Using Time Series Sentinel-1 Images for Object-Oriented Crop Classification in Google Earth Engine

Round 1

Reviewer 1 Report

Using time series Sentinel-1 images for object-oriented crop classification in Google Earth Engine

This revised manuscript utilizes Sentinel-1 (S1) time-series imagery (composites of 10, 15 and 30 days) for crop classification in Heilongjiang Province, China. Furthermore, pixel-based versus object-based classifications were compared in Google Earth Engine (GEE), as well as different sizes of the simple noniterative clustering (SNIC) method.

The topic of this study is acutal, however, still the manuscript needs some extensive editing of English language. My major concerns are, as follows:

• First of all, English must be extensivly edited and changes in reporting style are required
• Still terms about optical/SAR imagery are confusing. Single temporal should be replaced with single-date or monotemporal imagery (refer to Gasparovic and Dobrinic, 2020). In LN 59 single polarization and multi-polarization SAR dana is mentioned. Does it refer to co- and cross-polarized dana?
• LN 48 and 49: „which have high resolutions“. What resolution – spatial, spectral, temporal?
• LN 52: „put into use“ does not fit in this context
• LN 61 and 63: adjust the term: based on pixel classification or pixel-based classification
• LN 89: „different treatments“, the Authors had more classification scenarios in this manuscript, not treatments
• LN 129-131: pre-processing of the SAR imagery is very poorly described, and as the reference is mentioned Google? Gasparovic and Dobrinic (2020) describe these steps in detail, and LN 129 needs to be modified since terrain-corrected values are not converted to log scale, it is backscatter coefficient
• LN 133 and 134: depending on filter used, every speckle filter has it's purpose (to adjust to local image variations to smooth the values and thereby reduce speckle, and lines and edges are enhanced to maintain the sharpness of the imagery), so this lines need to be rewritten
• Furthermore, LN 138 and 139 also need to be rewritten, since interaction of SAR signal polarized in vertical or horizontal plane acts differently on objects of interest, it is not adjusted only for this study
• LN 190 – 197: some more details about used MDA for variable importance
• Section 3.3: please analyze more in detail PA and UA for different crops/study area
• Figures 3-10 please move sub-captions of the figure (a,b,c…) to the bottom of the figure, it is written with different fonts and placed on different sides of the figures, so make it uniform

I think that the research design of this manuscript is good, but still some additional changes need to be made, as well as changes in English language and style.

REFERENCES

Gašparović, M.; Dobrinić, D. Comparative Assessment of Machine Learning Methods for Urban Vegetation Mapping Using Multitemporal Sentinel-1 Imagery. Remote Sens. 2020, 12, 1952.

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 2 Report

Thanks for update. I have some additional comments.

Please in lines 71 add references. Also add reference in line 73 about object oriented classification.

SNIC algorithm has to be described fully. There is no description and also no reference.

Also Random Forest is described without necessary details.

Results demonstrate in figures 11 and 12 doesn’t demonstrate real progress in comparison with other methods

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 3 Report

This is an interesting study about "Using time series Sentinel-1 images for object-oriented crop classification in Google Earth Engine". The manuscript is well written with an adequate structure as a scientific paper demands. However, I have suggested a few modifications that can result in substantial improvement to this manuscript.

The Abstract "usually, it is one paragraph of 300 words or less, the major aspects of the entire paper in a prescribed sequence that includes: 1) the overall purpose of the study and the research problem(s) you investigated; 2) the basic design of the study; 3) major findings or trends found as a result of your analysis; and, 4) a brief summary of your interpretations and conclusions."
Furthermore, finding results (values) must be added to the abstract. Also, the objectives should be presented briefly and concisely. "Well defined research objectives will help the readers to clearly understand the study".

"In the "Introduction" section, this section can be improved to provide further background and include all relevant references. There is a need that you will use "recent publications" on the topic to make your research attractive. In this section, speaking about the development of modern technology and the advantage of using Remote Sensing (RS) technology, the authors should provide several references to substantiate the claim made in this section (provide references to other groups who have done research on this topic) to make the introduction more substantial for example GEE studies: 1. https://t.co/HgTHmILrN0?amp=1 2. https://ieeexplore.ieee.org/abstract/document/8009768  etc...

In the "Materials and Methods" section,
In the "Study area" section,
Figure 1. The map is too simplistic. The authors should add which coordinate system (projection, ellipsoid, and datum) is used here (see the example below). See here https://rb.gy/ww6kuu
Please include all the spatial reference properties see the example below:
Projected Coordinate System: WGS_1984_UTM_Zone_19N
Projection: Transverse Mercator
Linear Unit: Meter
Geographic Coordinate System: GCS_WGS_1984
Datum: D_WGS_1984
Prime Meridian: Greenwich
Angular Unit: Degree

Also, I strongly recommend that the author devise a flowchart that depicts the steps that they have processed in this study.

In the "Results and discussion" section, the author must extend the comparison between their approach and other ones that have been developed and used in the literature for the same or related purposes (I recommend increasing the number of Scientific articles cited, especially to compare the study context with similar studies).   Also, in this section, the authors should also highlight the current limitations and usefulness of the proposed research, and briefly mention some precise directions that they intend to follow in their future research work. speaking about the "Advantages of using GEE" Please conderd to read this article: https://t.co/HgTHmILrN0?amp=1  

About the Figures: 1. Figure 1, Please remove the label axes of the left right and the bottom top sides of the coordinate system and change the left label orientation to vertical. 2. Please increase the text/numbers in figures 1, 7, 8, 9, and 10.   

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 4 Report

According to the authors, the main objectives of this work were: 1) to evaluate the impact of different temporal resolutions and different segmentation size object-oriented methods on the accuracy of Sentinel-1 crop classification; 2) to study the key period of using Sentinel-1 images for crop classification; and 3) to compare the classification accuracy of two study areas with different plot sizes and evaluate the applicability of this method. The subject of the paper using the Sentinel-1 SAR images for crop analysis is very important. However it presents some concepts that are not clear for the research community. For example, the temporal resolution is specific for the sensors. Then the way the authors are dealing with it is not appropriated. It is not possible to change the temporal resolution of the sensor. But it is feasible to define time periods and take the average of the images available in that period and defined as in this case, 10d, 15d, monthly, etc. periods and not temporal resolutions. The authors should review this concept to make the paper clearer for the readers.

 

Some corrections and suggestions:

L.17 - The research time was two adjacent years (2018 and 2019),  ---  two consecutive years …

L.19- were composited with three temporal resolutions (10 d, 15 d and 30 d).  ---  were composited with three time periods (10 d, 15 d and 30 d).

L.26- the higher the temporal resolution of the composite  ---  the shorter time period of the composite

L.27- when the temporal resolution increased  ---  when the time period decreased

L.29- with different temporal resolutions  ---  with different time periods

L.34- temporal resolution  ---  time period

L.48- MODIS, which has a medium resolution, and Sentinel-2 and Landsat-8, which have high resolutions --- MODIS, which has a moderate resolution, and Sentinel-2 and Landsat-8, which have medium resolutions

L.63- Some scholars have proven  --- Some researchers have proven

L.70-few researchers has  --- few researchers have

L.80- can process millions of images quickly in  --can process several images quickly in

L.82- has good temporal and spatial resolution  --- has medium temporal and spatial resolutions

L.84- To study the potential of using object-oriented and time series Sentinel-1 in crop classification in GEE, two study areas with large differences in plot size were selected in this study. First, the Sentinel-1 images with different temporal resolutions were obtained by GEE processing, the images were segmented by the simple noniterative clustering (SNIC) image segmentation method, and then the processed images were classified using a random forest classifier. Finally, the classification accuracies of different treatments were compared to evaluate the effectiveness and applicability of the methods.  – This part is more related to the Materials and methods section.

L.98- Keshan farm (125°07′40″ - 125°37′30″ E, 48°11′15″ - 48°24′07″ N) and Tongnan town (124°54′15″ - 125°12′44″ E, 48°2′40″ - 48°15′13″ N) --- Keshan farm (125°07′40″, 125°37′30″ E, 48°11′15″, 48°24′07″ N) and Tongnan town (124°54′15″, 125°12′44″ E, 48°2′40″, 48°15′13″ N)

L.100 - the area is  --- the areas are

L.117- In order to reduce data redundancy, we will take --- Then we will take

L.120- Study area main crop calendar  --- Main crop calendar of the study area

L.126 - including 27 images in 2018 and 22 images in 2019.  --- The number of images refer to the May to September period ? I suggest to mention the temporal resolution of Sentinel-1 data.

L.140- To evaluate the impact of Sentinel-1 images with different temporal resolutions on ----– different time periods composite

L.145, 146 – 483 (2018) and 486 (2019) sample plots are different as described in the Table 2 --- Which are correct ?

L.150- randomly selected 30% of the sample points as the training samples and 70% of the sample points as the verification samples. --- These proportions are correct ? In general is the other way.

L.163- nearly uniform superpixels  --- nearly uniform polygons

L.166- Sentinel-1 time series images with different temporal resolutions --- Sentinel-1 multitemporal images with different time periods composites

L.173- "Seeds" does not need to be set in this article.?  --- It needs to be explained.

L.174- we directly use the accuracy of the classification results to evaluate the classification effect.  ?

L.177- three temporal resolutions  ?  ---  three time period composites

L.212- Figure 2 shows the time series curves  ---   multitemporal images curves.

L.213- is the time series of images at different temporal resolutions,  --- is the multitemporal images at different time period composites

L.215- temporal resolution of 30 d  ?

L.217- temporal resolution of 15 d ?

L.221- temporal resolution of 10 d ?

L.229- different temporal resolution times (A, VH_30d; B, VV_30d; C, VH_15d; D, VV_15d; E, VH_10d; F, VV_10d).  - ?

L.242, L.245- the higher the temporal resolution of Sentinel-1 ? The temporal resolution of Sentinel-1 is 12 days, right ? 6 days with Sentinel-1A and 1B together

L.247- When the temporal resolution increased from 15 d to 10 d  ?

L.277- Figure 7&8 shows  --- Figures 7&8 show

L.297-there are 2 rice sample sites  ---there are only 2 rice sample sites  

L.302- Sentinel-1 images with different temporal resolutions  ?

L.304 - Sentinel-1 images with 10-d temporal resolution  ?

L.307- temporal resolution of 15 d or 30 d  ?

L.320 - (A, MeanDecreaseAccuracy_2018_10d; B, MeanDecreaseAccuracy_2019_10d; C, MeanDecreaseAccuracy_2018_15d; D, MeanDecreaseAccuracy_2019_15d; E, MeanDecreaseAccuracy_2018_30d; F, MeanDecreaseAccuracy_2019_30d).  ?

L.328- temporal resolution images (A, MeanDecreaseAccuracy_2018_10d; B, MeanDecreaseAccuracy_2019_10d; C, MeanDecreaseAccuracy_2018_15d; D, MeanDecreaseAccuracy_2019_15d; E, MeanDecreaseAccuracy_2018_30d; F, MeanDecreaseAccuracy_2019_30d).

L.353-The theoretical revisit period of Sentinel-1 data used in this study is 3 d, ? 3d is in the Equator.

L.362-the time series images  ---  the multitemporal images

L.369- used refined Lee speckles  --- used refined Lee filter

L.372- all other processes were processed -- all other processes were performed

L.398- two adjacent Years  -- two consecutive years

L.399- The highest OA of object-oriented classification was approximately 95%, which was 10% higher than that of pixel-based classification. However, in Tongnan town, which has a small plot area, the highest OA of the object-oriented classification method was only approximately 2% higher than that of the pixel-based classification method.  – These comparisons are not supported by this work.

L.418- with the time series Sentinel-1 image  ---  with the multitemporal Sentinel-1 images

L.422- increase in temporal resolution from 15 d to 10 d,  ---  decrease in time period composite from 15 d to 10 d

L.430- the temporal resolution of SAR data for crop classification. ---  the time period composite of SAR data for crop classification

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Round 2

Reviewer 1 Report

The authors provide significant answers and updates from v1 to v2.

Therefore, I think it meets the criteria for publication.

Author Response

We thank the reviewer for this positive assessment of our manuscript.

Reviewer 2 Report

Dear authors, thanks for update. This update made many things clearer, nevertheless I have two comments:

1. Please add some explanation of Figure 1. Many things is later mentioned in text, but it is not easy to see relation to single part of text. It is also possibly understood, where SNIC algorithm is applied, but it is not explicitly in image.
2. This object oriented segmentation reduce soil pepper effect, nevertheless, if I am looking on figure 13, there is question, if such results are useful for practical purposes. Can you suggest, what could be next step to improve final output.

I understand, that you done good research, but this two topic probably need some explanation.

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 4 Report

According to the authors, the main objectives of this work were: 1) to evaluate the impact of different temporal resolutions and different segmentation size object-oriented methods on the accuracy of Sentinel-1 crop classification; 2) to study the key period of using Sentinel-1 images for crop classification; and 3) to compare the classification accuracy of two study areas with different plot sizes and evaluate the applicability of this method. The subject of the paper using the Sentinel-1 SAR images for crop analysis is very important. The paper has been improved after the revision.

Author Response

We thank the reviewer for this positive assessment of our manuscript

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

Using time series Sentinel-1 images for object-oriented crop classification in Google Earth Engine

The research utilizes Sentinel-1 (S1) time-series imagery (composites of 10, 15 and 30 days) for crop classification in Heilongjiang Province, China. Furthermore, pixel-based versus object-based classifications were compared, as well as different sizes of the simple noniterative clustering (SNIC) method. Results suggest using 10-day composites along with object-based method for the study areas with large plots.

The topic of this study is acutal, however, the manuscript was not well organized or written. Many contradictory statements and highly arguable were written in the manuscript, as follows:

• First of all, English must be extensivly edited and changes in reporting style are required
• Confusing terms regarding optical and SAR imagery are used in the manuscript; LN 48: „to classify single-stage or multistage optical images“, LN 57: „multiphase SAR data“, LN 58: „single-phase SAR data“, LN 59: „single polarization SAR data“, then again different term in LN 339-340 „single-polarized and dual-polarized radar data, multi-polarized RADARSAT-2 data..“, and also LN 223 „retroreflection coefficients“ for polarisation bands
• In abstract as the highest OA result is mentioned 10-d S1 composite with 95.49%, whereas in Table 2 the best result is 95.47%, also LN 148-149 „there were 493 sample plots“ but when adding 200 corn plots, 268 soybean and 15 rice plots, the sum is 483
• Furthermore, the scientific contributions in LN 87-92 overlap in 1) „to evaluate (...) different segmentation size object-oriented methods“ and 4) „to study the best segmentation size of object-oriented crop classification“
• Scientific contribution 2) „to study the key period of using Sentinel-1 images for crop classification“, LN 112 – 114 mention that crops are sown in April and harvested in October, but the authors analyze May to September, and their conclusion is that bands with higher importance were mainly distributed in three out of five analyzed months
• Some terms have to be rewritten since they are pretty abstract LN 398 „reasonable classification accuracy“, LN 413 „ideal accuracy“ and there is a lot of repetition „the average plot size was large/small“
• It is very ungrateful to say that: (LN 70 and 71) „no one has studied the impact of SAR data temporal resolution on crop classification accuracy [16]“ and then provide a reference at the end of the statement
• LN 134 – 145 have to be rewritten since they are very hard to read/follow
• Used keywords are highly correlated with the title, I suggest providing some additional keywords
• In this research class rice is pretty unrepresented, so the fact that LN 278 „object-oriented classification always had a higher UA, mainly due to the confusion between corn and soybean in pixel-based classification“ is pretty straightforward since these are only two classes left for evaluation
• Figures 3-6 are very hard to read, whereas Fig 7 and 8 have in legend class which is not shown on the graph

I think that the research design of this manuscript is good, but the results have not been clearly presented. Currently written manuscript does not provide significant scientific contributions in order to be published in this journal.

The authors should consider including some additional input features, such as band ratio or Sentinel-1 Radar Vegetation Index (Holtgrave et al. 2020), as well using deep learning based methods, like FCN, DeepLab.

 

REFERENCES

Holtgrave, A. K., Röder, N., Ackermann, A., Erasmi, S., & Kleinschmit, B. (2020). Comparing Sentinel-1 and-2 Data and Indices for Agricultural Land Use Monitoring. Remote Sensing, 12(18), 2919.

Reviewer 2 Report

My first concern is related to the quality of the figures and the information of the legends.

Using time series Sentinel-1 images for object-2 oriented crop classification in Google Earth Engine

In this manuscript the authors aim to evaluate the application of time series Sentinel-1 images for crop evaluation, in the study they use two sampling areas, different time series and plot sizes obtaining different type of results.

The abstract is clear, but in my opinion it should include a general conclusion of the main results.

Keywords should include an extra word making reference to time, as temporal scale or time series.

Introduction

The introduction is clear; the authors describe the importance to improve crop production if the world and enhance the importance of remote sensing technology to evaluate crop classification.

Instead of using optical images (which are dependent of clouds or other atmospheric phenomenon), the authors explain as radar images (SAR synthetic aperture radar images) show great potential for mapping crop distribution, more independent of external conditions.

With this idea, they present the methodology and the 4 objectives of the study, which are clearly explained.

Materials and methods

Lines 98-102. The size of the two study areas should be indicated in the text.

Figure 1. The legend of this figure should be more extended, explaining as in the left side is the map of China and in the right side if it is the province and the exact location of the study areas.

Lines 107-108. It should be interesting for the reader to have climatic information of the area, including temperatures and climate type.

The analysis of the methodology is clear, including preprocessing of the images and analysis of the information.

Results

In this section, the main concern is about the quality of the figures and the information of the legends.

Lines 209-210. In these lines the authors indicate that X axis of figure 2 is the time series of images, but as we don’t have any data of time in the axis, we never know as the differences among crops are in July or September.

In the same way, the legend of Figure 2 should be more clear and should include more information.

Crop classification is one of the main goals of this study, but the quality of figures 3-6 is very bad and it is hardly difficult to identify the crops and the information in the maps.

Line 273. There is a mistake, instead of having figure 5, for the user accuracy (UA) and producer accuracy (PA) for images with different 273 temporal resolutions should be figures 7 and 8.

In Figures 9 and 10 the meaning of X-axis should be included in the figure or in the legend.

The rest of the text seems clear, the discussion emphasizes the importance of the segmentation size and the relevance of the methodology and the conclusions are well described.

 

Reviewer 3 Report

Utilization of Radar data for crop classification is very useful and interesting topic and this could be very useful for future utilization of EO in Agriculture. But to bring real benefit for the community, paper need to be significantly improved.

Literature review is provided on very generic level and it is necessary extend this part and more detail describe, what was benefit of these papers for your work.  This description as is not give any clear, information how mentioned literature describe state of the art. Please extended this part and give more details

About study are please describe in more details, how the are was selected, to guarantee, that only these three crops will be present in this area. What will happened, if this method will be used for more heterogenous area.

Can you describe in more details preprocessing  methods and algorithms, which was used. There are mentioned only 1) thermal noise removal; 2) radiometric calibration; 3) terrain-corrected value that is logarithmically scaled. Please describe it in sufficient details.

Please describe in details segmentation and classification methods, which was used for both. For such type of paper and with the fact, that paper will read people, who don’t know these algorithms is necessary to provided their description. It is mainly important for this segmentation.

Please make figures 3, 4, 5, 6. On these image is difficult to see real details. Please put here bigger images with better resolution.

From images 11,  12 is not well visible advantage of segmentation,  objects are real fragmented, explain, how it is used for final classification and what is advantage.