Assessment and Comparison of Six Machine Learning Models in Estimating Evapotranspiration over Croplands Using Remote Sensing and Meteorological Factors

Round 1

Reviewer 1 Report

General comment:

The study evaluated six ML models for predicting crop ET against observed data from flux towers. The manuscript is well presented, and the findings could be useful for future researchers for exploring the potential of ML algorithms for estimating actual ET. However, the current manuscript version is missing key information in the methods section and requires additional analysis (Gs results/analysis). Please see specific comments for further details:

 

 

Specific comments:

Line 71: remove “What’s more” and reword the sentence

 

Line80-81: report the reason for estimating ET from cropland only, not from other land cover types

 

Line 80: report the temporal scale of ET: instantaneous, daily, monthly, annual?

 

Line 88-90: what were the criteria for selecting the five factors; especially, the use of carbon dioxide concentration and missing wind speed?  Report the temporal scale of these meteorological variables and time period (2015 only?)

 

Section 2.2.1: any filtering applied to quality control the flux tower data?  The use of bad quality flux tower data may affect the ML algorithms performances.

 

Table 1: A map would be a better representation

 

Line 152: R2 (coefficient of determination?) – used first time

 

Line 205: report the reason for applying AIC for ANN vs R2/Bias for other ML models

 

Line 159-261: reads like a comment but not results, remove this small paragraph

 

Line 274: report ET bias in water depths, e.g. mm/day not in W/m2 (usually used to report for latent/sensible/ground heat fluxes)

 

Fig 5. Add the unit for Bias

 

Line 320: remove “model” from “….Gs model,…”

 

Line 392-395: any reason behind the large difference in LSTM model performance between train/val. dataset vs test dataset?  How the dataset proportions (60%, 20%, 20%) were divided (random, temporal splitting, magnitude splitting)? 

 

Fig. 8: add the temporal graphs (ET predicted/observed vs time) for the better interpretation of ML algorithms performances. The best model provides less error (R2, RMSE) but also must capture the temporal ET variations, to be useful for practical applications

 

Line 526: the authors mentioned that LSTM is suitable for time-series data (line 215) but the LSTM model is the least performing ML model for time-series ET data, please discuss

 

Line 568: R2?

 

Results section: results section is missing Gs prediction results as the authors reported in the methods section.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper examined multiple AI methods to help estimate evapotranspiration, which was a good trial but had a crucial issue about the definition of evapotranspiration, that is, AET vs. ET0. This should be adjusted in the next round.

 

AET refers to the actual amount of ET, which can be measured by the eddy covariance. Without a flux tower, however, the AET can be calculated using a crop coefficient and the ET0 defined as the ET of well-irrigated hypothetical grassland under a specific meteorological condition. In fact, the PM equation is for the calculation of ET0. Therefore, the ET values colored in gray in Figure 1 should be an estimation of ET0, not AET. However, the estimated ET0 (by AI and PM equation) was compared with the is-situ AET (by eddy covariance) in Figures 8 to 10, which is not appropriate.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Well revised. Misunderstanding was solved.

Author Response

Thank you very much for the reviewer’ comments and suggestions on our manuscript.