A new bias-correction method for precipitation over complex terrain suitable for different climate states: a case study  using WRF (version 3.8.1)

Velasquez, Patricio; Messmer, Martina; Raible, Christoph C.

doi:https://doi.org/10.5194/gmd-13-5007-2020

Articles | Volume 13, issue 10

https://doi.org/10.5194/gmd-13-5007-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-13-5007-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 13, issue 10

Methods for assessment of models

|

26 Oct 2020

Methods for assessment of models |

| 26 Oct 2020

A new bias-correction method for precipitation over complex terrain suitable for different climate states: a case study using WRF (version 3.8.1)

Patricio Velasquez, Martina Messmer, and Christoph C. Raible

Download

Final revised paper (published on 26 Oct 2020)
Preprint (discussion started on 01 Jul 2019)

Interactive discussion

Status: closed

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

- Printer-friendly version

- Supplement

RC1: 'An interesting study subject to several shortcomings though', Anonymous Referee #1, 24 Jul 2019
- AC1: 'Response to Referee #1', Patricio Velasquez, 06 Oct 2019
SC1: 'Executive Editor Comment on gmd-2019-131', Astrid Kerkweg, 26 Jul 2019
- AC2: 'Response to Executive editor', Patricio Velasquez, 06 Oct 2019
RC2: 'review Velasquez et al', Anonymous Referee #2, 09 Sep 2019
- AC3: 'Response to Referee #2', Patricio Velasquez, 06 Oct 2019

Peer-review completion

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

AR by Patricio Velasquez on behalf of the Authors (18 Nov 2019) Author's response Manuscript

ED: Referee Nomination & Report Request started (24 Nov 2019) by Fabien Maussion

RR by Anonymous Referee #1 (22 Jan 2020)

Suggestions for revision or reasons for rejection

The authors put considerable efforts into improving their manuscript. The revisions clearly improved the quality of the work. However, I believe that the main concerns regarding the method applied and its general motivation have not been addressed in a satisfying manner. The authors' responses to the original comments are only partly appropriate, and the paper still suffers from major shortcomings in my opinion. I'm listing the most important points below (this is partly a repetition of my original review). I'm not listing further minor issues this time. Given these remaining concerns my suggestion cannot be anything else than to return the paper for major revisions and to reconsider a publication afterwards. Involving an additional reviewer and thereby a further view on the quality of the manuscript might be helpful. I'm sorry for this rather harsh recommendation, but I truly believe the paper needs some more work to be carried out on it in order to be a valuable contribution to the community. I hope my comments will be helpful for this exercise.

With kind regards.

Motivation of the chosen approach: In my opinion the authors still fail to convincingly show that the inclusion of surface elevation or further topography-related co-variates is helpful for bias-correcting the model output. Again, what is surely needed is an analysis of bias ranges in the different topography-related classes in order to answer the question how informative orography is for the model bias. Other than stated in the replies, this is not shown by Figure 3 (this figure does not show the ranges of biases in the two classes, but the ranges of mean precipitation intensity). Also Figure 4 does not show this analysis explicitly. It rather points to the fact that biases can differ a lot even within the same elevation class (strong winter overestimation in the mountains along the northern rim, strong underestimation in the mountains of Ticino, and further examples). Also figure R1 in the replies only shows the mean biases (which are for many classes actually very close to each other). Furthermore, the general motivation is frequently mentioned in the paper: The transferability of the bias correction to a different climate state with a potentially different surface topography. I still believe this applicability is not apparent at all, at least one further experiment is required for this (see my first review). Based on Figure 4, topography seems to be a weak co-variate even in the present day climate, there is no indication that application to even a different climate state would be safer compared to other approaches. Even in the present day climate and in the training period, model performance is partly worse after bias correction (compare Figs. 4d and 4e, or see the Taylor diagrams in Figure 7). The replies of the authors concerning the rainfall/snowfall relation in a different (e.g. glacial) climate are not satisfying either. I believe it is obvious that running the model in a much colder climate setting might introduce modification of the bias structure due to handling of rainfall/snowfall in the model's microphysics.

Standard EQM: As also pointed out by Reviewer #2, the standard application of EQM in a bias-correction setting is really the grid-point wise training of a transfer function (not one transfer function for the entire domain). This standard application should be the reference for comparing the new version against.

Second bias correction step (wet day frequency adjustment): The motivation and the details for the second step of the bias correction are still very unclear to me. Standard EQM (i.e. the third step) would account for this already. The motivation via the different thresholds used is very unclear to me, sorry. Furthermore, the authors explain that they are using the LOCI method (Schmidli et al. 2006) for this step. This is not true, as the method involves both the threshold adjustment and the subsequent intensity scaling. The present work only used the first LOCI step.

Validation and cross validation: First of all, the cross validation setup (and, in general, the entire validation setup) should be explained in the bias correction chapter (3), not in the result chapter. Then, most of the validation carried out yields results that are to be obtained by definition (no separate calibration period, no spatial transfer). They therefore only show that the method works as it should, but do not show an independent validation. The authors need to make this point much clearer. Only the spatial cross validation contains an independent element. The performance shown, however, is not too convincing. One reason might really be the shift of spatial scale (training on 5 km resolution and applying on 2 km resolution). Precipitation intensities will systematically differ between both resolutions, therefore some EQM trained at 5 km resolution is not to be applied at 2 km in a straightforward manner. In the revised version the authors mention an additional validation that does not mix spatial scales, but really need to show this explicitly in my opinion in the paper (Figure R3). Furthermore, for assessing the new method in the spatial cross validation a benchmark is required to show how meaningful the orography-based method is compared to simpler approaches. Such a benchmark could be a simple correction of the mean bias or, as used in the standard validation, one EQM transfer function for the entire region.

Sliced simulation: I do understand the reason behind slicing the 30-year long simulation into 3-year parts, but the justification in the replies and in the revised version is insufficient, if not wrong. Other than mentioned by the authors on page 5 lines 1-3 (version including tracked changes), WRF is not atmosphere-only but is running with its own land surface scheme that can definitely have longer equilibrium times than 2 months, even when initialised with spinned-up GCM fields.

Orographic classes: Some points of the definition of orographic classes are unclear (second paragraph of Chapter 3). How are these classes derived for the observational grid, for instance (a different grid than WRF). Is some spatial interpolation involved?

Hide

RR by Anonymous Referee #2 (31 Jan 2020)

ED: Reconsider after major revisions (01 Feb 2020) by Fabien Maussion

AR by Patricio Velasquez on behalf of the Authors (17 Apr 2020) Author's response Manuscript

ED: Reconsider after major revisions (14 May 2020) by Fabien Maussion

Dear Authors,

thank you for your revised manuscript. It took me longer than usual to assess your manuscript because, as you have seen, a previous reviewer expressed serious concerns about the method you propose. I, for myself, was worried about the main conclusion of the study being that the method is "suitable for different climate states" without showing a single experiment demonstrating that it is the case. You now submitted a revised study with an even stronger focus on this potential application, without adding substantial new content to the paper.

In your reply, I believe that you addressed some of the important methodological concerns raised by reviewer 1 (the lack of proper independent validation). However, you dismissed the reviewer's main comment 2 ("transferability of the bias correction to a different climate state with a potentially different surface topography") by providing elements of discussion in the response letter, but you chose not to add these LGM analyses to the paper.

Let me repeat here the last sentences of your abstract:
"A rigorous temporal and spatial cross-validation with independent data exhibits robust results. Thus, the new bias-correction offers a good alternative to the standard quantile mapping method during times where orography strongly deviates from the present-day state." -> the logical step ("Thus...") is still unclear to me based on the paper alone (remember that the response to reviewers is not part of the paper).

You then write: "The new bias-correction method certainly leaves some drawbacks under present-day conditions. Nevertheless, the assumption of stationarity that is inherent in all bias-correction methods is handled more flexible; and hence, it is predestined for studies focussing on colder climate states." -> This last sentence leaves me worried that the method will be applied without scrutinity later since "it has been shown in GMD that it works".

If your method is not very useful when orography doesn't change (your own statement) but becomes useful when orography does change, than why not adding a section and a few plots showing that the LGM case is more realistically handled (based on Figs R2 and R3), and make your study much more valuable?

Don't get me wrong: I appreciate the difficulty of validating bias-correction methods for LGM conditions, and I wish your project to succeed. I would like to give you the opportunity to reassess your decision of not showing any "changed topography" simulation results in your manuscript. You can choose to dismiss my suggestion (in which case your answer to this review can be brief), and it will be the tasks of the reviewer(s) to decide.

Best regards,

Fabien Maussion

Hide

AR by Patricio Velasquez on behalf of the Authors (18 Jun 2020) Author's response Manuscript

ED: Referee Nomination & Report Request started (22 Jun 2020) by Fabien Maussion

RR by Anonymous Referee #3 (03 Jul 2020)

ED: Reconsider after major revisions (03 Jul 2020) by Fabien Maussion

AR by Patricio Velasquez on behalf of the Authors (12 Aug 2020) Author's response Manuscript

ED: Referee Nomination & Report Request started (14 Aug 2020) by Fabien Maussion

RR by Anonymous Referee #3 (23 Aug 2020)

ED: Publish subject to technical corrections (24 Aug 2020) by Fabien Maussion

AR by Patricio Velasquez on behalf of the Authors (01 Sep 2020) Author's response Manuscript

Short summary

This work presents a new bias-correction method for precipitation that considers orographic characteristics, which can be used in studies where the latter strongly changes. The three-step correction method consists of a separation into orographic features, correction of low-intensity precipitation, and application of empirical quantile mapping. Seasonal bias induced by the global climate model is fully corrected. Rigorous cross-validations illustrate the method's applicability and robustness.