Prediction of Large-Scale Regional Evapotranspiration Based on Multi-Scale Feature Extraction and Multi-Headed Self-Attention

Round 1

Reviewer 1 Report (Previous Reviewer 3)

Comments and Suggestions for Authors

In my opinion the paper is very well organized and the improvements are well enough. The estimation of regional evapotranspiration throught the presented models are satisfactory exaplained and are statistically approved.

I think that the paper should be published.

 

Thank you

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (Previous Reviewer 4)

Comments and Suggestions for Authors

This manuscript presents a novel deep learning-based methodology for estimating monthly evapotranspiration within a specific region of China. It considers data from the past 12 months to predict evapotranspiration for the next month. The proposed method incorporates additional features, leading to slightly better results compared to existing techniques. This is a resubmission of a previous manuscript, in which several of the comments made in the first revision have been answered. Despite the improved explanation of the results, some additional clarifications would help the reader to understand the methods and the results. 

 

Comments: 

 

1. A general review of the text should be carried out, as there are some typing errors, especially in the newly added parts. 

 

2. In Table 2, in addition to the RMSE, the relative RMSE should also appear, for example 17.4 (xx %), since, in that way, the importance of the error with respect to the average values of the variable could be observed. 

 

3.  2.3.3 When introducing ConvLSTM it would be useful for the reader who is not so familiar with this type of system to clearly indicate the difference with CNN-LSTM, since in both cases convolution and short and medium term memory are used, but with a quite different structure.

 

4. Fig. 3 Also for users who are not so familiar with these methods, the meaning of the operations (C, O, …) should be indicated.

 

5. Lines 107-110: Please rephrase. For example: “the correlation between the correlation variable and ETa” ? Which variable? 

 

6. L 281 Particularly, the consistent lowest values ?

 

7. L 285. valuesin -> values in

 

8. L 304-310. On several occasions it appears  “predicting low elevation differences regions…” “predicting high elevation differences regions…” “predicting these regions”. I think that the regions are not predicted, but the values of ETa. “predicting ETa/values for/in these regions…”  

 

9. L 347. Fig 1 and Fig 6 show, The study area contains …???

 

10. L 353 and 355: “elevation above” … above what? Above mean sea level?

 

11. L 374: resultes -> results

 

12. L 404  the introduce of the SAM -> introduction? Adding?

 

13. L 542-544 :  The sensitivity of the variables has been studied, but an interpretability and explainability study has not been carried out (e.g. which variables and nodes of the input grid influence the prediction of ETa in some specific areas). I think that sentence should be deleted.

 

14. L 550 - 551 The study of the relationship between the input variables was eliminated in this version, that sentence should not appear now. 

Comments on the Quality of English Language

A general review of the text should be carried out, as there are some typing errors, especially in the newly added parts.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

Comments for author File: Comments.pdf

Author Response

Please 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.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Accurately predicting ETa at the regional scale is crucial for efficient water resource allocation and management. The establishment of regional scale ETa prediction techniques can help us understanding the role of ETa in hydrological processes. Deep learning is an emerging machine learning and has been widely used in remote sensing. In this manuscript, the authors have made an attempt to predict actual evapotranspiration based on remote sensing data and deep learning techniques. This manuscript explores the capability of existing deep learning-based actual evapotranspiration prediction mode for the first time and improves upon the original model. It was very interesting to go through the manuscript.

However, there are still some questions that need to be addressed before it can be published in this journal. I suggest a minor reversion for this manuscript before accepted for publication. My questions and comments are as follows.

1. The authors validated three existing deep learning-based methods, e.g. CNN-LSTM, ConvLSTM, SA-ConvLSTM. Why are these three methods were selected? Please give the detailed reason for methods selection.

2. Five parameters were used in the study to predict actual evapotranspiration. How were these five parameters determined? Why was the elevation mentioned in the study not used as one of the input parameters?

3. The first appearance of abbreviations requires a complete spelling.

4. The training of deep learning-based methods needs true labels. How were the true labels obtained? Please give the details.

5. Zoom out the Fig 9 and delete the squares with white color in this figure.

Comments on the Quality of English Language

Accurately predicting ETa at the regional scale is crucial for efficient water resource allocation and management. The establishment of regional scale ETa prediction techniques can help us understanding the role of ETa in hydrological processes. Deep learning is an emerging machine learning and has been widely used in remote sensing. In this manuscript, the authors have made an attempt to predict actual evapotranspiration based on remote sensing data and deep learning techniques. This manuscript explores the capability of existing deep learning-based actual evapotranspiration prediction mode for the first time and improves upon the original model. It was very interesting to go through the manuscript.

However, there are still some questions that need to be addressed before it can be published in this journal. I suggest a minor reversion for this manuscript before accepted for publication. My questions and comments are as follows.

1. The authors validated three existing deep learning-based methods, e.g. CNN-LSTM, ConvLSTM, SA-ConvLSTM. Why are these three methods were selected? Please give the detailed reason for methods selection.

2. Five parameters were used in the study to predict actual evapotranspiration. How were these five parameters determined? Why was the elevation mentioned in the study not used as one of the input parameters?

3. The first appearance of abbreviations requires a complete spelling.

4. The training of deep learning-based methods needs true labels. How were the true labels obtained? Please give the details.

5. Zoom out the Fig 9 and delete the squares with white color in this figure.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript introduces an innovative approach to predicting Evapotranspiration (ETa) by considering environmental variables and pixel spatial locations. The inclusion of SA-ConvLSTM and MulSA-ConvLSTM models, along with their sensitivity to varied climatic and topographic conditions, underscores the study's relevance. Significant contributions, especially in sensitivity analyses, are evident in the results. Publication is recommended with the implementation of suggested adjustments to enhance validity, interpretation, and scientific impact.

1. Introduction

Lines 82-93: Move this paragraph to the Materials and Methods section or remove it, as it is currently outside the context of the introduction.

Lines 94-104: I recommend a concise reformulation of this paragraph. The delineation of objectives, especially for objective number 1, is overly extensive. Clarifying them would enhance interpretability.

2. Materials and Methods

Table 1: The table presents information on the spatial resolutions of the products used in the study. However, ERA5-Land has a spatial resolution of approximately 9 km or (0.1° x 0.1°).  

See in this link https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land?tab=overview.

Lines 126-127. Specify which method was used for data remapping (resampling).

4. Discussion

4.3 Sensitivity Analysis of Features: Acknowledge that correlation doesn't imply causality. Recommend employing simple linear regression to quantify each variable's influence on ETa prediction, establishing more robust causal relationships beyond observed correlations.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In my opinion the idea of this paper is very good but the paper it is not easily readable. There are a lot of modules and abbreviations leading to misunderstanding. The coefficient R or R2 is not cleary referred. I mean that it is not clear where the reader can find the observed ET values. Is there any validation, and how? I am lost.

Also I should like to ask for a better syntax at lines 117-123, 125-126,157-161

Please correct ywhileàline 218

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

In this manuscript, a deep learning based method for estimating monthly evapotranspiration in a region of China (it predicts evapotranspiration of a particular month considering data of the previous 12 months) is presented. The proposed method builds on previous methods by adding some new features. The obtained results seem to slightly improve those of the other methods. The study seems interesting although the current writing makes it difficult to understand the results. I think that the work done in the development of the proposed system is important, therefore its validation and explanation of the results should be more adequate. 

 

Some aspects that, in my opinion, should be improved, studied in more depth or explained in more detail:

 

1) Data:

a) Lines 117-121 and table 1. There is an inconsistency, it is not possible to know which variables are obtained from ERA5-Land and which from TerraClimate. Those listed in the text and in the table do not match. 

 

b) Furthermore, the TerraClimate web site lists the available variables (Maximum temperature, minimum temperature, vapor pressure, precipitation accumulation, downward surface shortwave radiation, wind-speed). Why not use all the variables directly from TerraClimate as inputs to the deep learning systems? One could argue that this is an attempt to mimic a real case where actual evapotranspiration (observed, e.g. obtained from TerraClimate) is estimated from global model data. However, other variables are also obtained from TerraClimate…

 

c) TerraClimate data are interpolations of several datasets, taking into account orography, at a relatively high resolution. They are based on WorldClim data, which in turn uses CRU data. In my previous experience there are regions where the observations are very sparse and the interpolations, especially of precipitation, are quite different from other observations. It would be useful to know, in the study region, how many stations are used to calculate the gridded data. Perhaps it could explain the problems observed in the mountains. In addition, the used interpolation does not preserve the coherence between the different physical variables, since they are interpolated separately, so that the neural models may not be able to learn a pattern that does not really exist in the regions with complex orography.

 

d) Figure 2 refers to spatio-temporal resampling. The spatial one is explained in the text. What is the temporal one? Is it the average/accumulation of the ERA5 values to obtain the monthly data?

 

2) Results

a) Throughout the results section it is argued that the largest errors occur in the elevated areas. However, the worst results occur in the southernmost part of the region, both at high and low elevations (as shown in Figure 1).

The different statistics (MAE, RMS, bias, ) are not presented in relative form, but in absolute way, but the mean values are not presented, so it is not possible to know if the relative errors are high or low. Errors may be higher in the south because ETa is also usually higher in that area. I cannot deduce this from the data provided. It would be necessary to present and comment on the results in an appropriate way.

 

b) To study the effect of relief, the errors could be plotted on north-south transects or on some plot of the error versus terrain height. 

 

c) How is the time series in Figure 7 obtained? Is it for a particular grid point? Is it an average of the entire region or some subregion? The same applies to the results in Table 2, are they spatial averages? 

If both refer to spatial means of the same region, the average errors (RMSE) are of the same order of magnitude as the ETa observations (errors of about 23 mm/month and mean values, to a good eye, from Figure 7, of about 20 mm/month). Are the relative errors so large?

 

d) Lines 343-344 and Figure 7. It is argued that one of the models performs better than the other in some extreme months, but the fact is that the results of both are quite far from the observations, so it does not seem that the improvement is significant. For example for January 2018 both show an error close to 50%.

 

e) Table 3. Something is wrong with the number of parameters of CNN-LSTM. 

Perhaps the title “performance” is not the most appropriate. 

 

f) Throughout the entire text the authors use "elevation differences." Are they referring to areas with many elevation differences (complex orography) or simply higher areas?

 

g) Figure 9: What is the meaning of the square sizes? 

I understand that a correlation study is performed between the input variables to see if any of them are dispensable. 

I do not understand the reason for the study of the linear relationship between the input variables and the output (ETa), besides the correlations are performed directly and the partial correlations are not calculated. If we could take this assumption as valid, surely such a complicated, non-linear model would not have been used to estimate ETa.

 

e) Sensitivity analysis of features

If a downselection study is to be performed, more elaborate techniques should be used to explain the neural networks and the importance of the inputs. 

 

3) Model description

The description of the models is sometimes difficult to follow.

In the figures there are a multitude of subscripts that are not commented. The dimensions of the input data or of the different stages are not indicated. It is mentioned that the original grid is divided in "n" blocks, but it is not finally said how many blocks were used nor their size...

 

In summary, as I commented at the beginning, it seems an interesting study, but the explanation of the model, the data used, the presentation and discussion of the results, ... should be substantially improved to make it really useful to interested readers.

Comments on the Quality of English Language

The English writing is quite correct, although I could not understand some sentences and there are incorrect words that should be checked.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf