Semantic Segmentation of Cucumber Leaf Disease Spots Based on ECA-SegFormer

Round 1

Reviewer 1 Report

 

The paper presents a novel method, ECA-SegFormer, for semantic segmentation of cucumber leaf disease spots under natural acquisition conditions. The proposed approach shows improved performance compared to the original SegFormer and appears to be more suitable for segmenting cucumber leaf disease spots in real-world environments.

Here are a few remarks to help improve the article.

·        You must check grammar and sentence structure.

 

·        The paper stated that ECA-SegFormer performs better than SegFormer, it would be valuable to discuss the potential implications of this improvement. How could the increased accuracy benefit cucumber leaf disease evaluation and treatment? Highlighting the practical implications of the findings will strengthen the paper.

·        Authors should add more details on the dataset: acquisition, out or in door, the process of annotation, …

·        The choice of only “four cucumber leaf diseases” is not justified

·        Data augmentation should not be a part of the new dataset preparation and presentation.

·        Authors need to test and evaluate the proposed improvement on other public datasets.

·        References should have the same format. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

 

The article presents a novel approach, the ECA-SegFormer model, which combines the SegFormer model with the Efficient Channel Attention (ECA) module, offering a unique solution for cucumber leaf disease spot segmentation in natural environments. The construction of a dedicated image dataset for evaluating the proposed model demonstrates a meticulous approach to addressing the specific challenges and complexities of cucumber leaf disease spot segmentation, enhancing the reliability and relevance of the research. The article highlights the practical implications of the research by acknowledging the potential application of the ECA-SegFormer model as an effective tool for evaluating and treating cucumber leaf damage. Furthermore, the mention of future research possibilities for real-time disease diagnosis in natural environments indicates the article's contribution to advancing the field and its potential impact.

Some potential weaknesses or limitations:

·  Limited Dataset: The article mentions the construction of a new image dataset of diseased cucumber leaves for evaluation. However, the size and diversity of the dataset are not discussed, which could potentially limit the generalizability of the results. A small or biased dataset may not fully represent the variations and complexities present in real-world scenarios.

·       Lack of Comparison with State-of-the-Art: The article focuses on comparing different modules and configurations within the SegFormer model but does not mention any direct comparison with state-of-the-art approaches or existing benchmarks for cucumber leaf disease spot segmentation. Without such comparisons, it is challenging to determine the model's performance relative to other methods in the field.

·   Evaluation on Limited Metrics: The article evaluates the proposed model using metrics commonly used in semantic image segmentation. While Intersection over Union (IoU), mean Intersection over Union (mIoU), Pixel Accuracy (PA), and Mean Pixel Accuracy (MPA) provide useful insights, they might not capture all aspects of segmentation performance, such as edge accuracy or instance-level segmentation quality. A more comprehensive set of evaluation metrics could provide a more robust assessment.

·   Lack of Detailed Hyperparameter Analysis: The article briefly mentions the hyperparameters used but does not provide an in-depth analysis of their impact on the model's performance. A more thorough exploration of hyperparameter tuning and sensitivity analysis could offer a better understanding of the model's robustness and generalizability.

·    Limited Discussion on Computational Efficiency: The article primarily focuses on the segmentation performance of the proposed model but lacks discussion on its computational efficiency. Considering the potential deployment of such models in real-world applications, it is crucial to assess their inference speed and resource requirements, especially when dealing with large-scale datasets or real-time scenarios.

· Lack of External Validation: The article does not mention any external validation or validation on independent datasets. The performance of the proposed model should ideally be tested on unseen data or compared with results obtained from different sources to demonstrate its generalization capability and reliability.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear authors,

 

Thank you for your responses to our comments and for your efforts to improve the quality of your document. 

However, some points are not properly addressed.

 

In point 2 of your reply, it is known that in some cases of image classification, 'it is not always necessary to segment images to improve classification accuracy. Therefore, when I asked you this question, I meant that you should demonstrate experimentally the importance of segmentation in your case. Results with and without segmentation. 

 

But you only cite two references that don't deal with turmeric leaves.

 

In point 3, you should add more information about the expert who helped you in the annotation process.

 

In point 4, if the main objective of your article was "transfer learning" architectures, we can accept that. But the article is about detecting turmeric disease. So you must be working on all the most important major diseases.

 

In point 5, I mean that augmentation can be used in experimentation: training and/or validation and cannot be part of the initial data set.

 

In point 6, when I talked about other public datasets, I was talking about the whole process, not the model test.

 

Here are a few other comments:

 

In the segmentation comparison you added in the recent version, have you tested all these approaches on your dataset? You say in the text that

 

"... we compare it with representative segmentation models on the test set".

 

Does this mean that you haven't trained these architectures on your training dataset? So you've been using pre-trained architectures without fine-tuning?

 

In reference 2, use this

 

Martinelli, F., Scalenghe, R., Davino, S. et al. Advanced methods of plant disease detection. A review. Agron. Sustain. Dev. 35, 1-25 (2015). https://doi.org/10.1007/s13593-014-0246-1

Author Response

Please see the attachment.

Author Response File: Author Response.docx