Estimating Aboveground Biomass of Two Different Forest Types in Myanmar from Sentinel-2 Data with Machine Learning and Geostatistical Algorithms
Round 1
Reviewer 1 Report
Review of the manuscript “Estimating aboveground biomass of two different forest types in Myanmar from Sentinel-2 data with machine learning and geostatistical algorithms” by Wai et al.
The paper deals with the estimation of aboveground biomass (AGB) of forests in Myanmar, a country in the South-East of Asia which is involved in REDD+ programs. In particular, the aim of the work is to identify, among different methodological approaches which are exhaustively recalled in the introductory section, the best approaches to obtain based on high resolution satellite data (Sentinel-2 MSI) and ancillary information used as covariates. The numerous studies cited by the authors testify that this is a hot topic, highly investigated, with different levels of success. In this scientific context, the work contributes with an application of advanced statistical technique based on the use of high spatio-temporal remotely sensed data, in an under-investigated area. My comments/suggestions are made with the aim, on one hand, to clarify some aspects related to the application of remote sensing to forest classification, and on the other, to an easier reading of some sentences or figures.
The authors utilise two Sentinel-2 images, both of them are taken during winter (free from heavy rainfall I suppose, is it correct?); the use of imagery collected in different season might be helpful to separate evergreen and deciduous species. There are in fact evidences that the use of multi-temporal images can help capturing different phenological behaviours (examples can be found in Puletti et al., 2018, Persson et al., 2018, Axelsson et al., 2021….). Moreover, there are evidences that Sentinel-2 images collected during periods with low solar irradiation might have problems in the atmospheric correction due to the low sun inclination angle (Maselli et al., 2020); the authors are invited to consider and, if necessary, discuss this topic.
Regarding the result sections, I would suggest to shorten them following criteria: 1) avoid to include methodological explanations which might be deleted or moved to the methodology section, depending if these were not previously mentioned or not (see for instance lines 470-472 or lines 486-487); 2) avoid to repeat what is already clearly shown by the figures and/or tables (see for instance lines 466-469 and Figure 5 or lines 500-503 and Table 5).
The discussion section is very exhaustive and includes parts directly deriving from the achieved results and others (i.e sections 4.4 and 4.5) that help the reader to contextualize/interpret them. I appreciate also the conclusions which correctly point out the major findings of the manuscript.
Specific comments/suggestions are provided to contribute improving the quality of the manuscript, which is, in general, well written even if in some points there are minor English typos (e.g. line 217 “are shown”, line 245 “were”, etc.).
Specific comments:
- Line 20: the acronym SRTM should be specified the first time it is mentioned.
- Lines 36-37: some keywords (i.e. Sentinel-2 and AGB) might be removed because they are already mentioned in the title.
- Line 115-116: the sentence should be rephrased; what do you mean by “good correction”? Also, pay attention to the use of singular/plural nouns and verbs.
- Lines 184-185: maybe there are two typos in the scientific names: the correct ones should be “Macaranga denticulata” and “Ficus cuspidata”.
- Figure 2: the caption of this figure might be summarised avoiding repetitions. Maybe something similar might be enough: “Sentinel-2 L1C (a) and L2A (b) image of “DATA OF THE SCENE” with the corresponding spectral response curves of a vegetated pixel before (c) and after atmospheric correction (d).”
- Lines 300-302: avoid repetitions of the aforementioned equipment.
- Line 304: are subplots the aforementioned circular plots?
- Figure 4: if possible, improve the quality of this figure because, at least in my version, it is difficult to read all words. If I understand well from this figure, plots are sampled based on a regular grid, maybe you should include this important detail in the NFI description section.
- Line 311: what do you mean by “old dataset”? When does this dataset refer to? If I understand well, this is the first time the dataset is mentioned, maybe more information would be useful.
- Lines 317-324: only one general equation is used to quantify both evergreen and deciduous AGB. This is justified by a complete lack of information about some regions. Wasn’t it possible to detect tree volume and then compute AGB using basic wood density per evergreen and deciduous species?
- Line 462: correct the verb.
- Figure 8: please include the number of samples considered and the level of significance of the correlation.
- Line 506: replace with “show”.
References
Axelsson A. et al. (2021). Tree species classification using Sentinel-2 imagery and Bayesian inference. International Journal of Applied Earth Observation and Geoinformation, 100, 102318
Maselli F. et al. (2020). Use of Sentinel-2 MSI data to monitor crop irrigation in Mediterranean areas. International Journal of Applied Earth Observation and Geoinformation, 93, 102216
Persson M. et al. (2018). Tree species classification with multi-temporal Sentinel-2 data. Remote Sensing, 10(11), 1794.
Puletti et al. (2018). Use of Sentinel-2 for forest classification in Mediterranean environments. Annals Silvicultural Research, 42, 10.12899/asr-1463
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Major revision
This manuscript aims to consider the auto-correlation information of AGB and proposes a random forest based kriging (RFK) model to improve the AGB estimation of evergreen and deciduous forest using Sentinel 2 in Myanmar. The results show better performance in AGB estimation when using RFK. It is an interesting work to estimate the residuals of AGB using kriging based on the results derived from remote sensing. Still, there are some major problems to be addressed.
1. It is weird to have a larger nugget than sill for a semivariogram (see figure table 8), and different patterns of evergreen and deciduous forests in the study site. Moreover, a large ratio of nugget over sill means weak spatial correlation, which is contrary to the principle of auto-correlation.
2. The results of RF-Evergeen and RFOL-Evergreen show that kriging makes few contributions to AGB estimation for the evergreen forest. The reasons should be provided.
3. The second conclusion should be rephrased to differentiate from the description in the result part, not just list the results.
4. Non-normative figures:
In figure 4, the characters cannot be recognized. Similar problems can be found in the rest figures.
The description of the empty blue box is missing.
The alphabet and the position description are redundant for subfigure indication.
In figure 6, the legend of the color of the points distributed on the map is missing.
5. It is also suggested to illustrate the proportion of two types of forests in the dataset and results part. If there are two different models for two types of forests, the results (Figure 9a) should be produced based on the classification mapping of forests.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
It seems that the author does not take the major revision seriously. Although the author made some modifications, most of the problems still exist, and no extra reasons were given. For example,
- In Figure 3, the description of the empty blue box is missing.
- In Figure 9a, the bracket is not completed. Besides, the alphabet and the position description (i.e., a-upper left) are redundant for the subfigure indication. Similar problems can be found in figure 9b.
Importantly, the large ratio of nugget over sill (table 8) and range (i.e., 10747 meters) indicated that the kriging is not appropriate for the case study. The inconsistent results of the proposed method’s performance (table 6) and spatial correlation (table 8) also show the kriging does not work. Strong spatial correlation (0.73 nugget/sill) does not bring better results (0.47 R2), while less spatial correlation (0.75 nugget/sill and 9924 range) leads to better results (0.52 R2).
Besides, there are other problems to be addressed. For example,
- Table 6 lists the results of RFOK and RFCK, but the description lags behind the table, which is not a standard layout.
Consequently, I do not recommend this manuscript.
Author Response
Please kindly see the attachment.
Author Response File: Author Response.pdf
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.
