A Comparison of Different Machine Learning Methods to Reconstruct Daily Evapotranspiration Time Series Estimated by Thermal–Infrared Remote Sensing
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper compared four machine learning methods in gap filling the modeled daily ET from TSEB model. It is an interesting paper and can help us improve our understanding in the daily ET generation using satellite observations. Although the paper is generally well-written, it requires additional refinements before publication. I suggest minor revision. My comments are as follows.
1. line17, please provide full name of 'RF'
2. line 19-20, It is hard to understand this sentence, who yields gaps?
3. line 23, do them performed reliable, they can replace the physical based models ?
4. line 34, Please add 'several' before meters
5. line 34, what is the difficulty in larger scales?
6. line 70, Why these machine learning methods are rarely used to generate ET?
7. Line 103, Please provide the web page of TPDC
8. It mentioned TSEB model can yield reliable ET, which really can not generate ET due to the absent of land surface temperature.
9. Why they can be used to generate daily ET?
10. What are the training target of these machine learning methods?
11. Please provide the reference of TCH method!
12. line 238-242, what do the 'all samples' and 'reconstructed samples' mean?
13. Line 249-253, It has described that the RF performed the best in reconstructing daily ET under cloudy conditions!
14. Line 276-279, But the validation using ground measurements shows the DF yields highest errors!
15. line 291-294, Please delete the later 'different machine learning methods', it is repeat in the sentence.
16. Line 294-298,what does 'period' mean ? for what season?
17. What is the object of the figure 12? it showed the spatial distributions of ET in different season, what can we learn from these explanations?
18. line 328-329, why the TSEB estimated ET requires gap filling ?
19. Line 336-339, There is only one reference by it used plural form in the sentence.
20. what is the difference with Cui 2021?
21. Actually, all of the four machine leaning methods have very similar accuracy in this study.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors coupled TSEB model and Machine Learning methods for reconstructing daily ET rates. The topic is relevant for journal readers and in my opinion, the manuscript shows some potential. However, I found several concerns so I kindly ask to the authors to improve their study. Please see my comments/suggestions below:
Title
1. Lines 2-4. I would suggest modifying the title to: “Comparison of Different Machine Learning Methods to Reconstruct Daily Evapotranspiration Time Series Estimated by Thermal-Infrared Remote Sensing”. But this is just a suggestion.
Abstract
2. Line 9. I would specify you are referring to optical remote sensing-based models. In fact, microwave remote sensing is not affected by cloud cover conditions.
3. Lines 11-13. Not completely agree. There are several studies in the last years about the comparison among machine learning (ML) methods for reconstructing evapotranspiration rates.
4. Lines 13-14. Please specify here the acronyms (RF, DF, DNN, XGB) you use later in the abstract.
5. Lines 18-19. “…but also reconstructed ET throughout the basin”. Not clear in my opinion.
6. Line 24. Regional or basin scale?
Keywords
7. Line 25. I would avoid keywords that are in the manuscript title and acronyms.
Introduction
8. Line 54. You should also include a paragraph about meteorological parameters. Specifically, in your study, you are using climate reanalysis (i.e. ERA5-Land), and in my personal opinion, this is a kind of innovation, since you are overcoming the main limits of the observed meteorological data recorded by the traditional weather stations. In fact, the use of climate reanalysis guarantees to have continuous and spatially distributed meteorological estimates.
9. Lines 73-79. I would highlight better the innovation of this study and how you are overcoming the current issues. In my opinion, as far as I understood by reading the manuscript, the general aim of the study should be the generation of a continuous ET time series, overcoming the spatial issue of the traditional measurement systems and the temporal issue of the remote sensing models (e.g. due to satellite revisit period, cloud conditions, etc.). The specific aim is the implementation of different ML methods and their accuracy and effectiveness evaluation.
Materials and Methods
10. Lines 129-134. I would detail better how you managed ERA5-Land variables. Specifically, the native ERA5-Land dataset does not directly provide the wind speed (WS) and the relative humidity (RH), but the two components (U and V) of WS at 10 m of height, and the air and dew point temperatures at 2 m of height.
11. Lines 135-137. In my opinion, you could provide a Table with all the data used in the study and their native (spatial and temporal) resolution. Moreover, I would homogenize the unit of the spatial resolutions since you are currently using metric and degree systems.
12. Line 145. I would highlight that at the end of the process, you are obtaining a reconstructed daily ET time series.
13. Figure 2. Please, for an easier and faster understanding, avoid using acronyms in the figure.
14. Line 151. I would add the equation of TSEB model for a better understanding of how your input data are used in this model.
15. Lines 152-163. For which and how many dates did you apply the TSEB model? From paragraph 2.2, based on your input, perhaps I get that is a daily application, so you already have a daily ET time series. If so, which gaps are you reconstructing with ML methods? In my opinion, the “temporal” scale of your study is not completely clear.
16. Lines 165-166. How many gaps do you have in the ET time series estimated by TSEB model? As said in the previous comment, you should specify better this information.
17. Lines 176-188. These are preliminary results of the study, so I would move this paragraph into the result section.
18. Line 177. Are these correlations statistically significant? First of all, you must evaluate the statistical significance (in terms of p-value). After that, how are you using the results of the correlation matrix? Are you removing the non-correlated variables? What about the weak correlations? Are you fixing a Pearson correlation coefficient threshold for selecting the variables or not? Please try to face these key points in your study.
19. Figure 3. Please increase the font size of the variable names. Probably the use of the full names is not possible for a problem of space, but eventually use the same acronyms you are using in the text of the manuscript, some of them are not the same.
20. Lines 190-192. How many ground measurements are you considering for validating the ET estimates? You reported using a time range from 2012-2015 and 2013-2016 depending on the station (Table 1), but in my opinion, a clear statement with the number of measurements considered is missing (please see also the comments n. 15-16).
21. Line 229. Please check this line.
Results
22. Lines 237- 264. First of all, the difference between “generated” and “reconstructed” ET, is not clear. Please specify better (maybe in the paragraph 2.4).
Additionally, you are using linear regression for validating the generated/reconstructed ET with respect to the observed. But what about the statistical significance? You could include the p-value of the regression, as well as the number of observations.
23. Line 290. Please check the sentence (double “after TSEB estimation”).
24. Lines 291-294. Is this evaluation done at a yearly temporal scale? If so, why did you show the results for only 2016 (Figure 11)?
25. Line 296. Which period?
26. Lines 303-308. In my personal opinion, this paragraph does not show relevant results. It is just describing the general and well-known behavior of ET during the seasons and among the different land cover types.
Discussion
27. Lines 355-367. In my opinion, you should move this paragraph into the result section and introduce the SHAP method into the material and method section. Additionally, be careful with the term “significant”, since no statistical significance evaluation has been done. Moreover, which average impact threshold value should be used for considering whether a parameter is important or not?
Conclusion
28. Line 419. Which scales?
29. Lines 416-430. The conclusions section as it is currently, is a summary of the main results of the study. In my opinion, it should include some statement about the implications of the findings and how it may be extended.
Author Response
Thank you very much for taking the time to review this manuscript. We revised our manuscript.
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsI thank the authors for having responded to my concerns. I only have 1 minor comment:
- Lines 50-54 (new version of the manuscript): I would add some references (about climate reanalysis) for supporting this statement.
Author Response
Thank you again for taking the time to review this manuscript. Please find the detailed responses below.
Comments: Lines 50-54 (new version of the manuscript): I would add some references (about climate reanalysis) for supporting this statement.
Response: We added references for supporting this statement.
‘The input parameters of TSEB model include surface boundary parameters based on remote sensing and meteorological parameters [6]. Meteorological reanalysis data overcome the spatial limitations of the observed meteorological data recorded by traditional weather stations, and can be employed to drive TSEB model at large scales [20,21].’(L50-54)
References:
Norman, J.M.; Kustas, W.P.; Humes, K.S. Source Approach for Estimating Soil and Vegetation Energy Fluxes in Observations of Directional Radiometric Surface-Temperature. Agr Forest Meteorol 1995, 77, 263-293, doi:10.1016/0168-1923(95)02265-Y.
Amjad, M.; Yilmaz, M.T.; Yucel, I.; Yilmaz, K.K. Performance evaluation of satellite- and model-based precipitation products over varying climate and complex topography. Journal of Hydrology 2020, 584, 124707, doi:016/j.jhydrol.2020.124707.
Muñoz-Sabater, J.; Dutra, E.; Agustí-Panareda, A.; Albergel, C.; Arduini, G.; Balsamo, G.; Boussetta, S.; Choulga, M.; Harrigan, S.; Hersbach, H.; et al. ERA5-Land: a state-of-the-art global reanalysis dataset for land applications. Earth System Science Data 2021, 13, 4349-4383, doi:10.5194/essd-13-4349-2021.
