best values in the table be bolded to make it easier for the reader to read.

Although there are some areas that need to be improved, the overall logical rigor, detailed experimentation, and complete structure can give readers some inspiration.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

In this paper, the authors focused on a highly relevant topic: monitoring plant diseases in open-field settings. They addressed a pressing issue in modern automated plant disease diagnostic systems: long-tail (rare diseases) and the difficulty of recognizing early stages of tissue damage in plants.
The depth of the presented model and the significant attention paid to its optimization for use on edge devices are also worth noting. The authors successfully solved the complex problem of compressing the model without losing quality on specific data.

The paper was very impressive, but there are some recommendations for improvement:

1. In the "3 Results and Discussion" section, in addition to the results obtained, attention should be given to the discussion. In addition to citing specific results, describe the physical significance, for example, why this hypothesis works, and cite the work of other teams of authors whose results confirm the obtained results.
2. Perhaps it would be worthwhile to examine in more detail the results when recognition accuracy is low, for example, for "Tomato leaf yellow virus," and describe the reasons and methods for improving accuracy.
3. Describe the direction of future work, for example, field testing of the model on existing agricultural UAVs, since the model's capabilities to process data in the presence of external noise, such as vibration and image blur, are not entirely clear.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf