Leveraging Google Earth Engine and Machine Learning to Estimate Evapotranspiration in a Commercial Forest Plantation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper, the authors have put forth a comprehensive study leveraging modern technologies like Google Earth Engine and machine learning to predict crop coefficients and ET for a commercial forest in South Africa. On the whole, the manuscript is well-structured with a clear exposition of the methodology and a logical progression through the results. The use of ensemble machine learning models to integrate leaf area index (LAI) data for ET estimation is methodologically sound and the results show promising levels of accuracy.

 

However,  the manuscript does not make a significant contribution to the field from an ecological or innovative standpoint. The core approach largely hinges on the application of existing algorithms to existing datasets, without a substantive development or adaptation of these algorithms. This limits the novelty of the research. Moreover, while the results are relevant for the specific study area (regional scale), their generalizability and applicability to other regions or ecosystems are not convincingly demonstrated. In this case, the algorithms' performance might vary significantly under different ecological conditions or with different types of vegetation.

 

Please see other detailed comments below:

 

1. Diagrams and Figures: remove specific references to software or programming languages in the flow chart. 

2. Introduction Section: The paragraphs concerning data and methods within the introduction could be integrated into fewer paragraphs.

 

Author Response

Please refer to the attachment for responses to you comments and queries.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Estimation of evapotranspiration based on reference evapotranspiration and the crop coefficient remains one of the most widely used ET estimation approaches, yet its application in non-agricultural and natural environments has been limited. Authors utilized satellite-derived estimates of LAI acquired through the Google Earth Engine geospatial cloud computing platform and machine learning to quantify the water use of a commercial forest plantation situated within the eastern region of South Africa. The topic is interesting and important, but I think some shortcoming should prevent publication of the paper in its current form. In fact, the English writing, the presentation of the state of the art, the structure of analysis, and the discussion of results can be largely improved. Hence, my recommendation is for major revision. What follows are general comments and specific remarks, which I sincerely hope to be useful for the authors.

General comments:

1.  A careful and thorough revision of the English writing must be performed by the authors. There are so many writing errors and confusing statements that it became impossible to point them out individually. The authors should also ensure that the correct statistical terminology is adopted throughout the paper.

2.  2.2 Data collection and processing does not provide proper background to the readers. In fact, the logic of the method is vague. The connection between the topics and methods presented in this section could be improved.

3.  The organization of the manuscript could also be improved. The discussion section is too short and mainly comprises literature review and should be entirely rewritten. The discussion could also be improved by clearly presenting the limitations of the study and envisaged research developments, instead of merely repeating some results.

4.  Figs.3 and 4 are poor and should be redone.

Comments on the Quality of English Language

Estimation of evapotranspiration based on reference evapotranspiration and the crop coefficient remains one of the most widely used ET estimation approaches, yet its application in non-agricultural and natural environments has been limited. Authors utilized satellite-derived estimates of LAI acquired through the Google Earth Engine geospatial cloud computing platform and machine learning to quantify the water use of a commercial forest plantation situated within the eastern region of South Africa. The topic is interesting and important, but I think some shortcoming should prevent publication of the paper in its current form. In fact, the English writing, the presentation of the state of the art, the structure of analysis, and the discussion of results can be largely improved. Hence, my recommendation is for major revision. What follows are general comments and specific remarks, which I sincerely hope to be useful for the authors.

General comments:

1.  A careful and thorough revision of the English writing must be performed by the authors. There are so many writing errors and confusing statements that it became impossible to point them out individually. The authors should also ensure that the correct statistical terminology is adopted throughout the paper.

2.  2.2 Data collection and processing does not provide proper background to the readers. In fact, the logic of the method is vague. The connection between the topics and methods presented in this section could be improved.

3.  The organization of the manuscript could also be improved. The discussion section is too short and mainly comprises literature review and should be entirely rewritten. The discussion could also be improved by clearly presenting the limitations of the study and envisaged research developments, instead of merely repeating some results.

4.  Figs.3 and 4 are poor and should be redone.

Author Response

Please refer to the attachment for responses to you comments and queries.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript proposed to use remote sensing data and machine learning approach to estimate evapotranspiration (ET) in a non-agricultural and natural vegetation region. A number of machine learning approaches were investigated to model the relationship between satellite-based LAI and measured crop coefficient. The estimated crop coefficient was subsequently used to estimate ET by multiplying the reference ET. The ensemble machine learning model (EMLM) was demonstrated to provide the highest accuracy when compared with ground-based measurements. The manuscript is well organized and written clearly. Figures and tables are used appropriately. Some minor comments are given below:

 

1.     Could you please clarify why was LAI only selected for Kc estimation? Other variables such as precipitation and temperature are also related to ET, and their data can be easily acquired from either satellites or reanalysis data.

2.      Although discussed in the Discussion section, I am still worrying about the no water limit assumption made in the study. I wonder if there is any method to test the establishment of the assumption in the study, or to what extend the violation of the assumption affect the results?

3.     Could you please explain why the Acacia mearnsii is selected particularly for analysis? What about other species?

4.     Lines 301: what is the possible reason for the overestimation in September and October, and the underestimation in other months?

5.     Line 79: “respiration [23, 25] Since” “respiration [23, 25]. Since”

Author Response

Please refer to the attachment for responses to you comments and queries.

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

The presented study uses machine learning and satellite-derived leaf area index observations to estimate ET in a South African commercial forest plantation. The study introduces the prediction of crop coefficient using by integrating remote sensing and data driven methods to calculate ET. Although the topic is very interesting the adopted methods and presented results are limited.

Comments:

- Line 36: I would use "specifically" instead of "vis".

- Figure 1: Please include the source of the background map used in the main map and the aerial view insert.

- Line 161: What are the image quality filtering methods employed in this study? Readers would be interested in additional details about this procedure.

- Line 166: Are the authors referring to the mean value of LAI? Was the spatial average calculated using arithmetic mean? If yes, please explain why not using a method that takes into account the heterogeneity of the land cover in the study area.

- in the "Data Collection and Processing" section, Please clearly indicate the study time period in the text.

- Figure 2: Under the in-situ data acquisition, authors mention long-term historical measurement (2007-2013). However, they used the k-fold cross-validation approach for the ML training due to the relatively small size of the datasets. why using K-fold in the study uses a long term dataset? MODIS observations are available from 2002 to date, why not expending the study period?

- Figure 2: How was the LAI retrieval aggregated to a monthly scale?

- Google Earth Engine usage was limited to data filtering and download. I don't think the title should include GEE, as long as the platform is used for limited processing.

- The method section should include a description of the LAI retrieved from MODIS observation with relevant references.

- Evaluation metrics should include at least another metric that shows if the model is underpredicting or overpredicting the target variables, e.g., percent bias.

- Table 2: Please clarify the p-value presented in Table 2. Is it representing the significance level of which test? The high value of the p-value raises a flag about the adopted methods and shows that the results are not statistically significant.

- The results section is poorly developed. Comments about the model's performance across seasons are missing. Why is the model overpredicting in February and underpredicting in December, for example?

- The term "cloud computing" was used. The study is not leveraging cloud computing capability. Please revise this term.

Comments on the Quality of English Language

Moderate editing of English language required

Author Response

Please refer to the attachment for responses to you comments and queries.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

It can be accepted for publication at this stage.

Author Response

Thank you for your careful reading of the manuscript and excellent recommendations which have contributed to an improved representation of our research in the revised version of the manuscript. 

Reviewer 4 Report

Comments and Suggestions for Authors

- Follow-up on Comment 2: If the map is generated using ArcGIS Pro, it is necessary to indicate the source of the base map. Alternatively, this information should be included in the figure's caption to maintain clarity.

- Follow-up on Comment 3: Please clarify the methods used for handling cloud-covered images (if the level 4 product does not exclude such scenes). Please clearly indicate in the text if the LAI does not include cloud covered images.

- Follow-up on Comment 6: Please incorporate the answer from comment 6 directly into the text. This clarification is essential for providing a comprehensive understanding to all readers.

- Follow-up on Comment 10: The provided metrics (MAE and RMSE) do not specify whether the model tends towards over or under prediction. Please expand on the model's performance using the existing metrics and other metrics (Percent Bias for example), noting that this section currently offers limited insight.

- While a satisfactory agreement was obtained from Figure 3, it would be beneficial to comment on how this agreement varied through different seasons/months. Was the model more accurate during the dry or wet season? Detailing this performance in relation to LAI temporal patterns will provide deeper insight into the model's effectiveness across seasons and LAI conditions/patterns.

- In lines 203-204, the use of the Kruskal-Wallis significance test for assessing uncertainty should be supported with appropriate equations. Please add details of the applied test and explain the use of the test results to infer uncertainty.

- Finally, regarding the usage of Cloud computing terminology, I suggest expanding the scope of the introduction. Authors should expand the introduction to include examples of leveraging Google Earth Engine in earth science applications such as: real-time monitoring of natural hazards, climatological analysis, and water resources management. Building on previous studies, authors should introduce the novelty of this study-using GEE/cloud computing resources to advance the estimation of evapotranspiration in a commercial forest plantation.

Author Response

Please see responses in the attached document.

Author Response File: Author Response.docx