EGO: a global 0.05° hourly GPP dataset for monitoring diurnal photosynthesis dynamics

Liu, Xi; Li, Xing; Hao, Dalei; Xiao, Jingfeng; Zhou, Yanan; Zhao, Cenliang; Diao, Zikang; Qu, Fuqiang; Lin, Shangrong; Liu, Xiangzhuo; Zhang, Zhaoying; Liu, Xinjie; Zhang, Helin

doi:10.5194/essd-18-6613-2026

Articles | Volume 18, issue 9

https://doi.org/10.5194/essd-18-6613-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/essd-18-6613-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 18, issue 9

Data description article

|

09 Sep 2026

Data description article |

| 09 Sep 2026

EGO: a global 0.05° hourly GPP dataset for monitoring diurnal photosynthesis dynamics

Xi Liu, Xing Li, Dalei Hao, Jingfeng Xiao, Yanan Zhou, Cenliang Zhao, Zikang Diao, Fuqiang Qu, Shangrong Lin, Xiangzhuo Liu, Zhaoying Zhang, Xinjie Liu, and Helin Zhang

Download

Final revised paper (published on 09 Sep 2026)
Supplement to the final revised paper
Preprint (discussion started on 15 Apr 2026)
Supplement to the preprint

Interactive discussion

Status: closed

RC1:
'Comment on essd-2026-40', Anonymous Referee #1, 23 May 2026

Liu et al. develop a new global GPP model called EGO using a CKML-XGBoost approach and compare it against FLUXCOM. The new model is interesting, but generally it’s hard to make too much of a case for studying diurnal patterns of GPP given the massive uncertainties in partitioning GPP (e.g. the nighttime approach will give different patterns and new studies like those of Keenan et al. emphasize the importance of under appreciated processes like the Kok effect). But regardless it’s an interesting challenge for models. The major shortcoming in my opinion is that the authors did not benchmark improvements of the CKML-XGBoost model against traditional approaches (e.g. LSTM) or standard XGBoost models. How are we to know if it’s an improvement, and by how much? I don’t doubt that it makes an improvement, but without explaining what is really new (and how much the novelty makes a difference), it’s hard to know if CKML is really a future direction of an approach that might work in principle but may (or may not) be meaningfully better. Such an addition needn’t extend to the FLUXCOM comparisons but it is important from a model development perspective. Minor comments follow.

39: GPP is just canopy photosynthesis, the way it’s being described is like you’re trying to talk around something

46: these early studies like Wofsy et al. and Grace et al. and more should be cited

In general, fundamental (or at least more fundamental) references are missing throughout the manuscript. For example, both Ruehr et al. references cited beforehand are great to note, for example, but these are not the first times that people realized that GPP is the largest flux in the global C cycle or that vegetation responds to environmental variability at multiple (including short) time scales.

The PCMCI is interesting but how does this compare to KGML?

106: this paragraph makes a great point, but what metric will be introduced instead? I’ve always wondered why people don’t just use Nash-Sutcliffe modeling efficiency, and there would be a whole world of time and frequency (e.g. wavelet)-based approaches for choosing a superior metric.

111: Note Khan et al. (2022) here. The introduction is generally well reasoned but would benefit from a broader suite of references from the multiple groups who are working on these challenges, as well as the foundational papers only very briefly alluded to above.

140: this pixel/footprint matching requires more detail, especially given all the work that’s gone into this topic by Chu et al. (2021, Ag. For. Met.) and others.

Figure 1 is nice, would benefit from larger text in some of the subplots.

159-160: just say 2 m. Probably too technical a point but using non-breaking spaces on 164, 165 & elsewhere between mathematical characters and values would be an improvement, also a subscript on the 2 in CO2 around here. Throughout, there are a number of minor usage issues that just require a careful read to check.

3.2.1: I agree in principle with this approach but would be curious to know if a simple LSTM would fare much more poorly if the challenge is to incorporate causal inference, or, alternately, if XGBoost alone gives a reasonable depiction of instantaneous flux as it often does. I don’t question the CKML approach, rather I wonder if it’s the only thing that might beat XGBoost here because a number of products including FLUXCOM-X, CASS, ALIVE, and many studies (that didn’t give their model an acronym, e.g. https://doi.org/10.3390/land14010124) have arrived at XGBoost. But without benchmarking against XGBoost alone, the CKML-XGboost model’s improvements can’t be quantified.

3.2.2: were any sites fully held out for the training testing split (commonly 70:30 or 80:20 or similar) or was the fold approach adopted alone? I’m wondering if this isn’t a full fair comparison against FLUXCOM; how does the model perform against data from sites that it has never seen before?

3.2.3 (really before): how was flux data quality considered?

272: the relationship between GPP and radiation isn’t linear, it saturates.

320 and beyond: I’d be curious about the time dependencies of these variables as soil moisture should increase in importance at longer time scales and at annual timescales temperature will likely be more important following the findings of Jung et al. This is perhaps a different topic though.

Fig. 6 and related comparisons: as noted above can we be sure that the FLUXCOM products are treated to a fair comparison in these analyses?

For Figure 12, what time period is the reference? Also, this is more than the US, also a good chunk of Canada and Mexico

Citation: https://doi.org/10.5194/essd-2026-40-RC1
- AC1: 'Reply on RC1', Xing Li, 06 Jul 2026
  
  Thank you for the comments. We have uploaded a point-by-point response attached as a supplement.
  
  Citation: https://doi.org/10.5194/essd-2026-40-AC1
RC2:
'Comment on essd-2026-40', Anonymous Referee #2, 04 Jun 2026
This study aimed to produce eddy covariance (EC) site-based Global Hourly GPP estimates (named EGO) using a causal inference approach (CKML-GPP) spanning the period from 2000 to 2022. While the model's performance in regions with limited EC flux sites may remain challenging, I believe this attempt to improve hourly GPP estimates by combining high-frequency on-site EC data with remote sensing data carries important insights for global carbon cycle studies. Overall, the manuscript reads well. However, I felt the explanations regarding how the authors compared EGO with earlier products (i.e., FLUXCOM and X-BASE) are currently lacking and require clarification. Below are my specific comments.
L65: Please spell out “OCO” at its first occurrence.

L97: Lin et al. (2021) do not seem to use a causal inference approach in their analysis. Citing this paper here for the PCMCI method is somewhat confusing.

L165-166: Please subscript the 2 in “CO2”.

L207: Please add a space in “TableS2”.

Fig 1 (c): The legends are a bit small. Please increase their font size.

L249: The authors use the phrase “following previous studies,” but only a single study is cited. Please revise or add additional citations.

L250-252: Until reading this sentence, I expected the EGO outputs to be continuous hourly GPP estimates, similar to X-Base. It would be better to clarify this earlier in the manuscript, such as in the Abstract or the final paragraph of the Introduction.

L304-305: What does “all products were compared under identical conditions” actually mean? Specifically: (1) Did you rerun the other models to apply the same cross-validation approach used for EGO?, (2) Were the comparisons conducted over the exact same time periods (both across years and hours of the day), and were they all evaluated using monthly-averaged hourly GPP outputs?, (3) How were the observed GPP data sampled? Were only the EC tower sites included in each respective model plotted?

Adding a dedicated paragraph in the Methods section to clarify this comparison is necessary. Currently, the comparison feels like it abruptly begins in Results Section 4.1.

L320: Please spell out the “SHAP” method at its first mention.

Figs 6–8: “Local solar time” should include the unit (“Hour”) in the axis titles.

Fig 6: Please spell out the six vegetation types for each panel in the figure caption.

Fig 8: The subpanel title (a) overlaps with the y-axis values. Also, the statistics table in panel (a) overlaps the Case 3 line. Please relocate them for better visualizations.

Fig 9: The subplots span from 8:00 to 16:00, but the figure caption states they run from 8:00 to 18:00.

Fig 11: Please add parentheses for panels (c) and (d) in the figure caption.

Figs 11 & 14: Please add the appropriate units for each variable in the color scale bars.

5.1: The authors suggest that EGO performs better than past models (especially X-Base, given its use of hourly data inputs) mainly due to the inclusion of the aridity index and the implementation of PCMCI. While the authors have compared the model with and without AI in Fig. S6, proving that the inclusion of AI improves performance, how does the comparison look between PCMCI + XGBoost vs XGBoost alone?
Citation: https://doi.org/10.5194/essd-2026-40-RC2
- AC2: 'Reply on RC2', Xing Li, 06 Jul 2026
  
  Thank you for the comments. We have uploaded a point-by-point response attached as a supplement.
  
  Citation: https://doi.org/10.5194/essd-2026-40-AC2

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

AR by Xing Li on behalf of the Authors (06 Jul 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (14 Jul 2026) by Hao Shi

RR by Anonymous Referee #1 (27 Jul 2026)

ED: Publish as is (31 Jul 2026) by Hao Shi

AR by Xing Li on behalf of the Authors (03 Aug 2026)

Download

Article (17629 KB)
Full-text XML

Short summary

Vegetation photosynthesis, quantified as Gross Primary Productivity (GPP), changes rapidly throughout the day. We developed a global hourly GPP dataset called EGO (Eddy covariance site-based Global hOurly) by combining tower observations with advanced artificial intelligence that accounts for causal effects, outperforms existing hourly GPP products and well captures diurnal photosynthesis dynamics. It will provide a reliable foundation for investigating sub-daily ecosystem processes and benchmarking Earth system models.

EGO: a global 0.05° hourly GPP dataset for monitoring diurnal photosynthesis dynamics

Download

Interactive discussion

Peer review completion

Suggestions for revision or reasons for rejection