MSFCA-Net: A Multi-Scale Feature Convolutional Attention Network for Segmenting Crops and Weeds in the Field

Round 1

Reviewer 1 Report

agriculture-2405515-peer-review-v1

The authors present an interesting study of using CNN for feature extraction. The work is interesting and well-planned and implemented.  Few comments need to be considered before accepting the manuscript.

Introduction

The authors presented the background and justification of the problem in an acceptable way.

Materials and Methods

Do you think the number of images used in training is enough to build models that can be used with independent data sets?

 What do you think about transfer learning to other crops?

Figure 1: place labels on the soybean and weed (which colour is for which?)

Figures 2-4: same like Figure 1

What are the main advantages of MSFCANet over other feature extraction networks?

The authors have explained the MSFCANet in an adequate way.

Results

The authors sufficiently addressed the results.

Discussion

The authors still need to compare the results obtained in the study to those from other algorithms in the literature whether traditional or deep types.

Conclusions

The authors need to put some numbers that clarifies the advantages of their algorithm.

The English language is ok and no major errors were found.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Authors contributions:

The proposed MSFCA-Net in the research paper makes several contributions to the field of weed segmentation in agriculture.

The authors have the following contributions:

The paper introduces MSFCA-Net, which is a novel architecture specifically designed for crop and weed segmentation. It utilizes asymmetric large convolutional kernels and an attention mechanism to aggregate multi-scale features, enabling the model to capture both local information and global contextual information. This approach significantly improves the segmentation accuracy and the model's ability to handle details and edge segmentation.

The paper proposes a hybrid loss calculation mode that combines dice loss and focal loss. This hybrid loss function addresses the challenges of class imbalance and enhances the model's ability to learn from difficult samples. By designing separate loss functions for crops and weeds, the proposed model can effectively handle the segmentation of different classes.

Comparative evaluations with popular semantic segmentation methods on soybean, sugar beet, carrot, and rice weed datasets demonstrate that the proposed MSFCA-Net achieves higher mean intersection over union (MIoU) scores. The results indicate that the model outperforms other methods under the same experimental conditions and parameter configurations. It exhibits strong robustness and generalization ability, making it suitable for crop and weed segmentation in complex backgrounds.

The researchers captured and manually annotated a dataset of soybean seedlings and weeds in the field. This dataset enrichment contributes to the availability of agricultural weed data, providing researchers with a rich and effective resource for future investigations in the field of intelligent weed control and smart agriculture.

 

There are also limitations to consider:

The research focuses on soybean, sugar beet, carrot, and rice weed datasets. While these datasets cover important crops, the generalization of the proposed method to other crop types and weed species remains to be tested. The model's performance on datasets outside the scope of this study is unknown.

The paper mentions limited and difficult-to-obtain agricultural weed datasets as a challenge. Although the researchers manually annotated a dataset of soybean seedlings and weeds, the size and diversity of the dataset may still be limited. It is important to ensure the dataset's representativeness and scalability for broader applicability.

The proposed method may still face challenges related to complex changes in field lighting intensity, mutual occlusion between crops and weeds, and uneven size and quantity of crops and weeds. While the research addresses these issues to some extent, real-world field conditions can be highly variable and may introduce additional complexities not fully accounted for in the study.

The paper focuses on the development of a segmentation model, but it does not discuss the real-time implementation aspects of intelligent point-to-point spraying or weeding. Practical considerations, such as computational efficiency, hardware requirements, and integration with existing agricultural machinery, are important factors for the adoption and scalability of such methods.

 

I have some reviewer notes:

Abstract. How your work can be implemented in practice?

Introduction. Line 86. Write sentence with other words. The aim of this work is not clearly defined.

Line 114. “Anhui Agricultural University” (City, country).

How many plants are studied? Also, for 2.1.2. Sugar beet dataset. You have to describe it.

Line 120. Does the resolution 1024x768 is enough? You have to describe why do you use this resolution.

Line 134. How 70% training, 20% validation, and 10% test sets are selected? You have to describe it. Same for “Carrot dataset” and data processing of all datasets.

Equation (5). Does “Totalloss” has measurement unit. If it is dimensionless, you have to describe it.

Line 290. “Intel Core i9-13600KF”. It has to be: Model (Manufacturer, City, country of origin). Also for NVIDIA GeForce RTX 3090 GPU.

Line 292. “Windows 11, CUDA 11.3, Python 3.9, and TensorFlow 2.6. ” It has to be: Software, Version (Manufacturer, City, country of origin).

Table 1. A comparison of the proposed method with other state-of-the-art methods based on soybean

dataset. The criteria are in “%”. 0.9 means 0.9%. You have to correct it. Same for other tables.

Discussion. You have to compare your results with minimum 3 other papers. 6 or more will be better.

Conclusion. At the end of this part, you have to describe the limitations of your work. How the work will be continued?

 

I have some suggestions:

Improve the presentation of your results. Make better description of hardware and software that you use.

Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

First of all, I want to congratulate the authors for their efforts in this manuscript. They presented a mechanism for differentiating soybean, sugar beet, carrot, and rice from weed plants based on a convolutional-attention network. The topic is aligned with the journal's scope, and the paper is well-structured. In general terms, the paper has a high quality. There are some aspects which have to be improved before accepting the paper. Following, I include some comments aimed at enhancing the quality of the paper:

1-      In the introduction, the authors have mentioned other methods using ML to determine the presence of weed plants. Nonetheless, some proposals use basic image processing, such as vegetation indexes or edge detection with no ML methods (10.1016/j.compag.2020.105684 and 10.3390/agriengineering2020012). Authors should cite some examples of these techniques to contextualize the existing solutions for weed detection.

2-      Moreover, in the introduction, the paper's aim should be defined in a clearer and more straightforward way. I suggest using a single paragraph for this purpose. Consider using bullet points to highlight the novelty or main contributions.

3-      Authors have to justify why the equipment used for dataset generation is different for soybean.

4-      Regarding the carrot, rice, and sugar beet, no information about the distance from the camera to the soil. This information must be provided.

5-      In the case of the rice, the position of the camera seems to be different. While in all the datasets, the images were gathered with the camera in a zenith position, the case of the rice is different. Authors have to justify these differences and their possible impact on the performance of the proposal. What differences do the shadows and leaves overlap compared with the zenith position?

6-      The discussion must be extended. It is suggested to include different subsections. First, the authors should compare the results among the different datasets and assess the differences based on the model, the crops, and the image capture mechanism. In a subsequent subsection, authors have to compare their results with existing similar proposals and justify their results based on current literature to highlight which contributions are supported in the literature and which are new or break with existing knowledge. Finally, the main limitations of the datasets and the proposal should be discussed. A limitation I have detected is the extremely low distance between the camera and the soil. Will the system perform well when images are gathered with UAV, which will provoke a reduction of spatial resolution of the pictures? Or is the proposal limited to their use with terrestrial vehicles?

 

7-      Finally, IN the conclusions, the future work should be provided in a new paragraph. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The authors have corrected the mentioned issues and the paper is now ready to be accepted.