Non-Gaussian data assimilation of satellite-based leaf area index observations with an individual-based dynamic global vegetation model

Arakida, Hazuki; Miyoshi, Takemasa; Ise, Takeshi; Shima, Shin-ichiro; Kotsuki, Shunji

doi:https://doi.org/10.5194/npg-24-553-2017

Articles | Volume 24, issue 3

https://doi.org/10.5194/npg-24-553-2017

© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 3.0 License.

https://doi.org/10.5194/npg-24-553-2017

© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 3.0 License.

Articles | Volume 24, issue 3

Research article

|

06 Sep 2017

Research article |

| 06 Sep 2017

Non-Gaussian data assimilation of satellite-based leaf area index observations with an individual-based dynamic global vegetation model

Hazuki Arakida, Takemasa Miyoshi, Takeshi Ise, Shin-ichiro Shima, and Shunji Kotsuki

Download

Final revised paper (published on 06 Sep 2017)
Preprint (discussion started on 25 May 2016)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of Non-Gaussian data assimilation of satellite-based Leaf Area Index observations with an individual-based dynamic global vegetation model', Matthias Morzfeld, 01 Jul 2016
- AC1: 'Reply to referee 1', Hazuki Arakida, 30 Aug 2016
RC2: 'Referee's comments', Malaquias Pena, 05 Jul 2016
- AC2: 'Reply to referee 2', Hazuki Arakida, 30 Aug 2016

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Anna Wenzel on behalf of the Authors (28 Oct 2016) Author's response

ED: Referee Nomination & Report Request started (01 Nov 2016) by Amit Apte

RR by Matthias Morzfeld (16 Nov 2016)

Suggestions for revision or reasons for rejection

I thank the authors for addressing my previous concern about the manuscript. The additional numerical experiments are indeed very useful in accessing the validity of the approach and of the conclusions of this paper.

The authors state that they use an "efficient particle filter approach", sequential importance resampling (SIR). However, it is widely known and generally accepted that particle filters are not efficient in the sense that an (usually) excessive amount of particles is needed to solve a given problem, especially if the dimension of the system is large. If the number of particles is large, the computations required to perform particle filtering also become large, as each particle requires (at least) one simulation. This failure of particle filters can be seen in this application, even though relatively large numbers of particles (500-16,000) are used. The failure becomes apparent when looking at figures 4, 6, 8, where the mean and "error bars" are shown for the estimated parameters. It can be seen that the data assimilation by particle filters changes the mean and decreases the variance, however the variance decrease is largely dominated by the amount by which resampled particles are perturbed. These perturbations are required to avoid "particle filter collapse", as is shown by the numerical experiments with smaller (and larger) perturbations. However, the collapse is only avoided by allowing the particles to "spread out" to within the bounds dictated by the perturbations. These bounds are somewhat arbitrary, and, for that reason, the results and conclusions are also somewhat arbitrary. Specifically, I expect that figures 4, 6, 8 look very different when different perturbations are used during resampling. In short, the numerical experiments suggest that the particle filter picks "the best particle", i.e., the one that is closest to the observation, and then defines uncertainty estimates that are largely dominated by the perturbations of the resampled ensemble.

The authors state in their conclusions that uncertainties of the state variables and model parameters are greatly reduced, and that, overall, uncertainty is significantly reduced. This is true, however the reduction is dominated by the perturbations applied to the resampled ensemble and the paper makes no suggestions as to how to chose these perturbations. It merely tests a few settings (small, moderate and large), and from these tests concludes that "moderate" is a good choice. In this way, the method yields results that are largely influenced by certain choices of parameters in the data assimilation method, rather than that the data assimilation method produces useful results without excessive tuning.

On the other hand, many data assimilation in daily use, e.g., the ensemble Kalman filter are known to "work" only after extensive tuning. However, the tuning is motivated by the physics of the problem or corroborated by theory and extensive numerical experiments. I recommend that the authors be more cautious about the tuning choices they make, to explain better why these choices are necessary and how guidelines for this tuning process can be established, at least for this or closely related problems (such as the one that perform this data assimilation on a global scale).

Hide

RR by Malaquias Pena (05 Dec 2016)

ED: Reconsider after major revisions (further review by Editor and Referees) (16 Feb 2017) by Amit Apte

AR by Hazuki Arakida on behalf of the Authors (20 Apr 2017) Author's response Manuscript

ED: Publish subject to minor revisions (further review by Editor) (13 May 2017) by Amit Apte

ED: Referee Nomination & Report Request started (16 May 2017) by Amit Apte

RR by Matthias Morzfeld (10 Jun 2017)

ED: Publish subject to technical corrections (24 Jul 2017) by Amit Apte

AR by Hazuki Arakida on behalf of the Authors (28 Jul 2017) Author's response Manuscript

Short summary

This is the first study assimilating the satellite-based leaf area index observations every 4 days into a numerical model simulating the growth and death of individual plants. The newly developed data assimilation system successfully reduced the uncertainties of the model parameters related to phenology and carbon dynamics. It also provides better estimates of the present vegetation structure which can be used as the initial states for the simulation of the future vegetation change.