A machine learning approach targeting parameter estimation for plant functional type coexistence modeling using ELM-FATES (v2.0)

Li, Lingcheng; Fang, Yilin; Zheng, Zhonghua; Shi, Mingjie; Longo, Marcos; Koven, Charles D.; Holm, Jennifer A.; Fisher, Rosie A.; McDowell, Nate G.; Chambers, Jeffrey; Leung, L. Ruby

doi:https://doi.org/10.5194/gmd-16-4017-2023

Articles | Volume 16, issue 14

https://doi.org/10.5194/gmd-16-4017-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-16-4017-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 16, issue 14

Development and technical paper

|

17 Jul 2023

Development and technical paper |

| 17 Jul 2023

A machine learning approach targeting parameter estimation for plant functional type coexistence modeling using ELM-FATES (v2.0)

Lingcheng Li, Yilin Fang, Zhonghua Zheng, Mingjie Shi, Marcos Longo, Charles D. Koven, Jennifer A. Holm, Rosie A. Fisher, Nate G. McDowell, Jeffrey Chambers, and L. Ruby Leung

Download

Final revised paper (published on 17 Jul 2023)
Supplement to the final revised paper
Preprint (discussion started on 04 Jan 2023)
Supplement to the preprint

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2022-1286', Anonymous Referee #1, 01 Feb 2023

In this paper, the authors employ a machine-learning approach to optimize parameters from a vegetation demography model - FATES. Their approach clearly shows the bright application of ML as a tool to improve the next-generation Earth system models. The paper is very well written and the question being addressed is novel. I really enjoyed reading the manuscript and learned a lot from the authors. I would recommend accepting this paper in its current form.

Citation: https://doi.org/10.5194/egusphere-2022-1286-RC1
- AC1: 'Reply on RC1', Lingcheng Li, 31 Mar 2023
  
  RC: In this paper, the authors employ a machine-learning approach to optimize parameters from a vegetation demography model - FATES. Their approach clearly shows the bright application of ML as a tool to improve the next-generation Earth system models. The paper is very well written and the question being addressed is novel. I really enjoyed reading the manuscript and learned a lot from the authors. I would recommend accepting this paper in its current form.
  AC: Thank you for your positive feedback and for suggesting acceptance of our submission for publication.
  
  Citation: https://doi.org/10.5194/egusphere-2022-1286-AC1
RC2:
'Comment on egusphere-2022-1286', Gregory Duveiller, 15 Feb 2023

This is an interesting study making use of machine learning to improve modelling parametrization in a complicated set-up of tropical forests with competing PFTs adopting different strategies. The manuscripts reads well and is clear and exhaustive. The paper is well suited for GMD and should be published. But before, I have the following suggestions to make, hopefully to improve it ...

L242: L_leaf is not defined yet in the text, and neither is WD

L260: perhaps explain what the "ensemble" is before. Help readers understand what makes XGBoost different from a standard Random Forest

The following is a bit subjective, but I find the structure of the Methodology section slightly sub-optimal. I would have prefered to have an overall experimental design at the beginning, e.g. before talking about the SHAP values. This could be just an overview, not necessarily with all the details in section 2.4, but rather a general understanding of what will come up after. Furthermore, section 2.4 is very long compared to the other ones, so splitting it or rebalancing them would be good.

Also, when you present the observational relaitonships (Eqn 1 to 3) it does not seem entirely clear why this is done. While after the difference between Exp 1 and Exp 2 is mentioned, that is clear. I this it would read better if these observatinal relationships are introduced when you explain the experimental design (around line 300).

L313: why is this in bold?

L348: why 15%?

L349: why [0.3, 0.7]?

L374: What does this incapacity of Exp2 (with the observational constraints) to reproduce realistic coexisting ratios tell us about the standard model itself (ELM-FATES without the ML)? Does it reveal some underlying biases/problems in the structure of ELM-FATES itself that makes it less adapted to this tropical context for some reason? It would nice to get some insights on this to guide model developement.

L422: The main problem here is probably the paucity of ELM-FATES simulations that are available to train the XGBoost, rather than the ML algorithm itself. To get highly non-linear behaviour, these require many more training samples. These would typically be trainend with many 1000s of simulations. You could mention that other ML techniques work better for sparser data (even when highly non-linear). I believe Guassian Processes would be one good candidate to explore. (I am not saying this is needed in this paper, but it shoudl be discussed + acknowledged).

L477: Why not place the orange coexistence above the green XGBoost, to increase clarity? or perhaps try some transparnecy?

L525: Fig 9: There is something I am not following anymore... I must be missing something... why are there are both "early" and "late" boxes when the measure/index is precisely a difference between late and early? Not sure what I need to take home as a message from this plot.

L549: Fig 10: Perhaps it would be wise to also add a bundle for the "standard" ELM-FATES without ML (ie. Exp 1? or also 2?) just to illustrate the improvements that the study proposes brings things closer to the observations.

The discussion is quite nice and exhaustive. However, it stays very focused on this modelling experiment and context. This is nice, but it would be (in my opinion) better to also have a more general overview going beyond this specific case of ELM-FATES on tropical forests of Manaus. To increase the breadth of the paper, I think a more general discussion on how to incorporate ML with process-based models would be very welcome. Inspirations could come from the perspective paper of Reichstein et al (2019) [https://doi.org/10.1038/s41586-019-0912-1]. For instance, it would be interesting to know if here the present study could fit in this logic of "hybrid-modelling" . There are also there other strategies to combine ML with ELM-FATES (as discsussed in the perpective paper) that could be evoked to outline further perspectives of this GMD manuscript. Finally, I would also encourage some more words of how much the auhtors beleive their approach is transferable to other contexts beyond tropical forests.

Citation: https://doi.org/10.5194/egusphere-2022-1286-RC2
- AC2: 'Reply on RC2', Lingcheng Li, 31 Mar 2023
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2023/egusphere-2022-1286/egusphere-2022-1286-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2022-1286-AC2

Peer review completion

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

AR by Lingcheng Li on behalf of the Authors (31 Mar 2023) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (12 Apr 2023) by Klaus Klingmüller

RR by Anonymous Referee #2 (25 May 2023)

ED: Publish subject to technical corrections (07 Jun 2023) by Klaus Klingmüller

AR by Lingcheng Li on behalf of the Authors (09 Jun 2023) Author's response Manuscript

Short summary

Accurately modeling plant coexistence in vegetation demographic models like ELM-FATES is challenging. This study proposes a repeatable method that uses machine-learning-based surrogate models to optimize plant trait parameters in ELM-FATES. Our approach significantly improves plant coexistence modeling, thus reducing errors. It has important implications for modeling ecosystem dynamics in response to climate change.