Parcel-Level Mapping of Horticultural Crops in Mountain Areas Using Irregular Time Series and VHR Images Taking Qixia, China as An Example

Round 1

Reviewer 1 Report

Please see the attached file.

Comments for author File: Comments.pdf

Author Response

Point 1: The introduction is well written and it has sufficient information on the scientific problem and how to solve it. The literature review is good, though, authors should put more spotlights on the developed approach in the highlights.

Response 1: Thank you for your suggestions. According to your suggestions, we have revised the relevant content in the introduction and added the introduction to the method.

 

Point 2: It is highly recommended to add more details on phenological information of

different crops based on Figure (2).

Response 2: Thanks for your suggestion, we modified the Figure 2.

 

 

Point 3: The task of overlapping the Sentinel-2 images used in the study lack to some details such as Resampling level of the scenes to fit on each other, pre-processing level of each image and the outcome is better given in a visual form (figure or image).

Response 3: Thank you for your suggestion. The Sentinel-2 images used in the study is Level -2A, which can be directly downloaded from GEE. The data has been radiometric calibrated and atmospheric corrected. The bands used in the study are visible bands (RGB) and NIR (B2, B3, B4, B8), with a resolution of 10m, so resampling is not required. The data is then re projected. These data need to be re projected. UTM projection is used. UTM projection adopts 6-degree zoning. The study area is located in the Northern Hemisphere, with the belt number of 51. Therefore, the coordinate system of the re projection is “WGS_1984_UTM_Zone_51N”.

 

Point 4: Please provide more details about the training and validation samples of crop classification in the case study. Moreover, the sampling type is not defined.

Response 4: Thank you for your suggestion. We added an introduction to sample data and sample expansion in the revised manuscript.

 

Point 5: Please elaborate more on Figure 6.

Response 5: Thank you very much for your advice. We added more explanations to Figure 6 in the revised manuscript to make it easier for readers to understand the definition of irregular sequences in this study.

 

Point 6: Figure 12 (a,b,c): The labels are hard to read. Please consider revising. Response 6: Thank you for your suggestion. We modified the label in Figure 12. In addition, relevant descriptions in the “comparative experiment” were modified to make the readers better understand.

 

Point 7: It would be more useful if the authors add a comparative execution time for the methodologies.

Response 7: Thank you very much for your advice. We are sorry that we did not add a comparison of the execution time of different methods in the manuscript. The main reasons are as follows.

• It takes different time to build different time series data. When building irregular time series, each grid should be processed separately, which is more time-consuming than building regular time series data.
• Different equipment is used in the process of model In order to speed up the experimental progress, we conducted the model training in parallel, using different equipment. At this time, the execution time of the model is meaningless.

To sum up, we are sorry that we cannot add the comparison of execution time of different methods in the manuscript.

 

Point 8: Typical references are need here.

Response 8: Thank you for your suggestion. We added references to the manuscript.

 

Point 9: Please add a table or any visualizations for the metrics mentioned in “Analysis

of classification experiment results”.

Response 9: Thank you for your suggestions sincerely. I'm sorry, we didn't catch your meaning. If the matrix you mentioned refers to the confusion matrix in the manuscript, (a, b, c) in Figure 12 refers to the confusion matrix of three classification experiment results. If you have any questions or suggestions about our reply, please point them out in the subsequent revision process of the article. Thank you for your valuable suggestions on our manuscript.

 

Point 10: There should be a brief explanation of the limitations encountered.

Response 10: Thank you for your suggestion. We cut some content to make it more concise.

Author Response File: Author Response.pdf

Reviewer 2 Report

This study proposes a parcel-level mapping method for perennial orchard crops. This is an interesting study and the authors have conducted various experiments to prove a proposed framework is superior.

The manuscript is well structured and contains a large amount of very interesting data but an adequate review appears necessary before publication. I list some comments and suggestions for improvements below:

1. The phenological information of wheat and peanut shown in Fig. 2 is inconsistent with the facts. According to daily experience, in Shandong province, China, winter wheat is usually sown in October of the previous year and harvested in June of the following year, while peanuts are usually sown in April and harvested in September of the current year. Obviously, the phenological information presented in Fig. 2 for these two crops is wrong.

2. Sentinel 2 images were acquired in 2019, while field surveys were conducted in 2021. How to ensure that the ground samples and image features are consistent? Especially the seedling and greenhouse category, is there any change in the land use type in the same location in 2019 and 2021?

3. There are some irregular large-sized parcels in Fig. 9(a2), and they seem to each contain multiple real agricultural parcels, which do not exactly match the real parcels presented in Fig. 9(a1). This casts doubt on the accuracy of the parcels extracted by the algorithm. The manuscript does not specify whether the accuracy of parcel extraction was verified. In other words, whether the boundaries of the parcels generated by the algorithm can exactly match the parcels in real agricultural management.

4. This manuscript aims to map horticultural crops. It is mentioned that the natural forest is masked out through image texture information. I wonder how other land use types other than natural forest are treated. For example, whether land use types such as roads, water bodies, and rural settlements are also masked?

5. On lines 146-147, “other crops” are mentioned, and the sample in Fig. 4 also mentions “others”, but there are no “other crops” in the final crop classification results. I’d like to know whether “other crops” are involved in the classification process,  and how are they distinguished from the four types of crops in the manuscript?

6.The overall accuracy of the “irregular-time-series +2DCNN” method is only 1.83% higher than that of the “regular-544 time-series +2DCNN” method, and the kappa value of the former is only 0.0327 higher than that of the latter. Poor accuracy raises questions about the usefulness of using images with irregular time series.

Here are some minor comments:

1. On line 129, an error was reported.

2. Line 218 is duplicated with line 157, the former should be 2.3 instead of 2.2.

3. Figures should be numbered in the order in which they are mentioned in the manuscript, thus, Fig. 4 (mentioned at line 192) should come before Fig. 3 (mentioned at line 205).

4. I think Fig. 3 mentioned on lines 449 and 455 should be Fig. 5.

5. Please indicate where the three sub-regions (a1), (b1) and (c1) in Fig. 11 are located in the study area.

6. On line 513, how did the 10963 real values come from, please add an explanation.

7. On line 550, “Figure 12” appears twice.

8. The word of parcel and plot have different meanings in the land survey. The usage needs to be standardized throughout the paper. Similarly, the use of nurseries and seedling needs to be standardized.

9. The 39th reference format is not standardized

Author Response

Point by point response to the comments of the reviewer1:

Point 1: The phenological information of wheat and peanut shown in Fig. 2 is inconsistent with the facts. According to daily experience, in Shandong province, China, winter wheat is usually sown in October of the previous year and harvested in June of the following year, while peanuts are usually sown in April and harvested in September of the current year. Obviously, the phenological information presented in Fig. 2 for these two crops is wrong.

Response 1: Thank you for pointing out the mistakes in the manuscript. We are sorry for our carelessness. We consulted local farmers in Shandong, obtained local phenological characteristics, and updated the phenological information in Fig. 2. In addition, the corn is divided into spring corn and summer corn, which is more suitable for the local planting type.

 

 

Point 2: Sentinel 2 images were acquired in 2019, while field surveys were conducted in 2021. How to ensure that the ground samples and image features are consistent? Especially the seedling and greenhouse category, is there any change in the land use type in the same location in 2019 and 2021?

Response 2: Thank you for the concern. In fact, when we took samples on the spot, we also considered whether the authenticity of the samples could be guaranteed after two years. The following sampling measures have been taken to ensure the authenticity of the samples in 2019.

• Judge whether the fruit tree can blossom and bear fruit. Apple trees generally begin to bear fruit after three or more years of growth, and the flowering phase is from late April to early May. The field survey was carried out in the early and middle days of May 2021 (5.1-5.12), in the flowering period of apple trees. We mainly collect samples of adult flowering apple orchards to ensure that the samples also exist in 2019. Cherry trees generally bear fruit after 3-6 years of planting and gradually matured at the end of April and May. During the field survey, it is the maturation stage of the big cherry. Therefore, it can be judged whether the fruit tree has been planted for more than three years by judging whether there are fruits on the tree.
• High resolution images are used as reference data. When collecting samples, we compared the high-resolution image data in 2019 with the field crops. Objects with unique texture features on high-resolution images, such as greenhouses and seedling orchard, can be easily distinguished by visual interpretation, as shown in the following figure. Using high resolution images as auxiliary data during field investigation can ensure that the collected greenhouse samples and seedling samples also exist in 2019.
• The samples collected are mainly located in the interior of the large parcels, and the sample points are placed in the middle of the parcel to avoid the wrong sample category caused by positioning errors.

 

Point 3: There are some irregular large-sized parcels in Fig. 9(a2), and they seem to each contain multiple real agricultural parcels, which do not exactly match the real parcels presented in Fig. 9(a1). This casts doubt on the accuracy of the parcels extracted by the algorithm. The manuscript does not specify whether the accuracy of parcel extraction was verified. In other words, whether the boundaries of the parcels generated by the algorithm can exactly match the parcels in real agricultural management.

Response 3: Thank you sincerely for your question. We have modified Figure 9. Regarding the problem that the accuracy of the parcel extraction results has not been verified, we have added relevant verification in the manuscript.

 

 

Point 4: This manuscript aims to map horticultural crops. It is mentioned that the natural forest is masked out through image texture information. I wonder how other land use types other than natural forest are treated. For example, whether land use types such as roads, water bodies, and rural settlements are also masked?

Response 4: Thank you for the concern. In the manuscript, considering that the natural forest and orchard crops have similar phenological features but different texture features, as shown in the figure below, we mask the natural forests on the VHR image based on the texture features to avoid the impact of natural forests on the classification results.

As for the classification of other ground feature types you mentioned, we are sorry that our expositionin the manuscript is not clear. In fact, during model training, we set five labels, including the four categories of research objects (apples, cherries, seedling, and greenhouses) and the "other" category. The “other” label includes buildings, water, roads, bare soil, and other non-orchard crops (corn, wheat, peanut). Samples of these land use types can be obtained from VHR images. The characteristics of these surface objects differ greatly from those of orchard crops in time series, so whether LSTM or 2DCNN can easily distinguish them. Considering that other readers may have the same question as you, we have added an explanation to this question in the manuscript.

 

Point 5: On lines 146-147, “other crops” are mentioned, and the sample in Fig. 4 also mentions “others”, but there are no “other crops” in the final crop classification results. I’d like to know whether “other crops” are involved in the classification process,  and how are they distinguished from the four types of crops in the manuscript?

Response 5: Thank you for the concern sincerely. In this study, “other crops” refer to staple crops (wheat, corn, and peanut in this study area). These categories also participate in the classification, but different from the main orchard crops (apples, cherries), they are classified as "other" label in the classification and used as negative samples. In the results of this study, these staple crops are not mapped mainly caused by two reasons.

• These crops are not the main crops in Qixia City. According to the 2020 statistical data of Qixia City, in 2019, the grain crop output was 77200 tons, while the fruit output was 2029000 tons which is 26 times of the grain output [1]. Therefore, compared with orchard crops, the planting area of staple crops is very small.
• Unable to obtain accurate sample data. In the field investigation in 2021, samples of staple crop in 2019 cannot be obtained. Unlike garden crops, the growth cycle of staple crops is usually less than one year, so it is impossible to obtain the samples in 2019. On the high resolution image, it is easy to distinguish a parcel from an orchard parcel or a staple crop parcel by texture features, but it is impossible to judge whether it is corn, peanut or other crops by visual interpretation.

Therefore, “other crops” are not taken as mapping objects in our study.

 

Point 6: The overall accuracy of the “irregular-time-series +2DCNN” method is only 1.83% higher than that of the “regular-544 time-series +2DCNN” method, and the kappa value of the former is only 0.0327 higher than that of the latter. Poor accuracy raises questions about the usefulness of using images with irregular time series.

Response 6: Thank you for your question. Your guidance will be of great significance to our future research. For the problems that are not obvious in the accuracy improvement of the method in this paper, we make the following explanations.

(1) Although the length of the irregular time series is much longer than the regular time series, the main missing data in the regular time series are in July and August. In other months, the regular time series contains two or more images, which is sufficient to describe crop growth differences. In these months, although irregular time series can provide more effective observations, it will not cause strong differences in classification results.

(2) In the study area, we mainly classify four types of land surface, namely apple, cherry, seedling and greenhouse. From the observation of time series, we think that the characteristics of seedling and greenhouse are obvious, and their distinguishability is higher. Even if the data in July and August are missing, the difference in the regular time series is enough to distinguish them from other crops. In fact, apples and cherries are more difficult to distinguish. The two orchards are deciduous trees, both belong to Rosaceae, and their phenological characteristics are very similar. The classification accuracy of apple orchard and cherry orchard based on regular time series is 98% and 89% respectively. However, the proportion of these two crops is very uneven. According to the final statistics, the proportion of apple orchard is the largest, 14 times that of cherry orchard. In the regular time series, the classification accuracy of apples has reached 98%. In this case, even if the irregular time series can better distinguish apple orchards and cherry orchards, due to the small area of cherry orchards, more recognition of cherry orchards will not significantly improve the overall accuracy.

To sum up, we believe that although the kappa coefficient of the experimental classification results of "irregular time series+2DCNN" is less improved than that of the regular time series, this does not mean that the irregular time series is useless. The potential of irregular sequences in crop classification has also been demonstrated by many previous studies.

 

Here are some minor comments:

Point 7: On line 129, an error was reported.

Response 7: Thanks for your suggestions sincerely. Sorry for this error in the manuscript and we have revised the manuscript.

 

Point 8: Line 218 is duplicated with line 157, the former should be 2.3 instead of 2.2.

Response 8: Thank you for your suggestion. We have corrected this error in the manuscript.

 

Point 9: Figures should be numbered in the order in which they are mentioned in the manuscript, thus, Fig. 4 (mentioned at line 192) should come before Fig. 3 (mentioned at line 205).

Response 9: Thank you for your suggestion sincerely. We revised the manuscript to ensure that the figures are in order.

 

Point 10: I think Fig. 3 mentioned on lines 449 and 455 should be Fig. 5.

Response 10: Thank you for your suggestion. We carefully checked the manuscript. In fact, the double-layer structure mentioned in the manuscript is shown in Figure 3. We have revised the relevant content to make it easier to understand.

 

Point 11: Please indicate where the three sub-regions (a1), (b1) and (c1) in Fig. 11 are located in the study area.

Response 11: Thank you for your suggestion. In Figure 9 of the manuscript, we added the geographical locations of three sub regions (a1), (b1) and (c1) in Fig. 11.

 

Point 12: On line 513, how did the 10963 real values come from, please add an explanation.

Response 12: Thank you for your suggestion. Based on the field survey samples and VHR image, the point features were expanded into the polygon features. Then, we randomly selected some pixels from the polygon as samples. Some of them were used for training the classification model, and the other samples (10963 pixels) were used for the accuracy verification of the classification results. We also added the origin of the verification data in the manuscript.

 

Point 13: On line 550, “Figure 12” appears twice.

Response 13: Thank you for your suggestion. We are sorry for this mistake. We have corrected this error in the manuscript.

 

 

Point 14: The word of parcel and plot have different meanings in the land survey. The usage needs to be standardized throughout the paper. Similarly, the use of nurseries and seedling needs to be standardized.

Response 14: Thank you for your suggestion sincerely. According to your suggestion, we have revised the manuscript.

 

Point 15: The 39th reference format is not standardized.

Response 15: Thank you for your suggestion. We are sorry for this mistake. We have revised the format of references, thank you again.

 

References

[1] The People's Government of Qixia. Available online: http://www.sdqixia.gov.cn/art/2020/10/29/art_31425_2918296.html (accessed on 29 November 2022).

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, the authors map horticultural crops based on extracted parcel-level plot and irregular time series classifications using Qixia City, Shandong Province, China as an example. And the accuracy of regular time series and irregular time series classification is compared. Overall, this paper has problems in different parts.

1、Is the study area the whole Qixia City or a certain area of Qixia City? figure1 is not clear. the figures' quality is unacceptable.

2、Fig2 The figure illustrations are too brief. What is the difference between "growing (Red)" and "growing period (Green)"? The starting point and direction of DOY are not marked.

3、Why is the image acquisition time (2019) inconsistent with the sample acquisition time(2021)?

4、How is a regular time series constructed? The description 2.2.1 is not clear.

5、Is the sample database used points or polygon?

6、Section2.2.2 The basis of grid division? Are there no references?

7、What is the innovation of DCNN model? Is it different from what Diyou Liu proposed?

8、What is the sample distribution of each grid? How does that affect the classification results?

9、section3.4 accuracy evaluation PA and UA are introduced, but these two indexes are not used in the later results. Why? It is suggested to analyze UA and OA in the results section.

10、Figure10 where is C? Box out the location of ABC.

11、section 4.1 line502 and 503 What are the advantages over traditional methods? What are the other traditional methods? How effective are they? The article doesn't show that.

12、section 4.2 Pixel-based mapping only has precision comparison, but no mapping results? If so, can it be compared in terms of mapping results?

13、section4.3 Classification based on parcel-level has a good effect. Is there a comparison with the mapping based on pixel-level?

14、section 4.3 Are there mixed pixels present in other types such as greenhouse, cherry orchards and so on? Which type, is more likely to produce mixed pixels with apples in the same parcel?

15、There are also many chapter labeling errors, language issues, abbreviated image annotation issues, reference citation issues, and other detailed sections that need attention.

Generally, the language should be revised and improved, please pay attention to the details of the manuscript. See all my comments in the attached PDF file.

Comments for author File: Comments.pdf

Author Response

Dear reviewer,

 

Thank you for your approval of our manuscript, and we are glad to accept your guidance on our manuscript. We have carefully reviewed the comments and have revised the manuscript accordingly.

 

Our responses are given in a point-by-point manner below. Changes to the manuscript are highlighted using the "Track Changes" function in Microsoft Word.

 

The specific replies are as follows:

 

Point 1: Is the study area the whole Qixia City or a certain area of Qixia City? (L126)

Response 1: Thank you for your suggestion. The study area is the whole Qixia City.

 

Point 2: What is the difference between "growing (RDE)" and "growing period(GREEN)"? The starting point and direction of DOY are not marked.(L152)

Response 2: Thank you for your question. "Growing (RDE)" refers to the growth period of staple crops. "Growing period (GREEN)" refers to the growing period of orchard crops. In addition, according to your suggestion, we marked the starting point and direction of DOY in the figure.

 

 

Point 3: Why is the image acquisition time (2019) inconsistent with the sample acquisition time(2021)?

Response 3: Thank you for your question. The reason why we chose 2019 to carry out the experiment is mainly because in 2019, there are more missing data in regular time series, which can better reflect the advantages of irregular time series in classification. The mapping of crop distribution in 2019 with field survey data in 2021 may be inconsistent with the actual category. In this regard, we have taken some measures in the field sampling to ensure the authenticity of the classified samples in 2019.

• Judge whether the fruit tree can blossom and bear fruit. Apple trees generally begin to bear fruit after three or more years of growth, and the flowering phase is from late April to early May. The field survey was carried out in the early and middle days of May 2021 (5.1-5.12), in the flowering period of apple trees. We mainly collect samples of adult flowering apple orchards to ensure that the samples also exist in 2019. Cherry trees generally bear fruit after 3-6 years of planting and gradually matured at the end of April and May. During the field survey, it is the maturation stage of the big cherry. Therefore, it can be judged whether the fruit tree has been planted for more than three years by judging whether there are fruits on the tree.
• High resolution images are used as reference data. When collecting samples, we compared the high-resolution image data in 2019 with the field crops. Objects with unique texture features on high-resolution images, such as greenhouses and seedling orchard, can be easily distinguished by visual interpretation, as shown in the following figure. Using high resolution images as auxiliary data during field investigation can ensure that the collected greenhouse samples and seedling samples also exist in 2019.
• The samples collected are mainly located in the interior of the large parcels, and the sample points are placed in the middle of the parcel to avoid the wrong sample category caused by positioning errors.

 

Point 4: How is a regular time series constructed? The description 2.2.1 is not clear.

Response 4: Regular time series require the same length of time series at different spatial locations in the study area. Therefor, when the regular time-series were built, the whole study area is taken as the screening range, and the images with cloud cover less than 20% are selected. A total of 29 high-quality sentinel-2 data covering the study area are obtained. Calculate the vegetation index as the characteristic, and then stack them in chronological order. For each pixel, there is a one-dimensional sequence composed of vegetation indices at different times.

 

Point 5: Is the sample database used points or polygon?

Response 5: Thank you for your question. I'm sorry we didn't make this clear in the manuscript. The original sample data is point data from field investigation. In order to obtain enough sample data for model training, we expand the point data into polygons based on VHR images. Then, from the inside of the polygon, randomly select pixels as sample data. Through this process, the number of training samples can be enriched. We also revised this part in the manuscript to make it easier for readers to understand.

 

 

 

Point 6: Section2.2.2 The basis of grid division? Are there no references?

Response 6: Thank you for your question. Sorry, there are no references cited here. The grid division has no special purpose here and is only used to build the basic unit of irregular time series. When we divide the grid, the grid scope is determined according to the scope of the study area.

 

Point 7: What is the innovation of DCNN model? Is it different from what Diyou Liu proposed?

Response 7: Thank you for your question. We are sorry that there is no innovation in the network structure of the 2DCNN model, which is also a direction of our future research. The difference with Liu Diyou's proposal is that three indexes are used as three channels in the study to generate images that can be used by 2DCNN.

 

Point 8: What is the sample distribution of each grid? How does that affect the classification results?

Response 8: Thank you for your question. Theoretically, every grid should contain all classification categories. In fact, when expanding the sample data, we also practice this in most grids. However, in some grids, there are no greenhouses and fast nursery areas, so these two kinds of samples cannot be included. For grids with incomplete samples, the classification accuracy will be lower than that of grids with abundant samples.

 

Point 9: section3.4 accuracy evaluation PA and UA are introduced, but these two indexes are not used in the later results. Why? It is suggested to analyze UA and OA in the results section.

Response 9: Thank you very much for your suggestion. In fact, we use the overall accuracy and kappa coefficient to reflect the accuracy and reliability of the classification results. Therefore, we modified the description of precision evaluation indicators and deleted the reference to UA and OA. Thank you again for your suggestions.

 

Point 10: Figure10 where is C? Box out the location of ABC.

Response 10: Here is a mistake. We are sorry about that. We have corrected this error.

 

 

Point 11: section 4.1 line502 and 503 What are the advantages over traditional methods? What are the other traditional methods? How effective are they? The article doesn't show that.

Response 11: Thank you for your question. The algorithm used in this paper is deep-learning algorithm. Compared with the traditional algorithm, the effect of parcel extraction based on deep learning algorithm is more consistent with the actual parcel distribution. The deep learning algorithm learns the overall characteristics, and the texture in the seedling parcels will not be mistaken as the edge of the parcel. We have revised the content of this part. In addition, the advantages of this method are also discussed in the introduction

 

Point 12: section 4.2 Pixel-based mapping only has precision comparison, but no mapping results? If so, can it be compared in terms of mapping results?

Response 12: Thank you for your question. We obtained pixel-level mapping results in our experimental process, but considering that the research focus is parcel-level crop mapping, and pixel-level mapping result is only a part of the parcel-level mapping process. Therefore, parcel-level mapping results were not compared with pixel-level results.

 

Point 13: section4.3 Classification based on parcel-level has a good effect. Is there a comparison with the mapping based on pixel-level?

Response 13: Thank you for your question. Unfortunately, the study focused on parcel-level crop mapping. The mapping results were not compared with pixel-level classification results. This will be the focus of our follow-up research.

 

Point 14: section 4.3 Are there mixed pixels present in other types such as greenhouse, cherry orchards and so on? Which type, is more likely to produce mixed pixels with apples in the same parcel?

Response 14: Thank you for your question. We have explained in the manuscript that because the edge and texture features of the greenhouse and seedling on the VHR are very strong, they have a good extraction effect. The research regards them as pure parcels, and there is no mixed planting. In fact, during the field survey, we found that the main mixed planting situation in the study area is the mixed planting of apples and cherries.

 

Point 15: There are also many chapter labeling errors, language issues, abbreviated image annotation issues, reference citation issues, and other detailed sections that need attention.

Response 15: Thank you for your suggestion, and we carefully checked the manuscript again. The errors are revised.

 

Point 16: Generally, the language should be revised and improved, please pay attention to the details of the manuscript. See all my comments in the attached PDF file.

Response 16: Thank you for your suggestion. We have edited the language of our manuscript by the English editing service provided by MDPI, and have modified all the comments in the PDF attached.

 

Here is the modification of the annotation comments in the attachment PDF

Point 1: What does that mean? (L129)

Response 1: Thank you for your suggestion sincerely. We have corrected this error.

 

Point 2: Where is the point? Mark the figure a b.c Explain the figures one by one. (L138)

Response 2: Thank you for your suggestion. We have marked sample points in the figure and explained different figures one by one.

 

Point 3: The references are not standard, and the language needs to be improved. (L148)

Response 3: Thank you for your suggestions. We have revised the format of the references and improved the English of the whole manuscript.

 

 

Point 4: ???(L175)

Response 4: I'm sorry that our statement is not clear, which makes it difficult for you to read. We have revised the relevant description.

 

 

Point 5: Is the image distribution irregular why is it a regular time series? How is a regular time series constructed? (L178)

Response 5: Thank you for your question. I'm sorry that we didn't fully express our ideas in the manuscript. “Irregular”, what I want to express here is that when filtering the regular time series, it is found that the date interval between images is uneven. For example, there is only one image in June, but there are five images in October. It does not mean that the time sequence characteristics constructed between each pixel are different in length.

Point 6: Reference? (L196)

Response 6: Thank you for your suggestion. We added references to the manuscript.

 

Point 7: Is the section marked incorrectly? (L218)

Response 7: Thank you for your suggestion. I'm sorry for this mistake in the manuscript and we have revised the manuscript.

 

Point 8: Point or polygon? (L219)

Response 8: Thank you for your question. I'm sorry we didn't make this clear in the manuscript. The original sample data is point data from field investigation. In order to obtain enough sample data for model training, we expand the point data into polygons based on VHR images. Then, from the inside of the polygon, randomly select pixels as sample data. Through this process, the number of training samples can be enriched. We also revised this part in the manuscript to make it easier for readers to understand.

 

Point 9: The figures' quality is unacceptable, need to improve the resolution. (L234)

Response 9: Thank you for your suggestion. We updated the picture. The resolution is changed to 300 dpi.

 

Point 10: Language improvement. (L266)

Response 10: Thank you for your suggestion. We have improved the English in the whole manuscript.

Point 11: References？ (L293)

Response 11: Thank you for your suggestion. We have added relevant references here.

 

Point 12: pay attention to details! (L347)

Response 12: Thank you for your suggestion. We have corrected this error and we are so sorry for our mistakes.

 

Point 13: Is it still 3.2.2? (L352)

Response 13: Thank you for your suggestion. We have corrected this error.

 

Point 14: Where is C? (L490)

Response 14: Thank you for your suggestion. We have corrected this error.

 

Point 15: What are the advantages over traditional methods? (L502)

Response 15: Thank you for your question. Compared with the traditional algorithm, the effect of parcel extraction based on deep learning algorithm is more consistent with the actual parcel distribution. The deep learning algorithm learns the overall characteristics, and the texture in the seedling parcels will not be mistaken as the edge of the parcel.

 

Point 16: pay attention to the detail. (L521)

Response 16: Thank you for your suggestion. We have corrected this error.

 

Point 17: pay attention to the details. (L550)

Response 17: Thank you for your suggestion. We have corrected this error.

Author Response File: Author Response.pdf

Reviewer 4 Report

Please see the attached file...

Comments for author File: Comments.pdf

Author Response

Dear reviewer,

 

Thank you for your approval of our manuscript, and we are glad to accept your guidance on our manuscript. We have carefully reviewed the comments and have revised the manuscript accordingly.

 

Our responses are given in a point-by-point manner below. Changes to the manuscript are highlighted using the "Track Changes" function in Microsoft Word.

 

The specific replies are as follows:

 

Point by point response to the comments of the reviewer4:

Point 1: The main objective of the research needs to be clearly included in the abstract. The readers do not need to scan the whole abstract to see the motivation. Please update this part.

Response 1: Thank you very much for your suggestions. We have revised the abstract and highlighted the research objective in the manuscript.

 

Point 2: The abstract is too long, no need to provide data processing details. The data, objective, name of the methods and outcomes have to be included in the abstract.

Response 2: Thank you for your advice. We deleted some useless information in the abstract and reorganized its logic so that readers can fully understand the research content of this paper through the abstract. The modified abstract is shown in the following Figure.

 

Point 3: Data spesifications of Sentinel-2, along with the details of all types of resolution (i.e. spectral, radiometric incl.) has to be provided with a table. Not everyone should remember the such details.

Response 3: Thank you sincerely for your suggestions We have enriched the introduction of data in the manuscript.

 

 

Point 4: Data pre-processing steps were not provided and should absolutely included in the paper. This is important and critical because you are using time-series and generating the vegetation indices. The atmospheric condition is not always same for the different dates and might influence the pixel brightness values. Such details needs to be included.

Response 4: Your good advice was very much appreciated. The optical data used in the study was Sentinel-2 Level -2A and can be easily downloaded from GEE. The data has been radiometric calibrated and atmospheric corrected. The downloaded image needs to be re projected. UTM projection is used. UTM projection adopts 6-degree zoning. The study area is located in the Northern Hemisphere, with the belt number of 51. Therefore, the coordinate system of the re projection is “WGS_1984_UTM_Zone_51N”, as shown in the following Figure. We also added details of data processing in the manuscript.

 

Point 5: What level of Sentinel-2 data you have used? Level-1 or Level-2? Depending on your choose, the pre-processing level can vary and the algorithm that you us efor atmospheric and radiometric correction might change. Please double check this and include these critical details. Otherwise, the pixel values do not represent “correct” values therefore the the experimental results mislead the readers. It’s VERY CRITICAL!

Response 5: Thank you for your question. Images used in our study are Level -2A which have been radiometric calibrated and atmospheric corrected. We also explained this problem in the revised manuscript.

 

Point 6: Why did you not include red-edge NDVI? It’s the ultimate advantage of Sentinel-2 satellite. Please explain the reason of not including it.

Response 6: Thank you for your suggestion sincerely. I'm sorry that the study did not take into account the advantages of the red-edge in vegetation recognition. The purpose of this study is to explore the utilization of fragmented optical data by irregular time series. Sentinel-2 was selected to construct crop phenological characteristics because of its short revisit cycle, which can provide sufficient data support for the experiment. Therefore, only the most common red band is selected for the study, and the red-edge band is not used. The red-edge band is a sensitive band indicating the growth status of green plants and has a good correlation with the important biochemical parameters representing the growth status of plants. It can effectively monitor the growth of vegetation and is widely used in vegetation identification and health monitoring and other aspects. In the future research, we will add the red-edge band to the vegetation recognition to obtain more accurate mapping results.

 

Point 7: Please check the following papers.

Forkuor, G., Dimobe, K., Serme, I., & Tondoh, J. E. (2018). Landsat-8 vs. Sentinel-2: examining the added value of sentinel-2’s red-edge bands to land-use and land-cover mapping in Burkina Faso. GIScience & remote sensing, 55(3), 331-354.

Ustuner, M., Sanli, F. B., Abdikan, S., Esetlili, M. T., and Kurucu, Y.: Crop Type Classification Using Vegetation Indices of RapidEye Imagery, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL-7, 195–198, https://doi.org/10.5194/isprsarchives-XL-7-195-2014, 2014.

Sun, Y., Qin, Q., Ren, H., Zhang, T., & Chen, S. (2019). Red-edge band vegetation indices for leaf area index estimation from sentinel-2/msi imagery. IEEE Transactions on Geoscience and Remote Sensing, 58(2), 826-840.

Response 7: Thank you for recommending these articles. We read them carefully. The red edge is very effective in land use, vegetation monitoring, etc. Compared with the traditional band, the classification algorithm can obtain higher classification accuracy after including the red-edge. We are very willing to use the red edge in subsequent studies to further improve the classification accuracy of crop classification.

 

Point 8: Also please see the following papers for addressing the correct references for vegetation indices used in remote sensing, I can not see reference papers for NDVI or other indices that you have used?

Hadjimitsis, D., et al. "Atmospheric Correction for Satellite Remotely Sensed Data Intended for Agricultural Applications: Impact on Vegetation Indices." Natural Hazards Earth System Science 10 (2010): 89-95. 

Jackson, R., and A. Huete. "Interpreting Vegetation Indices." Preventive Veterinary Medicine 11 (1991): 185-200.

Price, J. "Calibration of Satellite Radiometers and the Comparison of Vegetation Indices." Remote Sensing of Environment 21 (1987): 15-27.

Response 8: Thank you very much for your advice. We added references to the vegetation indexes in the revised manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Dear authors,

Thank you for addressing all my comments. I understand your viewpoint and agree with the current version of the article. I would recommend it for publication in its present form.

Author Response

Dear reviewer,

 

It's my honor to receive your approval of our manuscript. Your suggestions played an important role in the manuscript revision process. Thank you again for your suggestions on our manuscript. Regarding the question "English language and style are fine/minor spell check required", I'm sorry that our mother tongue is not English, which makes it more difficult to read the manuscript. In fact, we have revised the language of the manuscript using the English editing services provided by MDPI. Furthermore, there will be another final English editing provided by the in house editors who will check your paper before proofread. I hope this will address your concerns. If you have any other questions about our manuscript, you are welcome to put forward your valuable suggestions. We will try our best to solve the problem so as to improve our manuscript.

 

Yours sincerely,

Shuhui Jiao

Reviewer 2 Report

Although the author has revised the manuscript, there are still some minor errors. I think the author should proofread the whole text carefully. Here are a few comments:

1. The F1 score and IoU values in lines 23 to 24 need to be specified.

2. The experimental classification results of "irregular time series+2DCNN" is less improved than that of the regular time series. However, the author believes that the irregular time series is useful. Please specifically point out its usefulness in the manuscript.

3. Please supplement the manuscript with a description of the authenticity of the ground samples, which is related to the reliability of the experimental results.

4. I maintain that the reference to Figure 4 in lines 473 and 479 is incorrect. In the revision, Figure 4 depicts irregular time series data information rather than method flow, and it is clear that lines 473 through 479 have been describing the classification implementation steps.

Author Response

Dear reviewer,

 

Thank you very much for your suggestions, which will help us to further improve the manuscript. We have carefully reviewed the comments and have revised the manuscript accordingly.

 

Our responses are given in a point-by-point manner below. Changes to the manuscript are highlighted using the "Track Changes" function in Microsoft Word.

 

Point by point response to the comments:

Point 1: The F1 score and IoU values in lines 23 to 24 need to be specified.

Response 1: Thank you very much for your suggestion. We have revised the manuscript and hope it can solve your concerns.

Point 2: The experimental classification results of "irregular time series+2DCNN" is less improved than that of the regular time series. However, the author believes that the irregular time series is useful. Please specifically point out its usefulness in the manuscript.

Response 2: Sincere thanks for your suggestion. The main advantage of irregular time series is that more fragmented optical data can be used to construct crop growth curves. For some areas with poor climatic conditions or when the scale of the research area is relatively large, it is unavoidable that the time series length and data date of pixels at different locations are difficult to keep consistent. This results in less data available in regular time series. Irregular time series can solve this problem well. Based on your suggestion, we have added the advantage of irregular time series to the manuscript, which we hope will address your concerns.

 

Point 3: Please supplement the manuscript with a description of the authenticity of the ground samples, which is related to the reliability of the experimental results.

Response 3: Thank you very much for your suggestion. In fact, we have given relevant information about ground samples, training samples and verification samples in the last version of the manuscript. In Figure 1, we added the categories and distribution of ground samples.

Point 4: I maintain that the reference to Figure 4 in lines 473 and 479 is incorrect. In the revision, Figure 4 depicts irregular time series data information rather than method flow, and it is clear that lines 473 through 479 have been describing the classification implementation steps.

Response 4: Thank you very much for your advice. We carefully examined the manuscript and revised the reference to the picture. What I want to express here is that the specific flow chart of the parcel filling strategy is shown in Part 3 of the overall flow chart (Figure 5). Thank you again for your suggestions.

In addition, in English language and style, you think that "Modify English changes required". In response to this problem, we edited the manuscript with the English editing service provided by MDPI. Furthermore, there will be another final English editing provided by the in-house editors who will check your paper before proofreading. I hope this will address your concerns.

Author Response File: Author Response.pdf

Reviewer 4 Report

The authors replied all my concerns and revised the manuscript accordingly. The revised manuscript can be accepted in its current form. 

Author Response

Dear reviewer,

 

It's my honor to receive your approval of our manuscript. Your suggestions played an important role in the manuscript revision process. Thank you again for your suggestions on our manuscript.

 

Yours sincerely,

Shuhui Jiao