UAV, a Farm Map, and Machine Learning Technology Convergence Classification Method of a Corn Cultivation Area
Round 1
Reviewer 1 Report
This work describes the use of certain Machine Learning methods for corn cultivation area classification. Some of my concerns are:
- Abstract -
* "The existing object classification method takes a lot of time and effort to optimize because segmented objects can use a large number of derived functions" : Which method? You refer to something which has not been yet introduced. Please rephrase this part to include more details about what this statement refers to.
* "[...] of classification techniques, i.e., a Support Vector Machine (SVM) and Random Forest (RF)" : This statement suggests that only two classification techniques exist. Please rephrase.
* "The main research results obtained are as follows." : Remove
- Introduction -
* What is "DT; Digital Transformation" and how does it relate to classification algorithms?
* I would include a paragraph at the end of this section explaining the structure of the manuscript.
- 2.4 Crop classification ... -
* In the paragraph starting with "Machine learning is a..." you are confusing two terms. Machine Learning is a branch and Deep Learning is another branch within the former one. Thus, you should not write about Machine Learning and then compare its limitations with those of Deep Learning. Deep Learning may be limited by data scarcity (what you mention about spatial information) but this may not be true for general Machine Learning methods.
* "The kernel method is a method" : Rephrase. Maybe "The kernel trick..."
- 3.3 Crop Classification Results and Accuracy Verification -
* "The overall accuracy of RF algorithm application was 98.84% higher than that of SVM application." I do not follow this claim. RF obtained a 98.84% of overall classification accuracy whereas SVM achieved a 95.58%. Where is than increase with one with respect to the other one? 98.84% is the absolute classification rate of one of the proposals not the (relative) improvement
- Overall points to discuss -
* I understand the different metrics considered for the evaluation of the task but I think there is an underlying problem with them. The task at hand, I guess, depicts an imbalanced distribution of the classes at issue (e.g., corn VS not corn). In those cases classification accuracy is not considered an appropriate metric for assessing the scenario but one would usually resort to other figures of merit as, for instance, the F-measure.
In this regard I think the authors should discuss about whether there is an imbalanced class distribution among the data of the problem and, if so, clearly state why you use these metrics. In any case, if these distribution is not balanced, classification is not an appropriate figure of merit.
* Do you consider any train/validation/test partitions?
* Do you consider any cross-validation procedure? There is a single mention to it the caption of Figure 11 for tuning the SVM method, but I could not find any other reference to it.
* The manuscript mentions quite often Deep Learning but there is no work related to this. I am not claiming that you consider SVM and RF as Deep Learning methods since they are not, but I would recommend avoiding these points as there is nothing about it in the work.
* The title is too explicit and verbose. I would suggest making it more concise.
Author Response
Dear Reviewer.
Thank you for your meticulous, informative, and reasonable comments on our manuscript despite your busy schedule. I reinforced the manuscript as possible following your point.
-Abstract –
Point 1: The existing object classification method takes a lot of time and effort to optimize because segmented objects can use a large number of derived functions": Which method? You refer to something which has not been yet introduced. Please rephrase this part to include more details about what this statement refers to.
Response 1: Thank you for your thoughtful point and comment.
As you pointed out, I made the following context clear:
“Abstract: South Korea’s agriculture is characterized by a mixture of various cultivated crops. In such an agricultural environment, convergence technology for ICT (Information, Communications, and Technology) and AI (Artificial Intelligence) as well as agriculture is required to classify objects and predict yields. In general, the classification of paddy fields and field boundaries takes a lot of time and effort. The Farm Map was developed to clearly demarcate and classify the boundaries of paddy fields and fields in Korea. Therefore, this study tried to minimize the time and effort required to divide paddy fields and fields through the application of the farm map. “
Point 2: The abstract has to be rewritten. It reads like it is a summary of the methodology alone. The authors must provide an overview of the results too. "[...] of classification techniques, i.e., a Support Vector Machine (SVM) and Random Forest (RF)" : This statement suggests that only two classification techniques exist. Please rephrase.
Response 2: Thank you for your careful reading.
As you pointed out, I made the following context rephrase:
“This study aimed to evaluate the applicability and effectiveness of machine learning classification techniques using a Farm Map in object-based mapping of agricultural land using unmanned aerial vehicles (UAVs). In this study, the advanced function selection method for object classification is to improve classification accuracy by using two types of classifiers, Support Vector Machine (SVM) and Random Forest (RF).”
Point 3: "The main research results obtained are as follows." : Remove
Response 3: I have deleted that sentence as you pointed out.
-Introduction-
Point 4: * What is "DT; Digital Transformation" and how does it relate to classification algorithms?
Response 4: Thank you for your attentive review.
As you pointed out, this sentence has been deleted as it is not significantly relevant to this study.
Point 5: I would include a paragraph at the end of this section explaining the structure of the manuscript
- 2.4 Crop classification ... -
In the paragraph starting with "Machine learning is a..." you are confusing two terms. Machine Learning is a branch and Deep Learning is another branch within the former one. Thus, you should not write about Machine Learning and then compare its limitations with those of Deep Learning. Deep Learning may be limited by data scarcity (what you mention about spatial information) but this may not be true for general Machine Learning methods.
Response 5: Thank you for your attentive review.
“However, there is a limit to the application of deep learning technology in that there are not many spatial information data for a wide area at this stage. “
As you pointed out, this sentence was deleted because it was out of context and content.
Point 6: "The kernel method is a method" : Rephrase. Maybe "The kernel trick..."
Response 6: Thank you for your attentive review.
I almost made a big mistake.
Thank you for your careful review. The contents have been changed to Kernel trick.
- 3.3 Crop Classification Results and Accuracy Verification -
Point 7: "The overall accuracy of RF algorithm application was 98.84% higher than that of SVM application." I do not follow this claim. RF obtained a 98.84% of overall classification accuracy whereas SVM achieved a 95.58%. Where is than increase with one with respect to the other one? 98.84% is the absolute classification rate of one of the proposals not the (relative) improvement
Response 7: Thank you for your attentive review.
The purpose of this study is not to improve the accuracy of SVM and RF algorithms, but to propose a method to use Farm Map data for crop classification using UAV images and to apply SVM and RF algorithms. This means that the RF algorithm showed higher accuracy than the SVM as a result of applying the Farm Map.
- Overall points to discuss -
Point 8: I understand the different metrics considered for the evaluation of the task but I think there is an underlying problem with them. The task at hand, I guess, depicts an imbalanced distribution of the classes at issue (e.g., corn VS not corn). In those cases classification accuracy is not considered an appropriate metric for assessing the scenario but one would usually resort to other figures of merit as, for instance, the F-measure.
In this regard I think the authors should discuss about whether there is an imbalanced class distribution among the data of the problem and, if so, clearly state why you use these metrics. In any case, if these distribution is not balanced, classification is not an appropriate figure of merit.
Response 8: Thank you for your attentive review and advice.
The data imbalance problem you pointed out was supplemented by reflecting the contents of F-measure in Table 1, Table 5, and Table 6.
There is no difficulty in classifying other than two crops. However, this study was conducted to quickly explore the corn cultivation areas for a wide area and reflect it in the government policy.
We would like to present the points you pointed out in more detail through the following research.
Point 9: Do you consider any train/validation/test partitions?
Do you consider any cross-validation procedure? There is a single mention to it the caption of Figure 11 for tuning the SVM method, but I could not find any other reference to it.
Response 9: Thank you for your careful review.
The data used in this study were partitioned into Train and Test data.
Validation was divided into 5 folds through k-fold validation in the Train dataset, and cross-validation was performed by repeating training 10 times.
Relevant information is provided as a supplement to "3.2 Data Preprocessing and Hyperparameter Tuning"
Point 10: The manuscript mentions quite often Deep Learning but there is no work related to this. I am not claiming that you consider SVM and RF as Deep Learning methods since they are not, but I would recommend avoiding these points as there is nothing about it in the work.
Response 10: Thank you for your attentive review and advice.
As you pointed out, content that is not directly related to the contents of this study has been deleted or modified.
Point 11: The title is too explicit and verbose. I would suggest making it more concise.
Response 11: Thank you for your kind review.
As you pointed out, I changed the paper title to be concise as follows: “UAV, a Farm Map and Machine Learning Technology Convergence Classification Method of a Corn Cultivation Area”
Thank you very much again.
Sincerely.
Author Response File: 
Author Response.docx
Reviewer 2 Report
It is a valid manuscript, but some issues of scientific style need to be corrected. I make some suggestions to improve the manuscript.
In the introduction there is an excessive use of multiple citations. I advise authors to cite only one reference per sentence. Or otherwise develop the importance of all citations by developing what the actual content of the cited manuscripts is, and not in a general way as is currently done.
I propose to add a table of acronyms to help the reader quickly find the abbreviations used.
Since so many manuscripts on image classification are cited. I propose that these be put in table mode so that the reader can quickly find out the date, the crop on which it was used, the geographical area, etc.
*Table 3
2020-05-08= 2721 ha
2020-06-18= 2680 ha
Table 4
Corn = 109.94 ha
Others = 749.05 ha
What about all the missing surface area? Por favor comentese en el trabajo, esto es confuso para el lector Please comment on the manuscript, this is confusing for the reader.
Clarify also in the conclusions or limit to the area actually identified:
- “, a wide area of about 2,700 ha was studied”
- As a result of applying the SVM algorithm, the corn cultivation area was estimated to be 96.54 ha, showing an accuracy of 90.27%. The RF algorithm estimated the corn cultivation area to be 98.77 ha and showed higher accuracy than SVM at 92.36%.
The big question I have:
So only less than 100 ha have been classified? Why have a flight of more than 2,000 ha and then classify less than 100? If the idea is to see how the maize crop can be identified by the various techniques proposed, why have a flight of more than 2,000 ha? This should be clear, also in the abstract and conclusions.
Author Response
Dear Reviewer.
Thank you for your meticulous, informative, and reasonable comments on our manuscript despite your busy schedule. I reinforced the manuscript as possible following your point.
Point1: In the introduction there is an excessive use of multiple citations. I advise authors to cite only one reference per sentence. Or otherwise develop the importance of all citations by developing what the actual content of the cited manuscripts is, and not in a general way as is currently done.
Response 1: Thank you for your attentive review and advice.
I have modified it as much as possible to fit what you pointed out.
Point2: I propose to add a table of acronyms to help the reader quickly find the abbreviations used. Since so many manuscripts on image classification are cited. I propose that these be put in table mode so that the reader can quickly find out the date, the crop on which it was used, the geographical area, etc.
Response 2: Thank you for your kind review and advice.
For what you point out, we have added an abbreviation table in the appendix.
Point3: Table 3
2020-05-08= 2721 ha
2020-06-18= 2680 ha
Table 4
Corn = 109.94 ha
Others = 749.05 ha
What about all the missing surface area? Por favor comentese en el trabajo, esto es confuso para el lector Please comment on the manuscript, this is confusing for the reader.
Response 3: Thank you for your attentive review.
As presented in the paper, we used only the area corresponding to the Farm Map range among the UAV shooting areas for analysis. The other areas are non-agricultural land and consist of various types of cover such as villages, roads, mountains, and rivers. Since the purpose of this study is to quickly identify the cultivated area of ​​corn among agricultural land, non-agricultural land was omitted from the classification target.
Point4: Clarify also in the conclusions or limit to the area actually identified:
“, a wide area of about 2,700 ha was studied”
As a result of applying the SVM algorithm, the corn cultivation area was estimated to be 96.54 ha, showing an accuracy of 90.27%. The RF algorithm estimated the corn cultivation area to be 98.77 ha and showed higher accuracy than SVM at 92.36%.
The big question I have:
So only less than 100 ha have been classified? Why have a flight of more than 2,000 ha and then classify less than 100? If the idea is to see how the maize crop can be identified by the various techniques proposed, why have a flight of more than 2,000 ha? This should be clear, also in the abstract and conclusions.
Response 3, 4: Thank you for your careful review.
First, this study begins with a study to find the optimal conditions using various platforms.
In order to use it for agricultural statistics, it is necessary to identify various crops at various times. Because Korean agriculture cultivates a variety of crops in a small area, low-resolution images such as satellites make it difficult to distinguish corn among numerous crops due to the nature of Korean agriculture, so UAV images were used.
As presented in “2.1 Study Area” in the main paper, the total area of ​​Gammul-myeon is about 4280ha, and the image taken including the maximum amount of agricultural land is about 2700ha. The area of ​​pure farmland calculated based on the Farm Map in the UAV image is about 860ha, of which about 100ha is corn. To calculate the corn area of ​​100 ha, UAV images of the entire persimmon face are required. The reason for the flight over 2000ha is that agricultural land is widely distributed within the target area. In order to improve the problem you pointed out, this is a study to suggest that using Farm Map can save a lot of time and effort.
Thank you very much again.
Sincerely.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
All my concerns have been properly addressed, so I recommend the acceptance of the paper
Reviewer 2 Report
The authors have improved their manuscript following my recommendations. For my side the manuscript can be accepted for publication.
