Research on Identification Technology of Field Pests with Protective Color Characteristics
, Zhenhuan Long, Anwen Liu, Xiufeng Dai, Xiangming Lei and Zhenhui TangRound 1
Reviewer 1 Report
This article uses the particularity of infrared imaging to distinguish the presentation of cabbage and pest on the image and then uses yolov5 for training. Finally, it obtained a very high rate of pest detection. However, since the author has lost the description of critical experimental data, please submit it again after supplementing.
1. First of all, since this article lacks the content description of the data set used in the experiment (it seems that the author has deleted it), I urge the author to supplement it in detail before submitting it.
2. A dataset has three train/validation/test groups, and the number of images and instances should be specially annotated.
3. Figure 5, upper (a) to (d), has no description of them.
4. Figure 5. Crop and past method is a newer trend of data augmentation, but the author seems to oversimplify the spirit of crop and past. The instance should be crop and paste instead of multiple whole images. But still a good trick.
5. Figure 7. The author is requested to supplement the drawing of spectral difference values because it is not easy to get the correct difference value in Figure 7.
6. Figure 8. The classification graph doesn't seem to have any values. mAP? The author should note the number of pest categories. If there is only one pest category, mAP should be equivalent to AP.
7. Figure 9. The content has been changed to 99.3%, but Figure 9 is still not corrected?
Author Response
Dear Reviewer:
Thank you very much for your comments concerning our manuscript entitled ‘Research on identification technology of field pests with protective color characteristics’ (ID: applsci-1577984). Those comments (at the bottom of the cover letter) are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval.
The main corrections in the manuscript and the responds to the comments are as following:
1) First of all, since this article lacks the content description of the data set used in the experiment (it seems that the author has deleted it), I urge the author to supplement it in detail before submitting it.
Response: We thank you very much for your comments. We have added relevant contents to the original manuscript. And we also did relevant experiments. The supplementary contents are as follows:
“The original image data set was expanded through data enhancement to enhance the diversity of the data set, avoid overfitting, and boost the generalization ability and robustness of the identification algorithm [35-37]. Common data enhancement methods include rotation, flip, clipping, adding noise, jitter, blur, translation, and staggered transformation [38-40]. In this paper, the original image data set was expanded from 500 images to 1500 images through rotation, flip, translation, and changing brightness considering the factors such as the influence of camera angle and light intensity (including the lighting conditions simulating sunny or cloudy days and exposure or insufficient light) on the identification algorithm (Fig. 5).
Figure 5. Data enhanced pest image. (a) Image rotation, (b) Image flip, (c) Image trans-lation, (d) Change the brightness of the image ”
2) A dataset has three train/validation/test groups, and the number of images and instances should be specially annotated.
Response: We thank you very much for your comments. The pest image data set is divided into training set, verification set and test set according to the proportion of 6:2:2. The pest images of training set, verification set and test set are 900, 300 and 300 respectively. We have added relevant contents to the original manuscript. The supplementary contents are as follows:
“Then, the pest images and annotation files are divided into training set, verification set and test set according to the proportion of 6:2:2 respectively. The training set was employed to fit the detection network. The validation set was adopted to adjust the super parameters of the detection network and preliminarily evaluate the network performance. The test set is used to evaluate the generalization ability of the final model.”
3) Figure 5, upper (a) to (d), has no description of them.
Response: We thank you very much for your comments. We have added relevant contents to the original manuscript. The supplementary contents are as follows:
“Figure 5. Data enhanced pest image. (a) Image rotation, (b) Image flip, (c) Image trans-lation, (d) Change the brightness of the image”
4) Figure 5. Crop and past method is a newer trend of data augmentation, but the author seems to oversimplify the spirit of crop and past. The instance should be crop and paste instead of multiple whole images. But still a good trick.
Response: We thank you very much for your comments. Data enhancement mentioned in Figure 5 in the original manuscript is a data enhancement method called mosaic data enhancement that comes with the yolov5 algorithm. The principle of this data enhancement method is introduced in the original manuscript. Since we have adopted common data enhancement methods to expand the original data set, we have deleted the introduction of mosaic data enhancement methods. Delete the following:
“YOLOv5 carries out mosaic data enhancement while loading image data, that is, randomly select several images in the training set and form new training elements through cutting and splicing, which is used to alleviate the lack of elements in the training set, so as to enhance the recognition ability of the model. The data en-hance-ment process in Yolo network is shown in Figure 5.
Figure 5. Data enhancement”
5) Figure 7. The author is requested to supplement the drawing of spectral difference values because it is not easy to get the correct difference value in Figure 7.
Response: We thank you very much for your comments. We have added relevant contents to the original manuscript. The supplementary contents are as follows:
“As shown in Figure 8, the spectral reflectance difference between cabbage and Pieris rapae is the largest at 823nm wavelength.
Figure 8. Curve of spectral reflectance difference between cabbage and Pieris rapae ”
6) Figure 8. The classification graph doesn't seem to have any values. mAP? The author should note the number of pest categories. If there is only one pest category, mAP should be equivalent to AP.
Response: We thank you very much for your comments. Given only one identification target (Pieris rapae) in this paper and no classification of multiple objects, there is no data on the curve of classification loss in the third column. And We have replaced all AP in the original manuscript with mAP.
7) Figure 9. The content has been changed to 99.3%, but Figure 9 is still not corrected?
Response: We thank you very much for your comments. And We are very sorry for the problems neglected in the last revision. 99.3% are the training results without data enhancement on the data set. Figure 9 in the original manuscript shows the training results after data enhancement. Since we re added the content of data enhancement, the result shown in Figure 9 is still 99.7%
The above are my main corrections in the paper and the responds to the Reviewer’s comments. Special thanks to you for your comments. We tried our best to improve the manuscript and made some changes in the manuscript. Any revised portion made to the manuscript are marked up using the “Track Changes” function. At the same time, we appreciate for Editors and Reviewers warm work earnestly, and hope that the correction will meet with approval.
Once again, thank you very much for your comments and suggestion. If you have any queries, please don’t hesitate to contact me at the address below.
Thank you and best regards.
Yours sincerely,
Zhengfang Hu
E-mail: [email protected]
Author Response File: 
Author Response.docx
Reviewer 2 Report
Authors, in their paper, present a field pest identification method based on YOLOv5 and hyperspectral technology. The results have demonstrated that this method can effectively identify pests with protective color characteristics in the complex field environment.
The paper presents an important approach that combines hyperspectral imaging with a machine learning method based on the YOLOv5 algorithm. It is referred to a specific case. However, it can be more improved. Some comments and suggestions are the following:
_ In the text, please define abbreviations before their first use.
_ A few syntactic errors exist. Some phrases are not very clear. Please correct.
_ Please, add a short description of the next sections at the end of the Introduction.
_ According to camera specifications, the camera of Imaging Source company is coming with an IR cut filter. How did the authors handle this issue and finally added an 850 nm infrared filter? Please explain.
_ Please, improve Figures 6 and 7. Please, increase in size Figure 6 for better clarity of it and add units in axes of Figure 7.
_ Please, take care so that a Figure and its caption are found on the same page (Figure 1).
_ Authors in their paper present a method to detect a specific pest (Pieris rapae only) on cabbages combining hyperspectral imaging with a machine learning method and the results are promising. However, to increase the worth of their work, they have to compare these results with the results of other similar approaches. Moreover, authors can add a piece of text with details concerning the implementation of their approach (hyperspectral imaging with a machine learning method) to other kinds of pests and other host plants. It will be very important if almost the same setup could be implemented in many different cases.
Author Response
Dear Reviewer:
Thank you very much for your comments concerning our manuscript entitled ‘Research on identification technology of field pests with protective color characteristics’ (ID: applsci-1577984). Those comments (at the bottom of the cover letter) are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval.
The main corrections in the manuscript and the responds to the comments are as following:
1) In the text, please define abbreviations before their first use.
Response: We thank you very much for your comments. We are very sorry for such negligence. And we have defined all abbreviations that appear for the first time in the original manuscript.
2) A few syntactic errors exist. Some phrases are not very clear. Please correct.
Response: We thank you very much for your comments. We are very sorry for such negligence. We have tried our best to correct the ambiguous phrases. If you think there are other ambiguous phrases, we are happy to explain them to you in detail.
3) Please, add a short description of the next sections at the end of the Introduction.
Response: We thank you very much for your comments. We have added relevant contents to the original manuscript. The supplementary contents are as follows:
“Pieris rapae and its host plant (cabbage) with a similar color to Pieris rapae were se-lected as experimental objects in this paper (Figure 1). As an extension of computer vision technology, near infrared imaging technology, especially the conventional imaging in the first NIR (NIR-I) window of 700 to 900 nm [23], can distinguish the target objects similar to the background in appearance characteristics [24]. It is widely used in insect species identification [25,26] and plant disease monitoring [27], but there is little re-search on pest identification. Thus, near-infrared imaging technology and YOLOv5 have been used to the identification of pests with protective color. Firstly, the average spectral characteristic curves of Pieris rapae and cabbage were obtained by hyperspectral experiment. By analyzing and comparing these two curves, the wavelength with the largest difference in spectral reflectance is obtained. According to this wavelength, the appropriate infrared filter, ring light source and other image acquisition equipment are selected to build an image acquisition platform. Collect a large number of pest images, and construct pest image data set by optimizing and expanding pest images. Finally, the appropriate deep learning model (YOLOv5) is selected to achieve the identification of field pests with protective color characteristics.
Figure 1. Field pests with protective color characteristics ”
4) According to camera specifications, the camera of Imaging Source company is coming with an IR cut filter. How did the authors handle this issue and finally added an 850 nm infrared filter? Please explain.
Response: We thank you very much for your comments. We removed the infrared cut-off filter. In order to avoid incorrect imaging due to the change of refraction light path, we use coated white glass with the same thickness and size to replace it. Since ordinary glass has a reflectance of about 8%, it can not be used in the optical circuit, which is easy to cause light to reflect back and forth between CMOS and glass, resulting in blurred and unclear imaging. Therefore, it is necessary to coat the glass to reduce the reflectivity of the glass, and the reflectivity of the coated glass is only about 1%.
5) Please, improve Figures 6 and 7. Please, increase in size Figure 6 for better clarity of it and add units in axes of Figure 7.
Response: We thank you very much for your comments. We have revised the original manuscript as required.
6) Please, take care so that a Figure and its caption are found on the same page (Figure 1).
Response: We thank you very much for your comments. We have revised the original manuscript as required. We also made sure that no other pictures showed such negligence.
7) Authors in their paper present a method to detect a specific pest (Pieris rapae only) on cabbages combining hyperspectral imaging with a machine learning method and the results are promising. However, to increase the worth of their work, they have to compare these results with the results of other similar approaches. Moreover, authors can add a piece of text with details concerning the implementation of their approach (hyperspectral imaging with a machine learning method) to other kinds of pests and other host plants. It will be very important if almost the same setup could be implemented in many different cases.
Response: We thank you very much for your comments. We have added relevant contents to the original manuscript. The supplementary contents are as follows:
“In addition, although only one pest with protective color characteristics (Pieris rapae) is considered in this paper, the relevant literature has proved that the near-infrared technology can distinguish the target objects with similar appearance characteristics and background [24], so this method can still be used to identify other pests with protective color characteristics and their host plants. For different pests, only the wavelength with the largest spectral reflectance difference between pests with protective color characteristics and their host plants needs to be obtained through hyperspectral test, so as to select the appropriate infrared filter. Replace the original infrared filter on the original image acquisition platform. That is, almost the same setting can be implemented in many different situations. In the future, other pests with protective color characteristics will be tested to further improve the universality of this method.”
The above are my main corrections in the paper and the responds to the Reviewer’s comments. Special thanks to you for your comments. We tried our best to improve the manuscript and made some changes in the manuscript. Any revised portion made to the manuscript are marked up using the “Track Changes” function. At the same time, we appreciate for Editors and Reviewers warm work earnestly, and hope that the correction will meet with approval.
Once again, thank you very much for your comments and suggestion. If you have any queries, please don’t hesitate to contact me at the address below.
Thank you and best regards.
Yours sincerely,
Zhengfang Hu
E-mail: [email protected]
Author Response File: Author Response.docx
Reviewer 3 Report
I read the paper titled "Research on identification technology of field pests with protective color characteristics"
I found the case study interesting for the readers of Applied Science working on the identification of field pests. As a reviewer, I have some comments and suggestions to make the paper adequate to be re-submitted.
1. the text still contains revisions that should be removed before submission to the journal
2. the quality of the images is not adequate for publication in a research journal, axes labels are not reported and the size too small for an appropriate reading
before this work, I consider it impossible to perform a proper refereeing
Author Response
Dear Reviewer:
Thank you very much for your comments concerning our manuscript entitled ‘Research on identification technology of field pests with protective color characteristics’ (ID: applsci-1577984). Those comments (at the bottom of the cover letter) are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval.
The main corrections in the manuscript and the responds to the comments are as following:
1) the text still contains revisions that should be removed before submission to the journal
Response: We thank you very much for your comments. Because our manuscript is a resubmitted article, the editor asked to upload a revised version. If you have trouble reading for this reason, I am very sorry.
2) the quality of the images is not adequate for publication in a research journal, axes labels are not reported and the size too small for an appropriate reading
Response: We thank you very much for your comments. In view of the problems you mentioned, such as insufficient image quality, too small image size and no coordinate axis, we have made corresponding modifications in the article. The content related to the picture has also been changed.
3) before this work, I consider it impossible to perform a proper refereeing
Response: We thank you very much for your comments. We have tried our best to make modifications according to the requirements. If there are any bad modifications, we welcome your further suggestions.
The above are my main corrections in the paper and the responds to the Reviewer’s comments. Special thanks to you for your comments. We tried our best to improve the manuscript and made some changes in the manuscript. Any revised portion made to the manuscript are marked up using the “Track Changes” function. At the same time, we appreciate for Editors and Reviewers warm work earnestly, and hope that the correction will meet with approval.
Once again, thank you very much for your comments and suggestion. If you have any queries, please don’t hesitate to contact me at the address below.
Thank you and best regards.
Yours sincerely,
Zhengfang Hu
E-mail: [email protected]
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Thanks to the author for the response
Author Response
Thank you very much for your comments and suggestions.
Best regards.
Reviewer 3 Report
I read the revised version of the paper titled "Research on identification technology of field pests with protective color characteristics"
After the revision work, I found the paper well-written, goal-focused, and well-realized. for this reason, I suggest accepting the paper after some references revision.
please check journals, in some cases are not italic.
Author Response
We thank you very much for your comments. We are very sorry for such negligence. Through the format of the reference paper template, we have modified the “Abbreviated Journal Name” and “Volume” that are not italicized in the references. Thank you again for your advice.
Best regards.
Author Response File: Author Response.docx
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
The study was conducted in a topical issue.
The study is comprehensive. The results of the study are useful in practice.
Notes:
The conclusions could be more specific. At the moment, the conclusions are quite abstract.
Author Response
Response: We are very sorry for our negligence of the conclusion. The original conclusion has been modified as follow:
In this paper, a field pest identification method based on YOLOv5 and hyperspectral technology was proposed. The results have demonstrated that this method can effectively identify pests with protective color characteristics in the complex field environment.
In the process of collecting pest images, to realize the identification of pests with protective color characteristics, obtain the average spectral characteristic curve of the pest and its host plants through hyperspectral test before image acquisition, comparing these two curves to get the wavelength with the largest reflectivity difference (850nm), and an appropriate infrared filter and ring light source are selected to build an image acquisition system. In order to improve the accuracy of pest identification, we collect pest images in different situations, expanded the original pest data set by data enhancement and select the appropriate target identification algorithm (YOLOv5). The detection results of the test set showed that the combination of YOLOv5 and hyperspectral technology can effectively identify field pests with protective color characteristics. This paper takes Pieris rapae and its host plant (cabbage) as the experimental object, its mAP was 99.7%, and the average detection speed is 0.565s per image.
Considering the future application scenario of pest identification, the current algorithm has some limitations in detection speed. In order to improve the detection speed of target detection algorithm, efficient models can be designed to accelerate the algorithms, such as decreasing the redundancy in weights by network pruning and knowledge distillation. While improving the detection speed, how to ensure the detection accuracy is also an aspect to be considered in the future. In addition, this paper only considers one pest with protective color characteristics (Pieris rapae). In the future, the applicability of this method can be further improved by studying more kinds of pests with protective color characteristics.
Author Response File: Author Response.docx
Reviewer 2 Report
The manuscript is well written (consider that I am not a native speaker) and clearly structured and well organized. The case study is very current, innovative and, no doubt, a great deal of agronomic interest will arise from this study because of the major problem represented by pests in recent years. For this reason, pest identification is, perhaps, the most paramount aspect dealt by recent research studies in the agricultural field, and I think that detection by imaging represents a powerful analytical method in this context. Authors look very aware of the state-of-art and of the current issues which must be overcome yet.
I found it very ambitious undertaking a study aimed to identify pests with protective colour characteristics, it is undoubtedly very challenging, and I got very impressed from it.
I think that the Abstract and Introduction parts can be considered satisfactorily fine and ready for publication.
Few corrections are needed for the M&M part:
At section 2.1, I would specify the name of the instrument.
At line 94, I found not clear the expression “20 4th instar Pieris rapae were…”. Please, give some more details.
At lines 100-101, it is not clear how the two parameters ‘angle of viewing’ and ‘distance between lens and sample’ were a-priori decided.
At lines 107 authors write about specific segments of Pieris rapae. How many segments are there, in total, in this pest? Why exactly the 7th and 9th segments are taken into account?
Concerning the cabbage, it is not clear if the work was performed considering only sound cabbage leaves or also some leaves already affected, with some entity of damage, by pests.
The Results and Discussions part is, in my opinion, not that comprehensive. I mean, certainly the work demonstrates that very good results can be achieved by the developed method, but things are presented very roughly and with obvious lack of detail, thus not supporting at all the importance of a such interesting and ambitious work.
Why authors decided to collect the definitive images by just using the 850 nm wavelength? Without doubt they clearly shows that at 850 nm the difference between Pieris rapae and cabbage reaches its maximum. However, all the wavelengths from 700 to about 990 seem to give a good discrimination between pest and cabbage, thus resulting promising in the pest identification. Does the color camera used for the acquisition of the definitive images (the DFK 125 21BU618) could be used to collect images at more than one wavelength?
In section 3.2 authors say “As demonstrated in the Figure”, but without mentioning specific one/s. Do they refer to Figure 8? Or Figure 9?
Figure 8 is very “messy” and do not convey any kind of information. What about x and y-axis? What do they represent? What about the trend depicted by the blue ‘line’ in each plot? Moreover, why the Precision and Recall plot for related to Validation set are not equally named as the corresponding ones related to the Training set?
Anyway, consider to also provide more information about all these five parameters (Box, Objectness, Classification, Precision, Recall).
At lines 247-248, what authors mean by “The visual comparison is to obtain the missing detection and wrong detection of pests through comparison”? Are there some parameters that need to be visually inspected by operators? If so, which ones?
The expression “AP value” is introduced for the first time at line 257. What does this acronym stand for? Is it kind of model Accuracy-Precision?
Figure 9 is also definitely not interpretable.
The Conclusions section is quite redundant since authors just stress some considerations already stated in previous sections. Here authors should structure more comprehensive considerations focused on the comparison between the actual scenario and the improvement generated by the findings achieved and described in the present manuscript.
There is a hint future perspective, but it is not so clear what authors exactly mean by “In future research, the dependence of network parameters and the training model on the amount of data should be reduced as much as possible”? I suggest providing some more explanation.
Comments for author File: 
Comments.pdf
Author Response
1) At section 2.1, I would specify the name of the instrument.
Response: Because the instrument name will make the picture less beautiful when expressed in the picture, we have a detailed description of the name of each instrument after the picture name.
“Figure 2. Schematic diagram of the hyperspectral test. (a) Spectrometer, (b) Lens, (c) Halogen lamp, (d) Pest sample, (e) objective table.”
2) At line 94, I found not clear the expression “20 4th instar Pieris rapae were…”. Please, give some more details.
Response: We apologize for the line 94 of this expression. The "20 4th instar Pieris rapae were…" was corrected to "20 Pieris rapae of fifth larval instars were…".
3) At lines 100-101, it is not clear how the two parameters ‘angle of viewing’ and ‘distance between lens and sample’ were a-priori decided.
Response: These two parameters are referenced in a research article.
Reference：Amir, A.; Ehsan, A.; Seyedeh, M. K.; Yu H.; Seunghoon, H.; Andrei, F. Miniature optical planar camera based on a wide-angle metasurface doublet corrected for monochromatic aberrations. Nat. Commun. 2016, 7, 1-9.
4) At lines 107 authors write about specific segments of Pieris rapae. How many segments are there, in total, in this pest? Why exactly the 7th and 9th segments are taken into account?
Response: There are 5 instars of Pieris rapae, and the body length of the fifth instar is 28 ~ 35 mm. The segments of the body are organized into three distinctive but interconnected units: a head, a thorax and an abdomen. According to the book (General Entomology), an abdomen consists of 10 segments. According to the reference, the effect of the 7th and 9th abdominal segments are the best.
Reference：Li, Y.J.; Xiang, Y.; Yang, Z.X.; Han, X.Z.; Lin, J.W.; Hu, Z.F. A laser irradiation method for controlling Pieris rapae larvae. Appl. Sci. 2021, 11, 9533.
5) Concerning the cabbage, it is not clear if the work was performed considering only sound cabbage leaves or also some leaves already affected, with some entity of damage, by pests.
Response: The work has already considered sound cabbage leaves and affected cabbage leaves. As shown in the figure below, the picture on the left shows sound cabbage leaves and the picture on the right shows affected cabbage leaves. Note: the location of pests has been circled in red in the figure.
6) The Results and Discussions part is, in my opinion, not that comprehensive. I mean, certainly the work demonstrates that very good results can be achieved by the developed method, but things are presented very roughly and with obvious lack of detail, thus not supporting at all the importance of a such interesting and ambitious work.
Response: We are very sorry for our negligence of conclusion. The conclusion has been revised in the original text. The revised content is as follows:
“In this paper, a field pest identification method based on YOLOv5 and hyperspectral technology was proposed. The results have demonstrated that this method can effectively identify pests with protective color characteristics in the complex field environment.
In the process of collecting pest images, to realize the identification of pests with protective color characteristics, obtain the average spectral characteristic curve of the pest and its host plants through hyperspectral test before image acquisition, comparing these two curves to get the wavelength with the largest reflectivity difference (850nm), and an appropriate infrared filter and ring light source are selected to build an image acquisition system. In order to improve the accuracy of pest identification, we collect pest images in different situations, expanded the original pest data set by data enhancement and select the appropriate target identification algorithm (YOLOv5). The detection results of the test set showed that the combination of YOLOv5 and hyperspectral technology can effectively identify field pests with protective color characteristics. This paper takes Pieris rapae and its host plant (cabbage) as the experimental object, its mAP was 99.7%, and the average detection speed is 0.565s per image.
Considering the future application scenario of pest identification, the current algorithm has some limitations in detection speed. In order to improve the detection speed of target detection algorithm, efficient models can be designed to accelerate the algorithms, such as decreasing the redundancy in weights by network pruning and knowledge distillation. While improving the detection speed, how to ensure the detection accuracy is also an aspect to be considered in the future. In addition, this paper only considers one pest with protective color characteristics (Pieris rapae). In the future, the applicability of this method can be further improved by studying more kinds of pests with protective color characteristics.”
7) Why authors decided to collect the definitive images by just using the 850 nm wavelength? Without doubt they clearly shows that at 850 nm the difference between Pieris rapae and cabbage reaches its maximum. However, all the wavelengths from 700 to about 990 seem to give a good discrimination between pest and cabbage, thus resulting promising in the pest identification. Does the color camera used for the acquisition of the definitive images (the DFK 125 21BU618) could be used to collect images at more than one wavelength?
Response: The color camera used for the acquisition of the 850nm images (the DFK 125 21BU618) through 850nm infrared filter. It is possible to collect images of multiple wavelengths by replacing infrared filters of different wavelengths.
8) In section 3.2 authors say “As demonstrated in the Figure”, but without mentioning specific one/s. Do they refer to Figure 8? Or Figure 9?
Response: The original manuscript has been modified. In section 3.2“As demonstrated in the Figure” refers to Figure 8.
9) Figure 8 is very “messy” and do not convey any kind of information. What about x and y-axis? What do they represent? What about the trend depicted by the blue ‘line’ in each plot? Moreover, why the Precision and Recall plot for related to Validation set are not equally named as the corresponding ones related to the Training set?
Anyway, consider to also provide more information about all these five parameters (Box, Objectness, Classification, Precision, Recall).
Response: The original manuscript has been modified. The revised content is as follows:
“As demonstrated in the figure 8, the graphs of the first three columns are box loss, objectness loss and classification loss from left to right. The first row is the performance indicators of the training set, and the second row is the training indicators of the verification set. Box loss indicates the extent to which the algorithm can position the center of the target and the extent to which the predicted bounding box covers the target. The abscissa of the curve is the epoches of the algorithm, and the ordinate represents the value of box loss. The smaller the value of box loss, the more accurate the predicted bounding box is. Objectness loss measures the probability that an object exists in a proposed re-gion of interest. If the objectivity is high, the bounding box is likely to contain an object. The abscissa of the curve is the epoches of the algorithm, and the ordinate represents the value of objectness loss. The smaller the value of objectness loss, the more accurate the target detection is. Classification loss inspires how well the algorithm can predict the correct class of a given object. Given only one identification target in this paper and no classification of multiple objects, there is no data on the curve of classification loss in the third column.
The graphs in columns 4 and 5 in Figure 8 show the precision, recall, [email protected] and [email protected]:0.95 of the model. The model improved swiftly in terms of precision, recall and mean average precision (mAP) before plateauing after about 20 epochs. [email protected] means that when IoU is set to 0.5, calculate the average precision (AP) of all pictures of each category, and then all categories are averaged, that is mean average precision. [email protected]:0.95 represents mAP at different IOU thresholds (IoU from 0.5 to 0.95 in steps of 0.05).”
10) At lines 247-248, what authors mean by “The visual comparison is to obtain the missing detection and wrong detection of pests through comparison”? Are there some parameters that need to be visually inspected by operators? If so, which ones?
Response: It means that the false detection or missed detection of pests can be observed by comparing the recognized pest image saved in the training process with the prediction bounding box and the position of pests in the image. There are no parameters that require visual inspection by the operator.
11) The expression “AP value” is introduced for the first time at line 257. What does this acronym stand for? Is it kind of model Accuracy-Precision?
Response: We apologize for our negligence in expression. The original manuscript has been modified. AP means the average precision.
12) Figure 9 is also definitely not interpretable.
Response: Figure 9 is a P-R curve with a threshold of 0.5 generated in the training process. The abscissa is recall and the ordinate is precision, which represents the relationship between accuracy and recall. The larger the area between the curve and the coordinate axis, the better its recognition effect.
13) The Conclusions section is quite redundant since authors just stress some considerations already stated in previous sections. Here authors should structure more comprehensive considerations focused on the comparison between the actual scenario and the improvement generated by the findings achieved and described in the present manuscript.
There is a hint future perspective, but it is not so clear what authors exactly mean by “In future research, the dependence of network parameters and the training model on the amount of data should be reduced as much as possible”? I suggest providing some more explanation.
Response: We are very sorry for our negligence of conclusion. The conclusion has been revised in the original text. The revised content is as follows:
“In this paper, a field pest identification method based on YOLOv5 and hyperspectral technology was proposed. The results have demonstrated that this method can effectively identify pests with protective color characteristics in the complex field environment.
In the process of collecting pest images, to realize the identification of pests with protective color characteristics, obtain the average spectral characteristic curve of the pest and its host plants through hyperspectral test before image acquisition, comparing these two curves to get the wavelength with the largest reflectivity difference (850nm), and an appropriate infrared filter and ring light source are selected to build an image acquisition system. In order to improve the accuracy of pest identification, we collect pest images in different situations, expanded the original pest data set by data enhancement and select the appropriate target identification algorithm (YOLOv5). The detection results of the test set showed that the combination of YOLOv5 and hyperspectral technology can effectively identify field pests with protective color characteristics. This paper takes Pieris rapae and its host plant (cabbage) as the experimental object, its mAP was 99.7%, and the average detection speed is 0.565s per image.
Considering the future application scenario of pest identification, the current algorithm has some limitations in detection speed. In order to improve the detection speed of target detection algorithm, efficient models can be designed to accelerate the algorithms, such as decreasing the redundancy in weights by network pruning and knowledge distillation. While improving the detection speed, how to ensure the detection accuracy is also an aspect to be considered in the future. In addition, this paper only considers one pest with protective color characteristics (Pieris rapae). In the future, the applicability of this method can be further improved by studying more kinds of pests with protective color characteristics.”
Author Response File: Author Response.docx
Reviewer 3 Report
Dear Authors,
Comments to your MS entitled “Research on identification technology of field pests with protective color characteristics”.
Your research topics and ideas are very interesting and promising. I am less experienced in machine learning and more trained in spectral issues, so my comments are related to this field.
Some comments:
Line 72-73: “The average spectral characteristic curves of Pieris rapae and cabbage were obtained by hyperspectral technology” What do you mean? How did you take an average spectrum of Pieris rapae. They are changing over time, depending on how many spectra you take. It is not traceable what you mean. There is no one spectrum to describe an animal or even plants, because they are very diverse in time and shape, mostly we use a lot of spectra to try to characterise living beings.
Line 98-99: “The spectrogram was corrected with the reflection reference plate.” In field spectroscopy or remote sensing we do not commonly say “spectrogram”, just simple “Spectrum”. If you talk about a reflection reference plate, do you mean the white reference? White referencing is not just a correction, it is a basic operation to be able to turn the target reflected radiance into target reflected relative reflection compared to the maximum reflectance of the reference white panel. You likely used one from LabSpehere or SphereOptics. Please, refer to which one you used and what white reflectance it had.
Line 126: How did you figure out that 850nm is the best for you, please underline this statement with statistical proof and other evidence.
Line 216-224: Please clear more scientifically the behaviour of the spectrum, the reader can see the differences but what is behind? Tissue structure, cells, water and co?
Summary:
Topic is interesting and relevant, it is not clear how you extracted the 850 nm wavelength from the hyperspectral data set. I see that it worked with 850 nm but it should be a better reasoning way, other statistical or physical. Figure 7 is not convincing me that at 850 nm there is the biggest difference in curve. I strongly recommend major revision.
Author Response
1) Line 72-73: “The average spectral characteristic curves of Pieris rapae and cabbage were obtained by hyperspectral technology” What do you mean? How did you take an average spectrum of Pieris rapae. They are changing over time, depending on how many spectra you take. It is not traceable what you mean. There is no one spectrum to describe an animal or even plants, because they are very diverse in time and shape, mostly we use a lot of spectra to try to characterise living beings.
Response: We are very sorry for the negligence in the expression of this sentence. By consulting the references [1-2], we have revised the misstatements in the original text. The revised expression is as follows:
“Using the Pieris rapae and cabbage spectral information obtained by hyperspectral technology, specific spectral characteristics of Pieris rapae and cabbage can be formed on the spectral curves.”
References：
- Ren, D.; Yu, H.Y.; Fu, W.W.; Zhang, B.; Ji, Q. Crop diseases and pests monitoring based on remote sensing: a survey. In Proceedings of 2010 Conference on Dependable Computing, Yichang, China, 20-22 November 2010.
- Shi, Y.; Huang, W.J.; Luo, J.H.; Huang, L.S.; Zhou, X.F. Detection and discrimination of pests and diseases in winter wheat based on spectral indices and kernel discriminant analysis. Comput. Electron. Agric. 2017, 141, 171-180.
2) Line 98-99: “The spectrogram was corrected with the reflection reference plate.” In field spectroscopy or remote sensing we do not commonly say “spectrogram”, just simple “Spectrum”. If you talk about a reflection reference plate, do you mean the white reference? White referencing is not just a correction, it is a basic operation to be able to turn the target reflected radiance into target reflected relative reflection compared to the maximum reflectance of the reference white panel. You likely used one from LabSpehere or SphereOptics. Please, refer to which one you used and what white reflectance it had.
Response: I am very sorry for the negligence expressed in the paper. But in this paper, we have shown that the hyperspectral experiments of cabbage caterpillar and cabbage inner leaves were carried out indoors. And A Munsell grey panel with 18% reflectance to use as a standard to calibrate hyperspectral images.
3) Line 126: How did you figure out that 850nm is the best for you, please underline this statement with statistical proof and other evidence.
Response: We apologize for the negligence of the expression in the paper. It has been revised in the original text. The amendments are as follows:
“Based on the products of many filter production companies on the market, the optional filters in the near-infrared band range (780-1000nm) are divided into 850 nm and 950 nm. At the wavelength of 850 nm, the difference in the spectral reflectance between cabbage and Pieris rapae was the largest. Therefore, 850nm infrared filter and 850nm ring light source were selected to acquire pest images.”
The following figure shows the spectral reflectance difference curve between the spectral characteristic curve of inner leaves of cabbage and the spectral characteristic curve of Pieris rapae. The abscissa is the wavelength and the ordinate is the difference of reflectivity. It can be clearly seen that 823nm is the wavelength with the largest reflectance difference.
4) Line 216-224: Please clear more scientifically the behaviour of the spectrum, the reader can see the differences but what is behind? Tissue structure, cells, water and co?
Response: As mentioned in the next paragraph (line 225-229), figure 1 in the paper shows the pest image under visible light in natural environment, and figure 4 and figure 5 in the paper show the pest image under 850nm infrared filter and 850nm ring light source. The pests in figure 4 and figure 5 can find the location of pests more quickly and simply than figure 1. As for the reason behind it, it is more due to the 850nm infrared filter and 850nm ring light source, which has nothing to do with the pest itself.
5) Topic is interesting and relevant, it is not clear how you extracted the 850 nm wavelength from the hyperspectral data set. I see that it worked with 850 nm but it should be a better reasoning way, other statistical or physical. Figure 7 is not convincing me that at 850 nm there is the biggest difference in curve. I strongly recommend major revision.
Response: We apologize for the negligence of the expression in the paper. It has been revised in the original text. The amendments are as follows:
“Based on the products of many filter production companies on the market, the optional filters in the near-infrared band range (780-1000nm) are divided into 850 nm and 950 nm. At the wavelength of 850 nm, the difference in the spectral reflectance between cabbage and Pieris rapae was the largest. Therefore, 850nm infrared filter and 850nm ring light source were selected to acquire pest images.”
Author Response File: Author Response.docx
Reviewer 4 Report
This paper proposes to improve the accuracy of object detection by using the reflection differences of different targets in the near-infrared spectrum. This topic is of great interest to readers in this field, which can combine deep learning and invisible light imaging technology, as well as the availability of implanted low-power devices in the future. However, this paper is still insufficient in terms of theoretical assumptions and experimental methods, so it is expected that the author can re-examine the hypothesis (for example, supplement the recognition performance under visible light images to support the hypothesis.) and correct the training data before resubmitting.
Section 1. Introduction, As mentioned above, please consider the recognition performance under visible light. The theory will be more strongly supported if you can do a visible light image recognition performance test under the same data set.
Line 69, which should not appear in manuscripts submitted to journals. Authors should do their best to proofread.
Figure 1. Under today's powerful deep learning network, such backgrounds and targets can be easily identified. Please provide evidence that cannot be detected in relevant visible light images.
Section 2.2. Data augmentation should not be used in this way, you will be breaking similar copies of the original image into test and validation, which will cause over-learning and not represent true model performance.
Figure 7. Please reconfirm the location of the largest spectral difference. The biggest difference in this graph should not be at 850nm. If it is only due to the use of a wide range of supplier sources, just indicate that the existing parts are obtained.
Author Response
1) However, this paper is still insufficient in terms of theoretical assumptions and experimental methods, so it is expected that the author can re-examine the hypothesis (for example, supplement the recognition performance under visible light images to support the hypothesis.) and correct the training data before resubmitting.
Response: We very appreciate your suggestions. But we want to explain our ideas to you in more details. This paper mainly considers the recognition effect of other research articles on pest recognition with protective color features. Therefore, we think it can support the theoretical assumptions and experimental methods of this paper.
2) Section 1. Introduction, As mentioned above, please consider the recognition performance under visible light. The theory will be more strongly supported if you can do a visible light image recognition performance test under the same data set.
Response: The theory of this paper mainly compares the recognition effect of pests with protective color characteristics in the existing researches [1-3], but most of these researches are aimed at the situation with large differences between pests and background (such as the sticky card etc.), and the recognition effect of pests with protective color characteristics in these studies is poor, which can also give strong support to the theory of this paper.
References：
- Zhong, C.Y.; Li, X.; Liang, C.B.; Xue, Y.Z. A cabbage caterpillar detection method based on computer vision. Shanxi Electron. Technol. 2020, 01, 84-86.
- Gao, X.; Tang, Y.; Chen, Y.Y.; Cui, H.M.; Wang, H.B. Research on cabbage pest identification based on image processing. Jiangsu Agric. Sci. 2017, 45, 235-238.
- Wu, X.; Zhang, W.Z.; Qiu, Z.J.; Cen, H.Y.; He, Y. A novel method for detection of Pieris rapae larvae on cabbage leaves using Nir hyperspectral imaging. Appl. Eng. Agric. 2016, 32, 311-316.
3) Line 69, which should not appear in manuscripts submitted to journals. Authors should do their best to proofread.
Response: We are very sorry for the omission in the format of line 69. The original text “identification with protective color characteristics.” has been revised to that “At present, there are few studies on field pest identification with protective color characteristics.”.
4) Figure 1. Under today's powerful deep learning network, such backgrounds and targets can be easily identified. Please provide evidence that cannot be detected in relevant visible light images.
Response: According to the references, the effect of pest identification under visible light is poor. Wu et al. [1] proposed a pest identification method based on convolutional neural network (CNN) under natural light background, and the recognition rate of Pieris rapae has been increased to 81% by improving the model. Gao et al. [2] proposed a method which uses the plant pest identification and detection methods of Euclidean Distance to detect the plant pests. This paper takes Pieris rapae and its host plant as the experimental object, and the experimental results show that these two methods can correctly identify the disease area and the accuracy rate reaches 88.33%. Gao et al. [3] extracts the shape features of pests by preprocessing the pest image, and uses fuzzy recognition to distinguish the types of pests, so as to effectively identify some kinds of cabbage pests. The pest characteristics selected in this study are relatively bright, the leaf surface of the selected crop cabbage is relatively smooth and flat, the color change is uniform, the recognition effect is relatively ideal, and the recognition rate is higher than 80%.
References：
- Wu, X.; Zhang, W.Z.; Qiu, Z.J.; Cen, H.Y.; He, Y. A novel method for detection of pieris rapae larvae on cabbage leaves using NIR hyperspectral imaging. Appl. Eng. Agric. 2016, 32, 311-316.
- Gao, X.; Wang, H.C. Research on cabbage rapae pests automatic recognition system based on machine vision. J. Agric. Mech. Res. 2015, 37, 205-208+222.
- Gao, X.; Tang, Y.; Chen, Y.Y.; Cui, H.M.; Wang, H.B. Research on cabbage pest identification based on image processing. Jiangsu Agric. Sci. 2017, 45, 235-238.
5) Section 2.2. Data augmentation should not be used in this way, you will be breaking similar copies of the original image into test and validation, which will cause over-learning and not represent true model performance.
Response: Due to the small sample of the original data set in this paper, the traditional data enhancement method (such as flipping, clipping, rotation, etc.) has been adopted to expand the original data set. This method is suitable for the samples with small data set, and can realize the transformation of small samples into the large scale, good diversity and high-quality data set, so as to train the model with better performance and better generalization [1-3].
References：
- Gan, L.; Shen, H.F.; Wang, Y.; Zhang, Y.J. Data augmentation method based on improved deep convolutional generative adversarial networks. J. Comput. Appl. 2021, 41, 1305-1313.
- Chen, X.; Chen, M.X. SAR target recognition based on CNN trained by augmented training samples. J. Chongqing Univ. Technol. (Nat. Sci.) 2020, 34, 86-93.
- Jiang, J.; Xiong, C.Z. Data augmentation with multi-model ensemble for fine-grained category classification. J. Graph. 2018, 39, 244-250.
6) Figure 7. Please reconfirm the location of the largest spectral difference. The biggest difference in this graph should not be at 850nm. If it is only due to the use of a wide range of supplier sources, just indicate that the existing parts are obtained.
Response: We apologize for the negligence of the expression in the paper. It has been revised in the original text. The amendments are as follows:
“Based on the products of many filter production companies on the market, the optional filters in the near-infrared band range (780-1000nm) are divided into 850 nm and 950 nm. At the wavelength of 850 nm, the difference in the spectral reflectance between cabbage and Pieris rapae was the largest. Therefore, 850nm infrared filter and 850nm ring light source were selected to acquire pest images.”
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
Dear Authors,
thanks for your comments and references. I accept your comments in present form.
Reviewer 4 Report
Thank the authors for their replies. I still maintain the same opinion about the review comments. Except for the reason for the selection of the spectrum, this matter has been satisfied. The main issue is whether the recognition performance of the SOTA model in visible light is insufficient, which is not supported in the literature, so I expect the author to find newer literature or do experiments to confirm this argument. Moreover, for doubts about data augmentation, it is recommended that the author could consult other teachers or peers with deep learning experience to examine it.
