An Improved Approach Considering Intraclass Variability for Mapping Winter Wheat Using Multitemporal MODIS EVI Images
, Bo Tao 2, Wei Ren 2,*, Demetrio P. Zourarakis 3, Bassil El Masri 4, Zhigang Sun 5 and Qingjiu Tian 1Round 1
Reviewer 1 Report
In this study, the authors introduced the sub-classes training samples for interpreting the intra-class variability based on the angles and distances of EVI time series multi-vector in winter wheat area mapping. After that, the effects of landscape structure on remote sensing-based crop mapping accuracy were analysed using two landscape metrics FRG and PLAND. The evaluation results showed good performances of the method when applied in Kansas and NCP.
I have some concerns with this article:
1) From my opinion, I don’t think it is a novel approach to considering Intra-class Variability for Mapping Winter Wheat. The intra-class variability of winter wheat caused by field conditions and management might cause different spectral responses even in the same growing phase. When implementing supervised or unsupervised classification for winter wheat mapping, we often sampled and divided the types of winter wheat into winter wheat 1, winter wheat 2, and so on, according to their different characteristics, which implied the Intra-class Variability were taken into consideration. Besides, many classification methods also calculate the distances and angles between the pixels and the training samples. Also, there are many studies on the similarity between two time series based on distance or angle.
2) P2, line 65. I don’t agree on that “In most of these methods, training samples are generated from specific features different from other classes but ignored intra-class differences.”
3) How to evaluate the thresholds for segmentation of sub-classes due to the intra-class differences is reasonable? The author noted the segmentation points were determined to make sure each sub-class contained roughly the same amount of training samples (20% ~ 30%). Why?
4) The scheme of segmentation for four sub-classes based on the value ranges of two EVI peaks in the two study areas were determined by trial and error procedure. Did such scheme have any biophysical meanings for winter wheat growing? In other words, did the scheme or intra-class variations reflect the climate, soil condition, or planting date lag across the region?
5) P7, Table 2, the number of training and validation samples for Kansas and NCP should be wrong written.
Author Response
Response to Reviewer 1 Comments
In this study, the authors introduced the sub-classes training samples for interpreting the intra-class variability based on the angles and distances of EVI time series multi-vector in winter wheat area mapping. After that, the effects of landscape structure on remote sensing-based crop mapping accuracy were analysed using two landscape metrics FRG and PLAND. The evaluation results showed good performances of the method when applied in Kansas and NCP.
[Response: Thanks for the reviewer’s positive and constructive comments on this study.]
I have some concerns with this article:
1) From my opinion, I don’t think it is a novel approach to considering Intra-class Variability for Mapping Winter Wheat. The intra-class variability of winter wheat caused by field conditions and management might cause different spectral responses even in the same growing phase. When implementing supervised or unsupervised classification for winter wheat mapping, we often sampled and divided the types of winter wheat into winter wheat 1, winter wheat 2, and so on, according to their different characteristics, which implied the Intra-class Variability were taken into consideration. Besides, many classification methods also calculate the distances and angles between the pixels and the training samples. Also, there are many studies on the similarity between two time series based on distance or angle.
[Response: Thanks for the reviewer’s comment. We agree that some studies have considered the intra-class variability for winter wheat mapping (such as Qiu et al., 2017; Wardlow et al., 2007). The novelty of our study is considering the calculation of angles and distances based on time series data for identifying sub-classes, which have not been used in the winter wheat mapping based on our literature review although they were widely used in remote sensing-based classification for other crop types, especially using hyperspectral remote sensing images.
In the revised manuscript, we revised the title as “An Improved Approach Considering Intra-class Variability for Mapping Winter Wheat Using Multi-temporal MODIS EVI Images”.]
2) P2, line 65. I don’t agree on that “In most of these methods, training samples are generated from specific features different from other classes but ignored intra-class differences.”
[Response: Thanks for the reviewer’s comment. We acknowledge that this statement was arbitrary. In the revised manuscript, we deleted this sentence and rewrote it as “Many factors may act to affect the remote sensing-based mapping accuracy, in which training samples have a more significant impact than the mapping techniques (Campbell 2003; Foody and Mathur 2004; Hixson et al. 1980).”]
3) How to evaluate the thresholds for segmentation of sub-classes due to the intra-class differences is reasonable? The author noted the segmentation points were determined to make sure each sub-class contained roughly the same amount of training samples (20% ~ 30%). Why?
[Response: Thanks for the reviewer’s comment. In the revised manuscript, we added Jeffries-Matusita Distance to evaluate the thresholds for segmentation of sub-classes in section 3.4.2. The JM distance is widely used for verifying the feasibility of training samples for remote sensing-based classification. We calculated the JM distance between four sub-classes and other land cover types for two study areas (Section 4.1, Table 6).
For the reviewer’s second question. We acknowledged that this statement was made more ambiguous than intended, and we revised this sentence to be clearer. We did not purposely make sure each sub-class contained roughly the same amount of training samples. It was the scheme for identifying thresholds that make sure each sub-class contained roughly the same amount of training samples.
Specifically, the training samples were randomly selected across the study area, which represent the winter wheat for the whole study area. In Figure 4 that shows the EVI time series of all training samples, we can see that the values of EVI at two peaks are distributed roughly even from low to high. The thresholds we used to divide each sub-class are nearly the middle of the ranges. Therefore, each sub-class contained roughly the same amount of training samples.]
4) The scheme of segmentation for four sub-classes based on the value ranges of two EVI peaks in the two study areas were determined by trial and error procedure. Did such scheme have any biophysical meanings for winter wheat growing? In other words, did the scheme or intra-class variations reflect the climate, soil condition, or planting date lag across the region?
[Response: Thanks for the reviewer’s comment. The EVI curve can reflect the influences of the climate, the growth condition, and the phenology stages of winter wheat (such as the planting date) to some extent. Therefore, the scheme or intra-class variations theoretically have some biophysical meanings. However, we have to collect more datasets, such as climate, field management practices, and crop phenology information to analyze and quantify this complicated situation, which involves much more efforts and is not the focus of this study.
We will collect the datasets and analyze how these factors affect the setting of thresholds under different conditions in the future study. We have included associated discussions in the section of uncertainty and future needs in the manuscript (section 5.4).]
5) P7, Table 2, the number of training and validation samples for Kansas and NCP should be wrong written.
[Response: Thanks for the reviewer’s comment. We corrected the numbers as the reviewer suggested.]
References:
Qiu, B., Luo, Y., Tang, Z., Chen, C., Lu, D., Huang, H., ... & Xu, W. Winter wheat mapping combining variations before and after estimated heading dates. ISPRS Journal of Photogrammetry and Remote Sensing. 2017, 123, 35-46.
Wardlow, B., S. Egbert, and J. Kastens, Analysis of time-series MODIS 250 m vegetation index data for crop classification in the U.S. Central Great Plains. Remote Sensing of Environment. 2007, 108, 290-310.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This manuscript is on mapping winter wheat using time series of EVI calculated from MODIS imagery. The authors used data from the state of Kansas (KS), USA, and the North China Plain (NCP) region in China. The main selling point of the manuscript is the use of intra-class variability to classify winter wheat from the time series data. It’s a good effort, although there are some major weaknesses in this manuscript, as described below. 1. In Figure 4 the authors show four types of intra-class variability in the EVI time series of winter wheat. But it is not clear at all how they concluded that the composite EVI time series (the top-level graph in Figure 4) can be broken down to just four types of intra-classes. Was there any statistical threshold that was followed to determine these four types? Clearly, the bottom level graphs in Figure 4 show that each intra-class can be divided into further intra-classes. This issue needs to be clarified. 2. A related section 3.5 (The Approach…) also needs clarification. It’s not clear how the parameters cos(theta) and D were derived. What do the ‘four standard vectors’ mean? In Euclidean geometry, there are only 3 standard vectors, if the authors meant the set of unit vectors in a Euclidian space. If some other ‘standard vector’ is being discussed, it should be clarified. This entire Approach section (3.5) needs to be rewritten so that it is clearly understandable what was exactly done with the time-series of EVI data to classify the intra-classes of winter wheat. It is not at all clear to this reviewer. Also, some minor issues need to be addresses: should the threshold (b1~b4) be (d1~d4)? And is it parameter ‘d’ or ‘D’ (as in Table 4)? 3. Although the authors present a section (5.3) on the comparison with other studies, they do not present how their study provides any superior results for the KS or NCP study region for the winter wheat areas of the years under the study, compared to other existing methods. How about the Landsat-based crop area maps that currently exist (such as the Global Croplands database), does this study provide a better estimate and classification? Finally, the Introduction and Discussion sections need to justify why this study was needed and what improvement over the existing methods are being provided in this study.Author Response
Response to Reviewer 2 Comments
This manuscript is on mapping winter wheat using time series of EVI calculated from MODIS imagery. The authors used data from the state of Kansas (KS), USA, and the North China Plain (NCP) region in China. The main selling point of the manuscript is the use of intra-class variability to classify winter wheat from the time series data. It’s a good effort, although there are some major weaknesses in this manuscript, as described below.
[Response: Thanks for the reviewer’s positive and constructive comments on this study.]
1. In Figure 4 the authors show four types of intra-class variability in the EVI time series of winter wheat. But it is not clear at all how they concluded that the composite EVI time series (the top-level graph in Figure 4) can be broken down to just four types of intra-classes. Was there any statistical threshold that was followed to determine these four types? Clearly, the bottom level graphs in Figure 4 show that each intra-class can be divided into further intra-classes. This issue needs to be clarified.
[Response: Thanks for this comment. Several statements that we made were more ambiguous than intended, and we revised this sentence to be clearer. In this study, the intra-class variability of winter wheat was fully considered in the process of building the sub-class sets of training samples. The over-wintering period is an obvious physical feature of winter wheat fields, which divides the entire winter wheat growth period into two phases and results in double peaks in the EVI curve. In Figure 4, the EVI peak values before the wintering phase change from 0.1 to nearly 0.65, and the second peak values vary from about 0.35 to around 0.85. The wide differences at each EVI peak exhibit the obvious winter wheat intra-class variability, which was adopted as the unique features to extract sub-class training samples. Therefore, we selected these two EVI peaks as the segmentation points to divide sub-classes.
The training samples were randomly selected across the study area, which represent the winter wheat crop for the whole areas. Figure 4 showed that the EVI values at two peaks were distributed roughly even from low to high. Taking the medians of EVI peak values as the thresholds can divide the total training samples into each sub-class with similar amounts. Thus, the segmentation thresholds of four sub-classes in two study areas were determined in Table 3.
To test the viability of our scheme for dividing into four intra-classes, in the revised manuscript, we introduced the Jeffries–Matusita (JM) Distance to evaluate the separability of four sub-classes, and added a table (Table 6) to display the feasibility of this segmentation. We also added more analysis to explain why we just divided into four sub-classes. ]
2. A related section 3.5 (The Approach…) also needs clarification. It’s not clear how the parameters cos(theta) and D were derived. What do the ‘four standard vectors’ mean? In Euclidean geometry, there are only 3 standard vectors, if the authors meant the set of unit vectors in a Euclidian space. If some other ‘standard vector’ is being discussed, it should be clarified. This entire Approach section (3.5) needs to be rewritten so that it is clearly understandable what was exactly done with the time-series of EVI data to classify the intra-classes of winter wheat. It is not at all clear to this reviewer. Also, some minor issues need to be addresses: should the threshold (b1~b4) be (d1~d4)? And is it parameter ‘d’ or ‘D’ (as in Table 4)?
[Response: Thanks for the review’s comments. In the revised manuscript, we reorganized the section 3.5 and divided it into four sub-sections. We added the description of standard vectors’ calculation, specified the derivation process of two parameters and added the sensitive experiments for determining the thresholds of parameters.
For the question of “standard vector,” we added some explanations in section 3.5.1. In this study, we took the EVI time series as the n-dimensional vector. The standard vector represents the averaged EVI time series respectively calculated from training samples of each sub-class.
We also changed the parameters “b1~b4” to “d1~d4” as the reviewer’s comments.]
3. Although the authors present a section (5.3) on the comparison with other studies, they do not present how their study provides any superior results for the KS or NCP study region for the winter wheat areas of the years under the study, compared to other existing methods. How about the Landsat-based crop area maps that currently exist (such as the Global Croplands database), does this study provide a better estimate and classification? Finally, the Introduction and Discussion sections need to justify why this study was needed and what improvement over the existing methods are being provided in this study.
[Response: Thanks for the reviewer’s comments. In the revised manuscript, we added section 3.8 to describe the comparative experiments, as the reviewer suggested. Based on the same training samples, we conducted experiments using some existing methods (including the maximum likelihood, support vector machine and artificial neural network), and evaluated the results against the same validation datasets (see Section 4.8, Page 21, Line 193-207).
The global cropland datasets (https://croplands.org/app/map?lat=0.17578&lng=0&zoom=2) show spatial distribution of cropland at 30m spatial resolution, but it do not classify into specific crop types. For the United States, they provide (same website) 250m MODIS-based crop types distribution during 2001-2013, but do not separate wheat into winter wheat and spring wheat. This situation justifies the importance of this study.
In the revised manuscript, we also added another test for evaluating superiority of considering the intra-class variability (i.e., the approach considering intra-classes vs. the one without intra-classes). We added to section 4.6 with two figures and one table to illustrate the comparisons. Our results showed that the approach considering intra-class variability improved the winter wheat mapping accuracies by 17% in Kansas and 15% in the NCP, respectively.
To clearly justify the objective of this research and to explain the improvement of our study as compared with the existing methods. In the revised manuscript, we added some sentences in Introduction section to emphasize the importance of this study. In addition, we added more descriptions and discussions to show the improvements in this study over existing methods. (See section 4.8 and section 5.3).
Author Response File: Author Response.docx
Reviewer 3 Report
This is an article that presents an interesting approach, the article is well written and well structured. The illustrations are relevant and make it possible to clearly show the approach proposed by the authors.
My main concern is that the article clearly claims in several passages the superiority of considering intra-class variability to improve the detection of wheat fields, but unfortunately it does not show this. The article shows that it is possible, and with a good level of certainty, to classify wheat plots in two situations, but to show the real interest, the results obtained should be compared to classical method (without taking into account intra-class variability) on the same examples.
The choice of some parameters is not well justified/discussed. I can understand that in a first approach, these parameters are adjusted in the best expert way, but this does not argue in favour of generalizing the approach or in favour of defining a robust and repeatable framework for use. This is the case, for example, of the EVIj and EVIi thresholds, the value of 0.53 for EVIj seems completely arbitrary to me. I think a sensitivity study to see if the approach is greatly or slightly affected by small variations in these parameters should be performed (if this has not already been done). This would then allow discussion on the need to pay particular attention to the choice of these parameters or not and, if necessary, to propose a method for their choice. Another aspect that is disturbing is the choice of the 4 classes. This aspect is never criticized or discussed. This choice also seems to me important for the method and here again the work lacks a robustness study at this level (perhaps such a study has been carried out and it is not presented, if this is the case, it must be mentioned to know if the number of classes impacts the quality of the results or not and if it is possible to propose an optimal number of classes).
The methods presented in section 3.5 are relatively traditional, it would be necessary to refer to reference work in this field. The way in which the method is presented leaves many questions and lacks clear explanations. For example, in equations 3, reference is made to the index "i" which is not defined (whereas in fact i = n-1, if as shown on line 231, n indicates the dimension of the vectors. I think what is explained between lines 241 and 246 is essential to understand the approach and how to implement it. This part deserves much more explanations so that it is much clearer for the reader. Equation 4 is also incomprehensible as presented.
Specific comments
abstract
line 24 : I would say agricultural practices instead of management,
line 26 : the angles and distances of MODIS EVI times series with intra-class variability, I think at this stage this is difficult to understand in the abstract. I would suggest the authors to stay general, I propose variables derived from MODIS EVI times series taking into account intra-class variability.
line 29 : please define the accuracy and R², this is not clear how what they refer to at this stage of the abstract,
line 31 : A comparison with other studies, I strongly desagree with this statement. We can talk about te meaning of comparison, but usually when two methods are compared, they have to be comparable (at least using the same data set to let the reader understand how better is the proposed approoach. This is not the case in this paper.
Introduction
line 64 : specific feature different from other classes = this is not clear what you want to mean,
line 72 : if I understood well, this is not different status, but different stages,
line 81 : english is not my native language, but I guess this is "considering intra-class variability" (without "of")
line 101 : I am surprised by the notation > 450 mm, I would remove the signs, "... precipitation ranges from 450 mm in the west to 1,200 mm... is enough I think (same lines after)
Figure 1.a : the weather stations are not positopnned on the map (contrary to figure 1.b)
line 128 : please define clearly what you call accuracy, PA, and UA (you can refer to the next sections may be),
line 140 : this is not clear to me wat are te collected observations (is it only meteorological data ?)
line 187 : I guess you chose a quadratic polynomial for the S-G fliter because this is common on other works ? Please cite these work,
Figure 3 : very clear, thank you,
Author Response
Response to Reviewer 3 Comments
This is an article that presents an interesting approach, the article is well written and well structured. The illustrations are relevant and make it possible to clearly show the approach proposed by the authors.
My main concern is that the article clearly claims in several passages the superiority of considering intra-class variability to improve the detection of wheat fields, but unfortunately it does not show this. The article shows that it is possible, and with a good level of certainty, to classify wheat plots in two situations, but to show the real interest, the results obtained should be compared to classical method (without taking into account intra-class variability) on the same examples.
[Response: Thanks the reviewer’s constructive comments. In the revised manuscript, we conducted experiments using some existing methods based on the same training samples (including the maximum likelihood, support vector machine and artificial neural network) (see Section 3.8, Page 16, Line 101-116), and evaluated the results against the same validation datasets (see Section 4.8, Page 21, Line 193-207).
We also added another test for evaluating superiority of considering the intra-class variability (i.e., the approach considering intra-classes vs. the one without intra-classes). We added to section 4.6 with two figures and one table to illustrate the comparisons. Our results showed that the approach considering intra-class variability improved the winter wheat mapping accuracies by 17% in Kansas and 15% in the NCP, respectively. ]
The choice of some parameters is not well justified/discussed. I can understand that in a first approach, these parameters are adjusted in the best expert way, but this does not argue in favour of generalizing the approach or in favour of defining a robust and repeatable framework for use. This is the case, for example, of the EVIj and EVIi thresholds, the value of 0.53 for EVIj seems completely arbitrary to me. I think a sensitivity study to see if the approach is greatly or slightly affected by small variations in these parameters should be performed (if this has not already been done). This would then allow discussion on the need to pay particular attention to the choice of these parameters or not and, if necessary, to propose a method for their choice. Another aspect that is disturbing is the choice of the 4 classes. This aspect is never criticized or discussed. This choice also seems to me important for the method and here again the work lacks a robustness study at this level (perhaps such a study has been carried out and it is not presented, if this is the case, it must be mentioned to know if the number of classes impacts the quality of the results or not and if it is possible to propose an optimal number of classes).
[Response: Thanks for the reviewer’s comments. Following the reviewer’s suggestion, we added a sensitivity experiment to examine if the robustness of our approach is affected by the variations of parameters and provided detailed explanations (see Section 3.5.3, Page13-14, Line 296-304).
In this study, the intra-class variability of winter wheat was fully considered in the process of building the sub-class sets of training samples. The over-wintering period is an obvious physical feature of winter wheat fields, which divides the entire winter wheat growth period into two phases and results in double peaks in the EVI curve. In Figure 4, the EVI peak values before the wintering phase change from 0.1 to nearly 0.65 and the second peak values vary from about 0.35 to around 0.85. The wide differences at each EVI peak exhibit the obvious winter wheat intra-class variability, which was adopted as the unique features to extract sub-class training samples. Therefore, we selected these two EVI peaks as the segmentation points to divide sub-classes.
The training samples were randomly selected across the study area, which represent the winter wheat crop for the whole areas. Figure 4 showed that the EVI values at two peaks were distributed roughly even from low to high. Taking the medians of EVI peak values as the thresholds can divide the total training samples into each sub-class with similar amounts. The median of second EVI peak values was 0.53, which was used for dividing all training samples to two first-level sub-classes. The thresholds 0.3 and 0.35 were the medians of two first-level sub-classes and subdivided training sets into four second-level sub-classes. Thus, the segmentation thresholds of four sub-classes in two study areas were determined in Table 3.
To address the question about the choice of four sub-classes, in the revised manuscript, we introduced the Jeffries–Matusita (JM) Distance to evaluate the separability of four sub-classes, and added a table (Table 6) to display the feasibility of this segmentation. We also added more analysis to explain why we just divided into four sub-classes.
In this study, we identified three thresholds (Figure 4, Table 3) based on two EVI peaks, which divided all training samples into four sub-classes. If we divide into more sub-classes, for example, 9 sub-classes (in this case, we need to identify thresholds using points of trisection of EVI peak values), the separability is too low to be accepted. The evaluation results are demonstrated below:
Kansas | Sub-class 2 | Sub-class 3 | Sub-class 4 | Sub-class 5 | Sub-class 6 | Sub-class 7 | Sub-class 8 | Sub-class 9 |
Sub-class 1 | 1.53 | 1.92 | 1.58 | 1.94 | 1.98 | 1.90 | 1.72 | 1.96 |
Sub-class 2 | 1.89 | 1.68 | 1.77 | 1.95 | 1.89 | 1.74 | 1.91 | |
Sub-class 3 | 1.93 | 1.88 | 1.56 | 1.97 | 1.96 | 1.85 | ||
Sub-class 4 | 1.84 | 1.97 | 1.82 | 1.72 | 1.96 | |||
Sub-class 5 | 1.94 | 1.89 | 1.97 | 1.78 | 1.90 | |||
Sub-class 6 | 1.99 | 1.97 | 1.70 | |||||
Sub-class 7 | 1.83 | 1.98 | ||||||
Sub-class 8 | 1.79 | |||||||
NCP | Sub-class 2 | Sub-class 3 | Sub-class 4 | Sub-class 5 | Sub-class 6 | Sub-class 7 | Sub-class 8 | Sub-class 9 |
Sub-class 1 | 1.52 | 1.73 | 1.62 | 1.49 | 1.84 | 1.79 | 1.54 | 1.79 |
Sub-class 2 | 1.55 | 1.51 | 1.38 | 1.68 | 1.42 | 1.50 | 1.62 | |
Sub-class 3 | 1.81 | 1.76 | 1.82 | 1.80 | 1.83 | 1.97 | ||
Sub-class 4 | 1.29 | 1.73 | 1.23 | 1.36 | 1.55 | |||
Sub-class 5 | 1.31 | 1.37 | 1.32 | 1.47 | ||||
Sub-class 6 | 1.69 | 1.54 | 1.58 | |||||
Sub-class 7 | 1.17 | 1.27 | ||||||
Sub-class 8 | 0.91 |
From the table, the lower JM Distances inhibits that almost no one sub-class is separable with each other in both study areas. Therefore, we just used four sub-classes for mapping winter wheat in this study.]
The methods presented in section 3.5 are relatively traditional, it would be necessary to refer to reference work in this field. The way in which the method is presented leaves many questions and lacks clear explanations. For example, in equations 3, reference is made to the index "i" which is not defined (whereas in fact i = n-1, if as shown on line 231, n indicates the dimension of the vectors. I think what is explained between lines 241 and 246 is essential to understand the approach and how to implement it. This part deserves much more explanations so that it is much clearer for the reader. Equation 4 is also incomprehensible as presented.
[Response: Thanks the reviewer’s constructive comments. In the revised manuscript, we reorganized the section 3.5 by citing relevant references in this field. We also carefully checked all equations and provided detailed explanations (see Section 3.5.1&3.5.2, Page 12-13, Line 28-48).]
Specific comments
Abstract
line 24 : I would say agricultural practices instead of management,
[Response: Replaced “management” with “agricultural practices” as suggested.]
line 26 : the angles and distances of MODIS EVI times series with intra-class variability, I think at this stage this is difficult to understand in the abstract. I would suggest the authors to stay general, I propose variables derived from MODIS EVI times series taking into account intra-class variability.
[Response: Revised as the reviewer suggested.]
line 29 : please define the accuracy and R², this is not clear how what they refer to at this stage of the abstract,
[Response: We have modified as the reviewer suggested in the abstract.]
line 31 : A comparison with other studies, I strongly disagree with this statement. We can talk about the meaning of comparison, but usually when two methods are compared, they have to be comparable (at least using the same data set to let the reader understand how better is the proposed approach. This is not the case in this paper.
[Response: Thanks for the reviewer’s comment. In the revised manuscript, we added some experiments to compare mapping accuracies between methods with and without considering the intra-class based on the same data. Accordingly, we revised that statement in Abstract as by “Comparisons with methods without considering the intra-class variability.”]
Introduction
line 64 : specific feature different from other classes this is not clear what you want to mean,
[Response: In the revised version, we rewrote this statement as “While, many factors may act to limit remote sensing-based mapping accuracy below. One particular is connected with the training sample sets. Studies found that the training process has a larger impact on mapping accuracy than the mapping techniques (Campbell 2003; Foody and Mathur 2004; Hixson et al. 1980).”]
line 72 : if I understood well, this is not different status, but different stages,
[Response: “status” is replaced with “stages” as the reviewer suggested.]
line 81 : English is not my native language, but I guess this is "considering intra-class variability" (without "of")
[Response: Thanks for the reviewer’s comment. We revised as the reviewer suggested.]
line 101 : I am surprised by the notation > 450 mm, I would remove the signs, "... precipitation ranges from 450 mm in the west to 1,200 mm... is enough I think (same lines after)
[Response 8: Thanks for the reviewer’s comment. We revised as the reviewer suggested.]
Figure 1.a : the weather stations are not positioned on the map (contrary to figure 1.b)
[Response: Thanks for the reviewer’s comment. The read points on Figure 1b are agro-meteorological stations across the NCP region, China (see the legend in Figure 1b). These stations belong to the China Agro-meteorological Observation Network, which provide cropland location and site-level phenological information. This information were used in this study for identifying winter wheat growth period (see Table 1 and associated statements in the text) and validating the mapping results.
Because there are not similar site-level information for Kansas State, we identified winter wheat growth period in Kansas based on the USDA statistical data. The results validation for Kansas was based on USDA CDL (site-level validation) and USDA statistical data (state- and county-level validation).]
line 128: please define clearly what you call accuracy, PA, and UA (you can refer to the next sections may be),
[Response 10: Thanks for the reviewer’s comment. In the revised manuscript, we clearly defined the PA and UA (Page 5, Line 140-144).]
line 140: this is not clear to me what are the collected observations (is it only meteorological data ?)
[Response: The observations mean the phenological and location information. Several statements that we made were more ambiguous than intended, and we revised this sentence to be clearer (Page 5, Line 155).]
line 187: I guess you chose a quadratic polynomial for the S-G filter because this is common on other works ? Please cite these work,
[Response: Yes. We used quadratic polynomial for the S-G filter. We added two references (Doraiswamy et al. 2007; Li et al. 2014) in the revised manuscript as the reviewer’s suggested.]
Figure 3: very clear, thank you,
[Response: Thanks for the reviewer’s positive comment.]
References:
Campbell, J.B. Introduction to remote sensing, 3rd ed.; London: Taylor and Francis, 2003.
Foody, G.M., and Mathur, A. Toward intelligent training of supervised image classifications: directing training data acquisition for SVM classification. Remote Sensing of Environment. 2004, 93, 107-117
Hixson, M., Scholz, D., Fuhs, N., and Akiyama, T. Evaluation of several schemes for classification of remotely sensed data. Photogrammetric engineering and remote sensing. 1980, 46, 1547– 1553
Author Response File: Author Response.docx
Reviewer 4 Report
The approach in this study to map winter wheat along with intra-class variability is noteworthy and well executed. As the authors expect, this work could pave way of similar studies on various crops and help in improving the crop mapping efficiency. There are a few points that need to be better explained in the manuscript.
· Lines 89, 90: Location of Kansas, USA is given as a point, while the location of NCP is given as a region. A common notation for both would be good.
· Lines 252, 253: What is being considered here for accuracy? Are the areas under winter wheat from statistical data and the results of this study being compared?
· Line 304: How were the maps of this study compared with the CDL maps? Were their values at validation points compared? In such a case, the figure 99.2% does not give a correct comparison, as the two maps are at very different resolutions – 30 m and 250 m. A high correlation of 99.2% for two maps of such different resolutions is not practical.
· Figure 6: In the graphs, does each point represent a county? Please include this in the figure description.
Author Response
Response to Reviewer 4 Comments
The approach in this study to map winter wheat along with intra-class variability is noteworthy and well executed. As the authors expect, this work could pave way of similar studies on various crops and help in improving the crop mapping efficiency. There are a few points that need to be better explained in the manuscript.
[Response: Thanks for the reviewer’s positive and constructive comments.]
· Lines 89, 90: Location of Kansas, USA is given as a point, while the location of NCP is given as a region. A common notation for both would be good.
[Response: Thanks for the reviewer’s comment. We revised the description of Kansas by replacing “41' 52.6128'' N, 97° 18' 53.4060'' W” with “94 - 102° W and 37 – 40° N” as the reviewer suggested.]
· Lines 252, 253: What is being considered here for accuracy? Are the areas under winter wheat from statistical data and the results of this study being compared?
[Response: In this study, we considered comparing our results with the statistical data for accuracy. PE is the percentage error of winter wheat areas between the statistical data and our results. In the revised manuscript, we added two sentences to explain the accuracy. (Page 14, Line 313-316)]
· Line 304: How were the maps of this study compared with the CDL maps? Were their values at validation points compared? In such a case, the figure 99.2% does not give a correct comparison, as the two maps are at very different resolutions – 30 m and 250 m. A high correlation of 99.2% for two maps of such different resolutions is not practical.
[Response: Thanks for the reviewer’s comment. In this study, we evaluated our results against the CDL maps at both the state level and the site level. For the state level, we calculated and compared the total areas of winter wheat from the CDL maps and our results. The accuracy was 99.2%, which is the comparison between the CDL maps and our results for the total winter wheat areas of Kansas State.
For the site level, we randomly selected 600 points including the winter wheat and no-winter wheat locations from the CDL maps. We then calculated the confusion matrix to evaluate our results using these 600 points. As we mentioned in section 3.4, we resampled the CDL maps to 250m spatial resolution for the evaluation purpose.]
· Figure 6: In the graphs, does each point represent a county? Please include this in the figure description.
[Response: Yes, each point represents a county. In the revised manuscript, we included this in the figure description as the reviewer suggested. (Figure 7, Page 20, Line 414)]
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The authors have addressed the main issues that I mentioned in my review. They have explained the methods better. The manuscript provides a relatively minor contribution to the remote sensing technology of classifying winter wheat (or any other crop). But it is publishable at this stage.Author Response
Response to Reviewer 2 Comments
The authors have addressed the main issues that I mentioned in my review. They have explained the methods better. The manuscript provides a relatively minor contribution to the remote sensing technology of classifying winter wheat (or any other crop). But it is publishable at this stage.
[Response: Thanks for the reviewer’s comments.
The intra-class variability of winter wheat caused by field conditions and agricultural practices might lower the mapping accuracy, which has received little attention in previous studies. This study introduced an improved method considered the intra-class variability for winter wheat mapping by integrating the angles and distances of EVI time series multi-vector. Comparison with other methods showed that our approach improved the winter wheat mapping accuracies by 17% in Kansas and 15% in the NCP, respectively. We also conducted a comparison to evaluate the superiority of considering the intra-class variability (i.e., the approach considering intra-classes vs. the one without intra-classes). The results showed that winter wheat mapping accuracies were increased by 9% in Kansas and 19% in the NCP after considering intra-class variability, respectively (see Table 11). Our study demonstrated the advantages of introducing intra-class variability to winter wheat mapping. We believe that the research of this manuscript would further benefit the remote sensing-based crop mapping in future research.]
Author Response File: Author Response.docx
Reviewer 4 Report
1. The authors should justify the use of very low resolution (250m) MODIS data, despite the availability of much better resolution datasets (Landsat and Sentinel-2). How were smaller fields and mixed pixels accounted for?
2. The winter wheat maps of both study areas seem to have large, continuous areas under wheat. The authors should explain concisely about the remaining area too – what part of it is non-cropped and what part is under a non-wheat crop. Also, a comment on how well this method would work in a non-wheat dominant area could be included.
Author Response
Response to Reviewer 4 Comments
1. The authors should justify the use of very low resolution (250m) MODIS data, despite the availability of much better resolution datasets (Landsat and Sentinel-2). How were smaller fields and mixed pixels accounted for?
[Response: Thanks for the reviewer’s comments. MODIS data has become the primary dataset since 2000 for satellite-based crop monitoring at large scales due to its high temporal resolution, huge observation scope, and timely availability. Compared with Landsat dataset (56m or 30m), 250 m MODIS EVI product holds a better temporal resolution (16-day composite period) which can better capture the dynamic changes of crop growth. In addition, MODIS dataset possesses higher availability of the cloud-free images for large-area coverage than Landsat. Sentinel-2 offers global coverage every five days (two twin S2) with a high spatial resolution (10 m), but the satellite was launched in 2015 and data is only available for the past several years. Time-series MODIS 250 m EVI product provides considerable promise for crop mapping at large scale with these features. In the future, the integration of multi-source remote sensing data (i.e., image fusion technology) might improve the spatial and temporal resolution for crop mapping over regions with fragmented crop distribution.]
2. The winter wheat maps of both study areas seem to have large, continuous areas under wheat. The authors should explain concisely about the remaining area too – what part of it is non-cropped and what part is under a non-wheat crop. Also, a comment on how well this method would work in a non-wheat dominant area could be included.
[Response: Thanks for the reviewer’s comments. In both study areas, winter wheat is the major crop type (see section 2.1 Study Area). For Kansas, there are other crop types, such as corn, soybean, and sorghum. Corn, rice, soybean, and rapeseed are also distributed in the NCP region.
This study mainly focused on extracting winter wheat based on MODIS EVI time series data by using an improved approach that combines the winter wheat phenological characteristics and considers the intra-class variability (see Section 3 – method). Crop mapping involved other crop types acquires much more effort (such as further extracting the EVI time-series characteristics of these crop types combined with phenology information), which goes beyond the objectives of this study. We will expanse our approach to more crop types and quantify the effects of intra-class variability on the mapping accuracies (of other crop types) in the next step.
For the mapping accuracy over the non-wheat dominant area, we conducted a landscape analysis to examine the effects of the fragmentation of winter wheat distribution on mapping accuracies. We used the landscape metric PLAND, which represents the proportional abundance of winter wheat at the county-level. Our results demonstrated that when the PLAND is higher than 10%, the winter wheat mapping accuracy is more than 80% (the average percentage errors of winter wheat mapping < 20%). Especially, winter wheat mapping accuracies are above 90% under higher PLAND values (> 30%) (Table 10). The table also indicates that only when PLAND is lower than 1% (i.e., the winter wheat area accounts for less 1% at the county-level), the average winter wheat mapping errors increase to 59%. This analysis concluded that our methods performed well in winter wheat dominant areas (lower fragmented areas in our analysis) and most non-wheat dominant areas. Relatively lower accuracies occurred only when the winter wheat areas account less than 1% at the county level, which is acceptable for crop mapping at large scales.]
Author Response File: Author Response.docx
