Crop Classification Based on Temporal Signatures of Sentinel-1 Observations over Navarre Province, Spain

Round 1

Reviewer 1 Report

The paper investigated the temporal signatures in high heterogeneity crop classification considering the agro-climatic and field size, which is a significant job. Whereas, there are some critical issues with regard to the methodology and results. The dataset, time duration, temporal signature, and classification method, results are not clearly clarified.

1.Line 180, what is the final result coverage since the ASC & DESC coverage area are not the same. In DESC data, there are some missing parts as you mentioned before. Besides, are you layer stacking the two mode (ASC & DESC) dataset according to their time? The explanation of these parts was not clear.

2. Section 2.3, line 190, how do you generate the 5 m buffer, within the boundary or outside the boundary? If it is outside the boundary, how to avoid the mixed pixels?

3. Line 200-201, are you comparing the temporal signature with the median signature of the declared class? How to judge the difference and How do you discard the 10% fields? Please clarified this point.

4. Line 213-214, the definition of temporal signature is not clear. What are the characteristics time series? In your case, do you mean the time series Sentinel-1 dataset?

5. There is no detail information of the supervised classification with only one reference, do you have any algorithm to support? Are you just using R2 and RMSE for classification? Why not use other supervised algorithms like support vector machine or random forest? How do you choose the calendars for comparison since we have no idea of the crop classes first. Do you do the classification (comparison) according to the selected calendars days or the whole time series? In addition, in line 228, we don't know what the mean fit is.

6. As for the time period in line 259, why choose 487 days but not in a yearly basis, that is 365 days.

7. In terms of the field size, do you mean the extent of the study region or training/validation samples? Do you have a distribution map of your declarations and inspections ground truth data? Besides, can you show the final classification map? We can only see the statistics from your manuscript.

Others related issues that should be revised:

1. Line 146-147, do you have any references that you used the seven agricultural regions with distinct Argo-climate conditions? Or you define it yourself. Please clarified this point.

2. Line 164, the statement of three relative orbits is misleading since sentinel-1 only has two kinds of orbits. And the 8DESC or 81DESC not just covered your study area.

3.line 247, table 2 doesn’t have marginal crop.

4. Line 63-66, the logic of the sentence is not correct. Before Sentinel, Landsat can also do the temporal analysis. In addition, how do you define the remote sensing era, which doesn’t make sense.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

I have reviewed the paper entitled "Crop classification based on temporal signatures of Sentinel-1 observations over a diverse region", authors conducted important work about using Sentinel-1 SAR data to identify crop types, the experiments contain multiple crop types, but I think the work could be improved to be more scientific-sounding.  

1. The objective of using SAR data for crop classification is that the optical data are seriously affected by cloud cover. It is widely known that Sentinel-1 data could be used for classification. So that I suggest the authors also use Sentinel-2 data to identify crop types with the same training and validation data sets and to compare the classification of optical-drived and SAR-derived results. Then we can see in the situation that optical data is missing, what accuracy will the SAR data achieve.   

2. Authors analyzed crop separability in section 4.1, but I think there should be a quantitative analysis of crop separability, such as using JM distance (Wardlow 2007, Hao 2015). Therefore, I suggest using JM distance to analyze crop separability (this analysis could be a part of the Result section).

Wardlow, B.D.; Egbert, S.L.; Kastens, J.H. Analysis of time-series MODIS 250 m vegetation index data 752 for crop classification in the U.S. Central Great Plains. Remote Sens. Environ. 2007, 108, 290–310.

Hao, P.Y., Zhan, Y.L., Wang, L., Niu, Z., & Shakir, M. (2015). Feature Selection of Time Series MODIS Data for Early Crop Classification Using Random Forest: A Case Study in Kansas, USA. Remote Sensing, 7, 5347-5369

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Overall a good paper, however a certain novelty is lacking, as crop type classification based on Sentinel-1 time series is a well-known topic from numerous other research.

2.2 - (optional), an interesting additional experiment would be to create 2 different datasets: 1 based on descending orbit data and 1 based on ascending orbit data and to compare the classification accuracies. As the different look angles might affect the backscatter time series signatures.

2.2 - the orbit file correction should be the first step of the workflow, followed by thermal noise removal and the rest

2.2 - just out of interest: did you try out using sigma naught instead of beta nought?

2.2 - why was the pixel size set to 20m?

2.3 - very good training and validation dataset, and also a good preprocessing of the data

2.3 - aggregating all pixels per object and obtaining median values seems insufficient. did you conduct an experiment with other features? e.g. stddev, mean, min, max, etc.. Also other statistsical features would be interesting, for instance area under curve, min, max, kurtosis, skewness of the respective time series. A subsequent analysis of the feature importance will also give insights on which features are more descriptive to derive a certain crop type class?

2.4 - did you try out other classification methods? Why did you prefer curve fitting rather than machine learning based classification approaches (SVM, Random Forest, Neural Networks)? A comparative analysis to other methods would be interesting and necessary! Also you would obtain posterior probabilities per parcel and per class.

2.4 - please describe the global radar feature scheme (g)

3.1. - Legibility of Figures 3,4,5 should be improved. 

3.2.5 - nice analysis of the influence of field sizes. However, change line chart style for Figure 7 to bar chart

4.6. - maybe a combination with optical data will yield higher classification accuracies.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The manuscript has been significantly improved for not only the methodology but also the results.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The author responded to my previous comment vaguely. 

1 The authors have provided a figure about the data availability of Sentinel-2 data in the study region, the figure showed that the figure showed the Sentinel-2 data are available in April, June, and August, I think these Sentinel-2 images are enough to generate crop type mapping. Or at least, authors should use these images to identify crop types and compare the Sentinel-2 data results with the SAR data result.  

2 Authors do not clear about the definition of separability and feature importance. JM distance is a kind of separability rather than feature importance.  

3 The author has calculated the JM distance by combination all HH/HV time series but this is not enough. Both JM distance and feature importance should be calculated for each temporal phase in the time series and find which temporal contributes more to the crop type classification. This result is scientific-sounding  

4 Authors only showed that the SAR time series could be used for crop type mapping, but this is not "new"! It is just like the homework of a college student and this kind of work has been widely conducted. I suggest the author treat the reviewers' comments carefully and improve the quality of the paper. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Nice improvement through adding the Jeffries–Matusita metric! Overall an interesting paper!

Author Response

Please see the attachment.

Author Response File: Author Response.pdf