Review Reports

Round 1

Reviewer 1 Report

In this manuscript, the authors proposed a method for plant leaf recognition based on several pre-trained CNN networks and feature recalibration. The proposed method has achieved excellent performance in identification and retrieval experiments. In my opinion, the contributions are enough, and the structure of the paper is reasonable and the writing is clear. 

I have some comments about this paper: 

(1) In Figure 2, are the feature maps extracted from a CNN network recalibrated together or separately?

(2) For VGG16 and VGG19, why the layers {9, 16, 23, 30}, {9, 18, 27, 36} are selected?

(3) Some middle-level leaf shape descriptors should be added in the Related Works section, for examples: 

Wang X, et al. Bag of contour fragments for robust shape classification. Pattern Recognition, 2014.

Zeng S, et al. Joint distances by sparse representation and locality-constrained dictionary learning for robust leaf recognition. 2017.

(4) The implementation process should be described in details, like software, computer configuration, etc. 

(5) How efficient and fast is the proposed method in training and testing stages?

According to the above, I think this paper can be accepted for publication after minor revision. 

Author Response

We have uploaded an attachment about response to the comments of reviewer1

Author Response File: Author Response.pdf

Reviewer 2 Report

The article give a new method to extract image features with application to plant leaf recognition. Impressive recognition accuracies and retrieval MAP scores have been obtained. The idea in the manuscript is novel and effective for plant recognition, the article is well organized and written.  

(a) What is the differences between the datasets ICL and ICL compound ?

(b) In Figure 4, the % is missing

(c) In Figure 6, why the result of configuration index 8 is lower than that of configuration index 7, even the former's feature length is longer. 

(d) There are several clerical errors, such as the sentence "our learned plant features is very effective" in section 4.9.

(e) It is better to give a summary table for the used eight datasets, for examples: categories, training/testing image number, etc. 

(f) The Turkey Plant is a complex dataset, it is better to give a confusion matrix to show the accuracy for each species.

Author Response

We have uploaded an attachment about response to the comments of reviewer2

Author Response File: Author Response.pdf

Reviewer 3 Report

In this manuscript, the authors introduce a feature extraction pipeline for plant recognition. The authors conducted extensive experiments for evaluation. However, there are some suggestions:

 

1. It would be better if the authors could clarify that the proposed method is used for plant recognition based on the traits like leaves since there are also research works on plant recognition based on the traits like roots or flowers.

 

2. Following comment 1, it would be better if the authors could discuss the existing datasets for plant recognition as well in the related works section, like [a][b][c]. Also, transfer learning works are recommended and worth to be discussed.

a. Kritsis, Kosmas, et al. "GRASP-125: A Dataset for Greek Vascular Plant Recognition in Natural Environment." Sustainability 13.21 (2021): 11865.

b. Xu, Weihuang, et al. "PRMI: A dataset of minirhizotron images for diverse plant root study." arXiv preprint arXiv:2201.08002 (2022).

c. Liu, Xinda, et al. "Plant disease recognition: A large-scale benchmark dataset and a visual region and loss reweighting approach." IEEE Transactions on Image Processing 30 (2021): 2003-2015.

 

3. It would be better if the authors could compare with more recent works in the experiment sections.

Author Response

We have uploaded an attachment about response to the comments of reviewer3

Author Response File: Author Response.pdf

Reviewer 4 Report

This paper presented a novel deep learning-based method for plant recognition, where the convolutional neural network was employed for the task of interest. In the proposed method, multiple CNNs were used for feature extraction of plan leaf, and linear SVM was used for recognition. The performance of the proposed method was validated using 8 open datasets, with satisfactory results. Overall, the topic of this research is interesting, and the manuscript was well organised and written. The detailed comments are provided as follows.

1.       The main innovation and contribution of this research should be well organised in abstract and introduction.

2.       Please broaden and update literature review on CNN or deep learning for practical applications. E.g. Torsional capacity evaluation of RC beams using an improved bird swarm algorithm optimised 2D convolutional neural network; Automated damage diagnosis of concrete jack arch beam using optimized deep stacked autoencoders and multi-sensor fusion.

3.       The effectiveness of both CNN and SVM is heavily dependent on the setting of hyperparameters. How did the authors optimise them in this research to achieve the best recognition accuracy?

4.       When the pretrained CNNs were used, transfer learning technology is included. Please give more details on how the authors transferred these pretrained models to the ones used in this research?

5.       Please give the features extracted from CNNs.

6.       Confusion matrix is suggested to be used for performance evaluation.

7.       A comparison between SVM and other ML classifiers is suggested to be included.

 

8.       More future research should be included in conclusion part.

Author Response

We have uploaded an attachment about response to the comments of reviewer4

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

This is a manuscript after revision. I am grateful for the authors' responses which have addressed all my concerns.