Mapping Winter Wheat in North China Using Sentinel 2A/B Data: A Method Based on Phenology-Time Weighted Dynamic Time Warping
and Xihong Cui 1,2Round 1
Reviewer 1 Report
The manuscript presented for evaluation is a well-written and interesting example of using multi-temporal data to identify winter wheat. Overall, I assess the presented text as very valuable and free of major errors or omissions.
Below are my specific comments and questions to the text:
- Figure 1b - Please add which VI the curves represent.
- Page 2 You write "Moreover, the collection of training
data covering a large area not only is time consuming, but also has a high field cost." - At the same time, in chapter 2.2. it is written that the reference can be obtained as a result of photo-interpretation. Please explain this discrepancy. - Figure 2 - I propose to change the background raster to the average annual air temperature or the length of the growing season.
- Page 5 You write "In total, 62,239 ground-truth pixels with 20 m spatial resolution were collected for accuracy evaluation including 27,669 winter wheat and 34,570 other pixels". Please add information what was the internal background class division into subclasses.
- Figure 3 shows that non-homogeneous pixels (containing pathways/roads) were used as reference for winter wheat. What was the impact of such artifacts on the modeling result? Why aren't such pixels excluded.
- Page 12 - You write "The zoomed-in images in Figure 10 show rich spatial details of the winter wheat cropland, and the cultivated parcels
are easily recognizable." I think the scale of their maps is wrong. There are no details on them to visually assess the correctness of the result." Please consider changing the scale. - Figure 11 - The word "Cities" should probably be changed into "regions"?
- Figure 12 - I suggest to change the scale as for Figure 10
- The analyses carried out in Chapter 5.1 should be transferred to results chapter. They are not a discussion but a description of results.
- Figure 15 - the column "Winter wheat change" shows significant changes in wheat cover in the two years analysed. Please expand on what this difference can result from?
Author Response
Response to Reviewer 1 Comments
The manuscript presented for evaluation is a well-written and interesting example of using multi-temporal data to identify winter wheat. Overall, I assess the presented text as very valuable and free of major errors or omissions.
Response: Thank you very much for your approval and very detailed comments. We have made revisions according to your suggestions, and the revised parts were highlighted in yellow in the revised manuscript.
Point 1: Figure 1b - Please add which VI the curves represent.
Response 1: We used NDPI for curves in figure 1a and figure 1b, we have added the annotations (Figure 1).
Point 2: Page 2 You write "Moreover, the collection of training data covering a large area not only is time consuming, but also has a high field cost." - At the same time, in chapter 2.2. it is written that the reference can be obtained as a result of photo-interpretation. Please explain this discrepancy.
Response 2: Sorry to confuse you with the discrepancy. Actually, the cost of sample collection depends on the number of samples. In page 2, we meant to emphasize the difficulty of collecting large number of training data. Whereas in Chapter 2.2, only a small number of reference (10 samples for each type) obtained by photo-interpretation were used for the method development. That is also the reason why we employed a sample expanding step to acquire larger training dataset for parameter optimization. We have rewritten the sentences to better convey our ideas and diminish the discrepancy between the two points (page 3 and chapter 2.2 in page 5).
On the other hand, it is true that the acquisition of validation set in this study is highly cost. However, the dataset is used only for validating the effectiveness of the proposed method. Such validation dataset will be not required when the proposed method is used for different area or different crop types.
Point 3: Figure 2 - I propose to change the background raster to the average annual air temperature or the length of the growing season.
Response 3: Thank you for this valuable comment. We have changed it as your suggestion (Figure 2).
Point 4: Page 5 You write "In total, 62,239 ground-truth pixels with 20 m spatial resolution were collected for accuracy evaluation including 27,669 winter wheat and 34,570 other pixels". Please add information what was the internal background class division into subclasses.
Response 4: Among the 34,570 non winter wheat pixels, there are seven subclasses including forest, barren land, artificial surface, vegetable, paddy rice, water, and grass/shrub land. We have added the subclass information to the revised manuscript (page 6).
Point 5: Figure 3 shows that non-homogeneous pixels (containing pathways/roads) were used as reference for winter wheat. What was the impact of such artifacts on the modeling result? Why aren't such pixels excluded.
Response 5: Sorry to confuse you with the mixed pixels in figure 3. Actually, figure 3 shows an example of collecting samples, most of the which were used for accuracy evaluation and only 80 samples were used for parameter determination. All the 80 samples were selected from pure pixels, so we believe the impact of mixed pixels on modelling result were reduced as far as possible. We have added sentences to emphasize the functions of large samples and the purity of the 80 samples (chapter 2.2 in page 5).
Regarding the accuracy evaluation, such mixed pixels are necessary because there is still land cover mixture even for the pixels at 20-m resolution. Therefore, when visually interpreting samples, the pixels with larger than 50% were defined as winter wheat, otherwise they are defined as non winter wheat. Such a strategy is commonly used in the accuracy evaluation of hard classification (Stehman and Foody, 2019).
Stehman S. V., Foody G. M. (2019) Key issues in rigorous accuracy assessment of land cover products. Remote Sensing of Environment, 231, 111199
Point 6: Page 12 - You write "The zoomed-in images in Figure 10 show rich spatial details of the winter wheat cropland, and the cultivated parcels are easily recognizable." I think the scale of their maps is wrong. There are no details on them to visually assess the correctness of the result." Please consider changing the scale.
Response 6: Revised (Figure 10).
Point 7: Figure 11 - The word "Cities" should probably be changed into "regions"?
Response 7: Thank you for your good suggestion. We have changed ‘Cities’ of the legends into ‘Regions’ (Figure 11).
Point 8: Figure 12 - I suggest to change the scale as for Figure 10
Response 8: Revised (Figure 12).
Point 9: The analyses carried out in Chapter 5.1 should be transferred to results chapter. They are not a discussion but a description of results.
Response 9: We prefer to think that putting section 5.1(Superiority of NDPI) in discussions is more appropriate than in the results with the following considerations. Firstly, given that the direct objective of this work is to map winter wheat rather than the comparison of before-winter-peak of multiple vegetation indices, we only show the mapping results and mapping accuracies of different methods in the result section to focus on our objective. Secondly, section 5.1(Superiority of NDPI) and 5.2(Benefits of the PT-DTW method) both explain the reason why the proposed PT-DTW method surpasses other methods, which further confirms the superiority of the proposed method in an indirect manner. Thus, we decided not to change this structure in the revised manuscript.
Point 10: Figure 15 - the column "Winter wheat change" shows significant changes in wheat cover in the two years analysed. Please expand on what this difference can result from?
Response 10: The winter wheat cultivation in this area located in Hebei province, experienced some changes because of the new crop rotation and fallow policy implemented since 2016 [40, 41] (section 5.3 in page 18).
- The implementation plan of agricultural project for comprehensive management of groundwater overexploitation in Hebei Province. Department of Agriculture and Rural Affairs of Hebei Province. China. 2016.
- The implementation plan of Hebei Province's seasonal fallow system in 2016. Department of Agriculture and Rural Affairs of Hebei Province. China. 2016.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The article entitled „Mapping winter wheat in North China using Sentinel 2A/B data: A method based on phenology-time weighted dynamic time warping” presents a new method for identifying and mapping winter wheat fields using medium spatial resolution satellite images. The Authors developed a new method of matching seasonal curves of vegetation indices. They used in the study NPDI index, which proved to be more useful to describe wheat development in the autumn than commonly used NDVI and EVI indices.
The work is well presented and fits into the scope of the journal therefore, in my opinion the manuscript is acceptable for publication. I have only a few questions:
- Figure 1. What is this „Typical VI”? NDVI or some other? Not all plant indicators have the same time course in the growing season and they don't all have values from 0 to1.
- Figure 1 What is the source of data on which the curves were drawn on the chart. The D curve looks strange, untypical. Why are VI values so high in August - are they higher than in May? Explanation later in the text is not enough. If NDVI is the index, wheat would have to be sown in May to achieve such high values in September.
- Figure 2. In what units is DEM given?
- Equation (1). It is necessary to explain why the image data were resampled to 20 m. Probably because the SWIR data are in this resolution. For another reason?
- Equation (1). The numbers and spectral ranges of the Sentinal sensor channels used to calculate the NDPI should be given.
- Figure 5. If wheat is sown in the study area in September, it should be explained why in Fig. 5a one of the three wheat curves has such high NDPI values (> 0.5) in September and why then they fall until October?
Author Response
Response to Reviewer 2 Comments
The article entitled „Mapping winter wheat in North China using Sentinel 2A/B data: A method based on phenology-time weighted dynamic time warping” presents a new method for identifying and mapping winter wheat fields using medium spatial resolution satellite images. The Authors developed a new method of matching seasonal curves of vegetation indices. They used in the study NPDI index, which proved to be more useful to describe wheat development in the autumn than commonly used NDVI and EVI indices.
The work is well presented and fits into the scope of the journal therefore, in my opinion the manuscript is acceptable for publication. I have only a few questions:
Response: Thank you very much for your approval and very detailed questions. We feel sorry to confuse you with some questions because of inappropriate statements. Revisions have been made and they are highlighted in yellow in the revised manuscript.
Point 1: Figure 1. What is this „Typical VI”? NDVI or some other? Not all plant indicators have the same time course in the growing season and they don't all have values from 0 to1.
Response 1: We used NDPI for curves in figure 1a and figure 1b, and we have added the annotations (Figure 1).
Point 2: Figure 1 What is the source of data on which the curves were drawn on the chart. The D curve looks strange, untypical. Why are VI values so high in August - are they higher than in May? Explanation later in the text is not enough. If NDVI is the index, wheat would have to be sown in May to achieve such high values in September.
Response 2: Firstly, NDPI curves on figure 1 were calculated from sentinel-2 MSI bands. We have added the statements of data source (Figure 1 annotation). Secondly, curve D in figure 1b is from a location with a rotation between winter wheat and summer corn. The high NDPI value in August results from summer corn, which in this location is sown in late May and harvested in September. In August, summer corn comes to the tasselling stage, reaching its highest leaf area index, even higher than that of winter wheat, so NDPI value in August and early September is very high, even higher than that in May. We have added the crop rotation background information in the study area description (section 2.1 in page 4). Also, we have changed the start time of NDPI curves from September to October in order to mitigate confusion (Figure 1b).
Point 3: Figure 2. In what units is DEM given?
Response 3: DEM unit is meter. However, in this revised manuscript, we changed the background raster from DEM to AAAT (annual average air temperature) according to the suggestion of another reviewer. We also agree that AAAT can better show the climate variation of the study area, which has a closer relationship to winter wheat cultivation than DEM (Figure 2 and section 2.1 in page 4).
Point 4: Equation (1). It is necessary to explain why the image data were resampled to 20 m. Probably because the SWIR data are in this resolution. For another reason?
Response 4: Your inference is right. We resampled image data to 20m because SWIR data are in this resolution.
Point 5: Equation (1). The numbers and spectral ranges of the Sentinal sensor channels used to calculate the NDPI should be given.
Response 5: We have added information of sentinel-2 channels used to calculate NDPI (section 3.1 in page 6).
Point 6: Figure 5. If wheat is sown in the study area in September, it should be explained why in Fig. 5a one of the three wheat curves has such high NDPI values (> 0.5) in September and why then they fall until October?
Response 6: With the same reason as for Point 2, high NDPI value in September and falling until October results from the growth of summer corn/paddy rice. We have added the background rotation information to the study area description (section 2.1 in page 4). Also, we have changed the start time of NDPI curves from September to October in order to diminish confusion (Figure 5).
