Extraction of Mountain Grasslands in Yunnan, China, from Sentinel-2 Data during the Optimal Phenological Period Using Feature Optimization

Round 1

Reviewer 1 Report

The manuscript presents a grassland classification study over Yunnan Province, applying Sentinel-2 data and using Google Earth Engine. The methodology implemented considers optimal phenological period identification and feature optimization for the classification. The topic of the manuscript is interesting but I believe that Introduction needs to be improved with a deeper analysis of the current state-of-the-art of the study topics (there are several studies about grasslands mapping and classification, including in mountainous areas) and of how this study advances and leverages the current knowledge. The manuscript falls short in explaining what's the novelty of this study. Also, in the sections of methodology, results and discussion there are several issues that are not well explained and/or grounded.

Regarding the manuscript presentation, major concerns are relative to:

(i)               In the Introduction section: It is my opinion that Introduction section needs to be improved. There are several publications on the use of RS for classifying grasslands and pastures. What already exists and what the present study innovates regarding the current state-of-the-art should be clear in the Introduction. The objectives of the study target a methodological innovation or simply an application of already existing methodological approaches to a geographical area with very specific characteristics? The Introduction must present a careful revision of the literature, making clear what are the gaps that are covered by the present study

(ii)              In the Results section: some of the results presented in Figures and Tables don’t seem to fit the arguments presented by Authors.

(iii)            In the Discussion section: some of the arguments discussed don’t seem quite aligned with the results

Detailed comments are provided below:

- Lines 40-41: It's not clear to me what Authors want to say with this sentence. The growing season and the senescent season aren't always distinct? What you mean is that, in some cases, one can have various grassland fields in the same region, which are in different phenological stages, but that's not the case in Yunnan Province? Please clarify

- Lines 47-50: It’s not clear to the reviewer why the Authors state that there is a need for accurate and timely assessment of optimal phenological period for extracting grasslands from remotely sensed imagery, in the following of citing the study by Royimani et al. (2022) where the most relevant vegetation indices for characterizing senescent grasslands in autumn were identified. Please clarify.

- Line 50: “there is a need of (...)”

- Lines 58-60: A reference should be included to support these sentences

- Lines 67-68: Please explain in the text what is Hughes phenomenon

-Lines 88-89: You are referring mountain areas in general or in Yunnan? There are various publications that use machine learning algorithms like RF for classifying grasslands/pastures in mountain areas, e.g.:

Filippa, G. et al. 2022. On the distribution and productivity of mountain grasslands in the Gran Paradiso National Park, NW Italy: A remote sensing approach, International Journal of Applied Earth Observation and Geoinformation, 108, 102718, https://doi.org/10.1016/j.jag.2022.102718.

 

Z. Guo, H. Liu, Z. Zheng, X. Chen and Y. Liang, "Accurate Extraction of Mountain Grassland From Remote Sensing Image Using a Capsule Network," in IEEE Geoscience and Remote Sensing Letters, vol. 18, no. 6, pp. 964-968, June 2021, doi: 10.1109/LGRS.2020.2992661.

 

Yuan, Y.; Wen, Q.; Zhao, X.; Liu, S.; Zhu, K.; Hu, B. Identifying Grassland Distribution in a Mountainous Region in Southwest China Using Multi-Source Remote Sensing Images. Remote Sens. 2022, 14, 1472. https://doi.org/10.3390/rs14061472

 

- Lines 124-125: Sentinel 2 has three visible bands with 10m resolution. Why did you use visible bands at 20m resolution and resampled to 10m? And also, you didn't use SWIR bands? Why not?

- Line 162: You indicate 3 key steps and then list 4

- Lines 191-192 and Table 1: Citations should be provided relative to the vegetation indices authors

- Lines 259-260: This sentence seems incomplete.

- Line 269: According to Fig.3, values are already decreasing in October, instead of being in the peak.

Lines 267-272: In line 269 Authors say “…reaching a peak in September to October, when the grassland was senescent“ and in line 271 “(…) and February to March was the senescent grassland season”. Information seems contradictory. Please clarify when is the senescent grassland season.

- Line 276 and Figure 3: The monthly NDVI values in Fig.3, for the various land cover types, show a very similar pattern for grasslands and croplands in the period February - March. The authors only refer in the text the differences relative to forests. Thus, it is not clear to the reviewer why this specific period was considered as the optimum time for extracting grassland. Between April and June, NDVI differences between grasslands and croplands are higher. Also, in line 277 Authors state that images from April were used as auxiliary in the classification. Thus, it is not clear what specific period of images was used. Please clarify.

- Line 289: SWIR band was not mentioned in section 2.2.1

- Figure 7: What exactly was the accuracy value obtained with 7 (and 10) features and with 21 features? The accuracy difference between 7 features (or 10) and 21 features seems to be very low, taking in consideration the high increase in the number of features. Why not using only the 7 features? In my opinion, so many additional features add too much redundancy without adding enough to the accuracy. For instance, aren't the various vegetation indices redundant (how do you prevent colinierity issues?

- Lines 342-352: I don't understand the results presented in this paragraph. In the features’ importance section, Authors say that 21 features were selected, according to importance score. If the Authors have a metric of the features’ importance, why then assessing classification accuracy based on combinations of features that ignore their importance score? Why are these combinations of features being analysed by groups of features (spectral indices, textural features, spectral bands)? What’s the OA if only using SWIR1, Blue, Elevation, Red Edge 1, Slope, Redc, and RVIc features, which were of features exhibiting higher importance scores?

- Lines 407-409: The observation of Fig. 10 doesn't show "accurate grassland extraction"; the accuracy is assessed by metrics. Fig. 10 allows a visual inspection of the results and an interpretation of spatial distribution

- Lines 430-431; 438-441: Despite the sentences is these lines, authors used a time period (February-March) when NDVI monthly values are similar for grasslands and croplands (Figure 3). Please explain.

-Line 455: The OA indicated in the results is 91.2%, not 96.7%

- Line 467: RVIc showed high importance score according to Fig. 8, when compared to EVIc, NDVIc, DVIc and MSAVIc. Considering that all these indices (NDVI, EVI, DVI, RVI and MSAVI) are obtained with the same spectral bands (but with different formulations) aren't they highly correlated?

- References section: References formatting should be uniform

Author Response

My reply is attached

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript is interesting, but at the same time needs some improvement.

Text formatting needs to be improved.

Abstract: not clear what is the source of textural and topographic features? From the abstract, it seems that all data comes from Sentinel-2. I would add user accuracy, not only producer accuracy.

39-51 the literature review regarding remote sensing techniques used for grassland analysis is quite short.

58-61 the definition of parametric and non-parametric classification is not clear; also I do not think they are necessary

61 why these two algorithms are mentioned?

61-63 this is one of the key elements

63-64 I do not agree. There are a lot of studies, which use data fusion

66 – reference at the end of a sentence is missing

52-75 This part of the text seems to be prepared based on too few scientific materials. There are no studies from North America or Europe, where classifying grassland is quite commonly used.

102-107 the aim of the study is not clear

Figure 1 – the grid is missing

126 – Was the resampling performed before the corrections? What method of resampling was done?

126-128 The images from Sentinel-2 are normally available as data radiometrically and geometrically corrected. What was the source of Sentinel-2 images? Also, using Sen2Cor only atmospheric correction can be performed.

129-141: The 318 points were verified during field research; what about the rest?

142-147: What spatial resolution was used? 30m?

Figure 2: too small font on the figure. What is the source of terrain data? From the graph, it seems that the optimal feature combination procedure does not involve any verification data. Do all validation and training datasets come from the field survey? I understood that only 318 pints are from field measurements. Grassland distribution is not the result of accuracy assessment, but the result of the classification.

175-184 – I do not understand, how the Authors determine the optimum phenological period. What were the values of NDVI and why are these?

191-193 Why these spectral indices were used?

Table 1 – reference of the indices are missing

208 What kind of resampling was performed?

219 the equation contains symbols, that are not explained in the text.

227-233 The procedure is not clear.

246  From the classification description it seems that the authors assumed, that RF is the best method of classification. If so, why did you have tested the other once? If four methods were tested, please describe each one in the same way and then explained, how the methods were compared. N what basis RF was chosen as the best one?

258-262 In my opinion cross-validation should be performed. Was it done?

264-282 The optimal period selection was described as a stage of the analysis, not the results. Please decide if this information is results, then change the method section, or only method – also change method and remove information from results. The same comment can be applied to lines 425-441.

Figure 4 – what is on the image: spectral bands, RGB composition, indices?

283-301 This section does not describe any type of optimization, there is an only comparison of spectral indices.

Figure 8 Unreadable. The legend is poorly constructed, the same colour is for two values. The chart should not be circular, but a bar chart.

Table 2 Are there presented two different datasets? It is not clear.

302-324 Why feature importance was used? Only 29 bands were used for the classification, which is not a high value. The final dataset consisted of 21 bands, which is only 8 less than the original.

Table 3 Add the number of features used for each dataset.

Section 3.4. Why are the RF results presented first, instead of four classifiers? This is incomprehensible. In the description of the results, there is information about the mixing of classes other than grassland. If the study aims to detect grassland, this information is not needed. More important is what classes the grassland is mixed with.

 Table 3 More informative would be adding confusion matrices. The information about PA and UA concerning other classes than grassland is not necessary.

442-491 There is a long description of feature optimization, but, based on described method and results, there was no optimization performed.

Why in the discussion there is no comparison of the accuracy of the classification results with other studies?

Author Response

My reply is attached.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The comments are in the attached file.

Comments for author File: Comments.pdf

Author Response

Our response is attached.

Author Response File: Author Response.docx