Near Real-Time Estimation of Daytime and Nighttime Evapotranspiration Using GOES-R Observations and Machine Learning Models

Ranjbar, Sadegh; Losos, Danielle; Hoffman, Sophie; Zhong, Yafang; Otkin, Jason A.; Desai, Ankur Rashmikant; Anderson, Martha; Hain, Christopher R.; Stoy, Paul Christopher

doi:10.22541/essoar.174792936.66373305/v1

Preprints

https://doi.org/10.22541/essoar.174792936.66373305/v1

Preprints

01 Dec 2025

| 01 Dec 2025

Status: this preprint is open for discussion and under review for Hydrology and Earth System Sciences (HESS).

Near Real-Time Estimation of Daytime and Nighttime Evapotranspiration Using GOES-R Observations and Machine Learning Models

Sadegh Ranjbar, Danielle Losos, Sophie Hoffman, Yafang Zhong, Jason A. Otkin, Ankur Rashmikant Desai, Martha Anderson, Christopher R. Hain, and Paul Christopher Stoy

Abstract. Evapotranspiration (ET) is a critical component of the water cycle, influencing climate, agriculture, and water resource management. However, most satellite-derived ET products are limited to daily or coarser temporal resolutions, despite the strong diurnal variability of ET processes. Existing satellite-based ET retrievals are largely restricted to daytime conditions, when nighttime ET is a small but often non-trivial flux. In this study, we introduce the Advanced Baseline Imager Live Imaging of Vegetated Ecosystems ET (ALIVE_ET), a near real-time, five-minute ET estimation framework, leveraging geostationary satellite observations from the GOES-R Advanced Baseline Imager (ABI) and machine learning models under both clear and cloudy conditions. We test Gradient Boosting Regression (GBR) and Long Short-Term Memory (LSTM) models to assess their ability to estimate ET variations across the diurnal cycle. GBR captures daytime ET with an R² of 0.74 (RMSE of 0.059 mm hh^-1 equivalent to about 74 W m^-2) while maintaining low computational cost. For nighttime ET, where R² decreases by about 0.50 compared to daytime, LSTM models trained on time-series observations perform better, achieving an R² of 0.24 (RMSE of 0.014 mm hh^-1) by leveraging temporal dependencies in land surface temperature (LST) and past ABI observations. Comparisons against daily ET estimates from the physically-based ALEXI remote sensing model demonstrates good agreement but opportunities for improvement. This study demonstrates the potential of integrating machine learning with geostationary remote sensing to advance high-temporal-resolution ET estimation.

Received: 08 Sep 2025 – Discussion started: 01 Dec 2025

Sadegh Ranjbar, Danielle Losos, Sophie Hoffman, Yafang Zhong, Jason A. Otkin, Ankur Rashmikant Desai, Martha Anderson, Christopher R. Hain, and Paul Christopher Stoy

Status: open (until 17 Feb 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2025-4400', Marloes Mul, 29 Dec 2025 reply
Dear authors,

I read your manuscript “Near real-time estimation of daytime and nighttime evapotranspiration using GOES-R observations and machine learning models” with much interest. It illustrates an interesting approach towards diurnal ET estimations for the CONUS region and it is an overall well written manuscript. I do have a few comments and suggestions for improvement as provided below.
Use consistent terminologies: in several figures you refer to ET estimations from EC towers as “calculated ET”, which to me is a bit confusing. In figure 2 you call it EC-derived ET, but in figure 3, the same (?) dataset is referred to as calculated ET, I think EC-derived ET is a better description.

Add number of observations per station used in the annex

Line 251: you used the normalised RMSE as an indicator, an alternative is to use the relative RMSE (divided by the median or mean instead of the maximum), this indicator is less influenced by extreme values (and is also a more used performance metrics (see figure 13, Tran et al 2023).

Line 264-269 seems to fit better in a discussion section (reflection on the computation time)

Figure 2: add number of observations presented in the figure (n=..)

Table 3: what does “prediction time” mean (called prediction speed in line 238- check consistency)? Also is this result based on the calibration or validation dataset (and how is it different for validation vs calibration)?

Figure 3: How is the day-time/ night-time defined? It seems the transition from day to night and night to day period is the most tricky one (and does this affect the training of the ML and in the end the performance of the model?). Also there seems to be an overlap in the night time – started at 4PM and end 8AM (perhaps related to winter? but this is not visible in the daytime model, which should have then have included the longer evenings?). The unit is in half hour, but the graph only presents hourly data points, I would suggest to make this consistent. Caption: what does “local hour” mean?

Line 301, do you mean the one year time series is an average across all sites, or one example year for one stations or??

Figure 3&4, how did you calculate the daily ET, did you use the two different ‘best’ models? How did you deal with the transition hours (between daylight and night time)?

Figure 6, why did you combine certain climate classes and not include the BSh climate class? How many stations are included in each climate class?

Figure 7, why were the other land cover classes not included?

Line 455: the bias comment is not really substantiated. The values during the night are generally much lower than at day time and the difference may look larger in figure 3, but this is not quantified. I would suggest to include bias as a performance indicator and add it to table 3 to support this comment?

Line 515 since you are explaining the results per climate, is the comment related to the vegetation cover really relevant here?

Line 517-520, but would you expect that the model would perform better at these locations if it is specifically trained for those conditions?

Line 571: which result show that ALIVEet underestimates ET for ‘high vegetation density and complex moisture dynamics’?

Line 573: which result show that ALIVEet ‘struggles to capture ET dynamics in the peak growing season’?

Edits:
Line 144, remove brackets from the reference.

Line 181, reduce number of names from the reference (check referencing style of journal)

Line 194 remove initials from reference (check referencing style of journal)

References:
Tran, B. N., van der Kwast, J., Seyoum, S., Uijlenhoet, R., Jewitt, G., and Mul, M.: Uncertainty assessment of satellite remote-sensing-based evapotranspiration estimates: a systematic review of methods and gaps, Hydrol. Earth Syst. Sci., 27, 4505–4528, https://doi.org/10.5194/hess-27-4505-2023, 2023.

Reply
Citation: https://doi.org/10.5194/egusphere-2025-4400-RC1
- AC1: 'Reply on RC1', S. Ranjbar, 13 Jan 2026 reply
  
  Thank you for your comments. We have carefully reviewed them and addressed each of your concerns. Please find our detailed responses below. We will upload the revised manuscript once we receive instructions from the editorial office to do so.
  
  Opinion: Dear authors, I read your manuscript “Near real-time estimation of daytime and nighttime evapotranspiration using GOES-R observations and machine learning models” with much interest. It illustrates an interesting approach towards diurnal ET estimations for the CONUS region and it is an overall well written manuscript. I do have a few comments and suggestions for improvement as provided below.
  Response: We sincerely thank the reviewer for the thoughtful and thorough review of our manuscript. We found the comments to be highly valuable and essential for improving the quality and clarity of the paper. We have carefully addressed all suggestions and believe the revisions have improved the manuscript.
  
  Comment: Use consistent terminologies: in several figures you refer to ET estimations from EC towers as “calculated ET”, which to me is a bit confusing. In figure 2 you call it EC-derived ET, but in figure 3, the same (?) dataset is referred to as calculated ET, I think EC-derived ET is a better description.
  Response: Thank you for this helpful comment. We agree that consistent terminology is important for clarity. We have revised the manuscript to use the term “EC-derived ET” consistently across the entire manuscript, including all figures and figure captions, to refer to evapotranspiration estimates from eddy covariance towers that result from latent heat flux measurements. This change has been implemented to avoid confusion and to improve clarity for the reader.
  
  Comment: Add number of observations per station used in the annex
  Response: We have added the number of observations per station to Table A1 (Annex).
  
  Comment: Line 251: you used the normalised RMSE as an indicator, an alternative is to use the relative RMSE (divided by the median or mean instead of the maximum), this indicator is less influenced by extreme values (and is also a more used performance metrics (see figure 13, Tran et al 2023).
  Response: Thank you for this valuable suggestion. We agree that using a relative RMSE normalized by the median or mean is less sensitive to extreme values than normalization by the maximum. Following this recommendation, we replaced the originally used normalized RMSE with nRMSE normalized by the median value. This choice reduces the influence of extreme values, improves robustness across conditions, and provides a more meaningful comparison, particularly for variables with strong diurnal variability and lower nighttime magnitudes. The manuscript text, figures, and tables have been updated accordingly to reflect this change.
  
  Comment: Line 264-269 seems to fit better in a discussion section (reflection on the computation time)
  Response: Thank you for this comment. We carefully considered moving Lines 264–269 to a Discussion section. However, we opted to keep this part in its current location because it strictly reports computational performance results (i.e., training and inference time and hardware usage) without interpretation or broader discussion. Since no reflection or conceptual analysis is provided in this paragraph, we believe keeping it within the Results section helps avoid confusion and maintains a clear separation between reported results and subsequent discussion.
  
  Comment: Figure 2: add number of observations presented in the figure (n=..)
  Response: We have revised the caption of Figure 2 to include the number of observations. Specifically, the figure now states that 17480 points are used for the daytime plot and 14304 points are used for the nighttime plot.
  
  Comment: Table 3: what does “prediction time” mean (called prediction speed in line 238- check consistency)? Also is this result based on the calibration or validation dataset (and how is it different for validation vs calibration)?
  Response: Thank you for this helpful comment. We have revised the manuscript to use the term “prediction time” consistently throughout the text and tables. We also clarified its definition in the manuscript. The reported prediction time is for the entire validation dataset, which allows for a consistent and fair comparison between models, particularly in the context of near real-time prediction applications. We use the validation set to better reflect operational performance during real-time deployment.
  
  Comment: Figure 3: How is the day-time/ night-time defined? It seems the transition from day to night and night to day period is the most tricky one (and does this affect the training of the ML and in the end the performance of the model?). Also there seems to be an overlap in the night time – started at 4PM and end 8AM (perhaps related to winter? but this is not visible in the daytime model, which should have then have included the longer evenings?). The unit is in half hour, but the graph only presents hourly data points, I would suggest to make this consistent. Caption: what does “local hour” mean?
  Response: Thank you for this thoughtful comment. Daytime and nighttime periods are defined using the solar zenith angle (SZA) rather than fixed clock hours, which allows for a physically consistent separation of day and night across seasons and latitudes. As a result, due to seasonal variations in solar geometry, some overlap in local clock hours (e.g., earlier night onset in winter and later night termination) is expected when data are aggregated by hour. This explains the apparent overlap in nighttime hours (e.g., 4 PM to 8 AM). The transition periods between day and night are indeed challenging; however, using SZA-based classification ensures that each data point is consistently labeled based on solar illumination conditions. This approach was applied uniformly during model training and evaluation, thereby minimizing any adverse impact on model performance. We have clarified this point in the manuscript.
  Regarding temporal resolution, the underlying data (EC towers) are at 30-minute resolution, while Figure 3 presents hourly aggregated values for clarity of visualization. We have revised the figure and caption to clearly state this to avoid confusion.
  
  Comment: Line 301, do you mean the one year time series is an average across all sites, or one example year for one stations or??
  Response: Thank you for this comment. The one-year time series shown corresponds to year 2023 and represents values averaged across all sites, rather than a single station example. We have clarified this in the manuscript to avoid ambiguity.
  
  Comment: Figure 3&4, how did you calculate the daily ET, did you use the two different ‘best’ models? How did you deal with the transition hours (between daylight and night time)?
  Response: Good question. Yes, we used two separate “best” models for daytime and nighttime ET. For Figure 3, we aggregated the half-hourly model outputs to hourly values and then averaged across all sites for each hour. For Figure 4, we combined outputs from the two models and averaged them to compute daily ET. In both cases, the underlying units remain mm hh-¹.
  
  Comment: Figure 6, why did you combine certain climate classes and not include the BSh climate class? How many stations are included in each climate class?
  Response: Thank you for this helpful comment. In the initial version, we focused on a subset of climate classes that were most representative based on spatial coverage in order to maintain figure clarity. Following this comment, we reconsidered this choice and expanded the analysis to include all Köppen climate classes, including BSh, to provide a more comprehensive evaluation. Figure 6 has been revised accordingly, and Figure A1 now presents the complete analysis, as including all classes in a single figure would have reduced readability and image quality. The associated text has also been updated to reflect the expanded set of climate categories. The number of stations contributing to each climate class is now reported in the revised tables and described in the text.
  
  Comment: Figure 7, why were the other land cover classes not included?
  Response: Thank you for this helpful comment. In the initial version, we focused on a subset of land cover classes that were most representative based on EC tower coverage in order to maintain figure clarity. Following this comment, we reconsidered this choice and expanded the analysis to include all available IGBP land cover classes to provide a more comprehensive evaluation. Figure 7 has been revised accordingly, and Figure A2 now presents the complete analysis, as including all classes in a single figure would have reduced readability and image quality. The associated text has also been updated to reflect the expanded set of land cover categories.
  
  Comment: Line 455: the bias comment is not really substantiated. The values during the night are generally much lower than at day time and the difference may look larger in figure 3, but this is not quantified. I would suggest to include bias as a performance indicator and add it to table 3 to support this comment?
  Response: Thank you for this helpful comment. We agree that the statement regarding bias needed stronger quantitative support. In practice, the bias values in our results are generally close to zero, which is expected given the nature of the machine learning modeling. For this reason, bias did not provide strong additional discriminatory power across conditions. Following your suggestion in another comment, we instead adopted the use of nRMSE normalized by the median value, which substantially improved the presentation and interpretability of the results. This metric more appropriately accounts for the lower nighttime values and quantitatively reflects the relative errors under those conditions. We have revised the analysis and tables accordingly and clarified this point in the manuscript.
  
  Comment: Line 515 since you are explaining the results per climate, is the comment related to the vegetation cover really relevant here?
  Response: Thank you for this comment. While this section primarily discusses model performance across climate classifications, we believe that referencing vegetation cover is relevant because climate and vegetation are inherently coupled drivers of evapotranspiration. Differences in vegetation density, phenology, and sub-grid heterogeneity systematically co-vary with climate regimes and directly influence ET partitioning between transpiration and soil evaporation.
  
  Comment: Line 517-520, but would you expect that the model would perform better at these locations if it is specifically trained for those conditions?
  Response: Thank you for this insightful comment. We agree that training the model specifically on sites within a given climate or environmental condition could potentially improve performance locally. However, our goal was to develop a generalized model applicable across diverse climates and land cover types. The current strong performance in Mediterranean and humid subtropical climates suggests that even without climate-specific training, the model effectively captures ET dynamics where relationships among radiation, temperature, and vegetation indices are stable.
  
  Comment: Line 571: which result show that ALIVEet underestimates ET for ‘high vegetation density and complex moisture dynamics’?
  Response: Thank you for this comment. The underestimation of ALIVE ET in regions with high vegetation density and complex moisture dynamics is illustrated in Figure 8, where we compare ALIVE ET with ALEXI ET. The difference map shows that ALIVE ET values are generally lower along the East Coast, and the red colors in the figure indicate areas where ALIVE ET is underestimated relative to ALEXI ET. We have clarified this point in the manuscript.
  
  Comment: Line 573: which result show that ALIVEet ‘struggles to capture ET dynamics in the peak growing season’?
  Response: The observation that ALIVE ET struggles to capture ET dynamics during the peak growing season is supported by the day-of-year (DOY) 149 and 179 comparisons shown in Figure 8, where underestimation is most pronounced. In these comparisons, ALIVE ET shows lower values than ALEXI ET in regions with high vegetation density, indicating that the model may not fully capture the elevated ET rates typical of the peak growing season. Additionally, the scatterplots in Figures 6 and 7 show that very high ET values deviate slightly from the 1:1 line, further indicating that ALIVE ET may not fully capture extreme ET rates.
  
  Edits: Comment: Line 144, remove brackets from the reference.
  Response: Thank you for this comment. We have removed the brackets from the reference and revised it accordingly throughout the manuscript, ensuring that the formatting of all references now aligns with the journal style.
  
  Comment: Line 181, reduce number of names from the reference (check referencing style of journal)
  Response: We have reduced the number of authors displayed in the reference to conform with the journal’s referencing style and revised it consistently throughout the manuscript.
  
  Comment: Line 194 remove initials from reference (check referencing style of journal)
  Response: We have removed the initials from the reference as per the journal’s style guidelines and updated the formatting throughout the manuscript.
  
  Comment: References: Tran, B. N., van der Kwast, J., Seyoum, S., Uijlenhoet, R., Jewitt, G., and Mul, M.: Uncertainty assessment of satellite remote-sensing-based evapotranspiration estimates: a systematic review of methods and gaps, Hydrol. Earth Syst. Sci., 27, 4505–4528, https://doi.org/10.5194/hess-27-4505-2023, 2023.
  Response: Thank you for suggesting this reference. We found it highly relevant and have now cited it in the manuscript to acknowledge the systematic review of uncertainty assessment in satellite-based evapotranspiration estimates.
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2025-4400-AC1

Sadegh Ranjbar, Danielle Losos, Sophie Hoffman, Yafang Zhong, Jason A. Otkin, Ankur Rashmikant Desai, Martha Anderson, Christopher R. Hain, and Paul Christopher Stoy

Viewed

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 122 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
121	0	1	122	0	0

HTML: 121
PDF: 0
XML: 1
Total: 122
BibTeX: 0
EndNote: 0

Views and downloads (calculated since 01 Dec 2025)

Month	HTML	PDF	XML	Total
Dec 2025	112	1	113
Jan 2026	9	0	9

Cumulative views and downloads (calculated since 01 Dec 2025)

Month	HTML	PDF	XML	Total
Dec 2025	112	1	113
Jan 2026	9	0	9

Viewed (geographical distribution)

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 117 (including HTML, PDF, and XML) Thereof 117 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 13 Jan 2026

Short summary

Water moves from land to air in a process called evapotranspiration, which affects weather, crops, and water supply. Using satellites and AI, we created a system that tracks this water movement every five minutes, day and night, even through clouds. This provides continuous insights that can help manage water, predict weather, and better understand the water cycle.


Total:	0
HTML:	0
PDF:	0
XML:	0