Projections of changes in extreme storm surges for European coasts using statistical downscaling

Apecechea, Maialen Irazoqui; Melet, Angélique; Menendez, Melisa; Lobeto, Hector; Valle-Rodriguez, Jonathan B.

doi:10.5194/nhess-26-2269-2026

Articles | Volume 26, issue 5

https://doi.org/10.5194/nhess-26-2269-2026

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

https://doi.org/10.5194/nhess-26-2269-2026

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

Articles | Volume 26, issue 5

Research article

|

15 May 2026

Research article |

| 15 May 2026

Projections of changes in extreme storm surges for European coasts using statistical downscaling

Maialen Irazoqui Apecechea, Angélique Melet, Melisa Menendez, Hector Lobeto, and Jonathan B. Valle-Rodriguez

Download

Final revised paper (published on 15 May 2026)
Supplement to the final revised paper
Preprint (discussion started on 10 Oct 2025)
Supplement to the preprint

Interactive discussion

Status: closed

CC1:
'Comment on egusphere-2025-3558', Tim Hermans, 17 Oct 2025
The authors present a very interesting analysis on statistical modeling of dynamic changes in storm surges. I have two specific questions that may be helpful while revising the paper:
The authors mention the performance of their statistical model in terms of the projected changes in the 1-in-10 year return height is 'satisfactory'. My question here would be: could you define more quantitatively what you consider to be satisfactory? If I look at Figure 6, I do see agreement, but I also see rather large differences. For instance, between (d), (e) and (f) around the UK coastline, and between (d), (h) and (i) at the west coast of France. It would be helpful to summarize these differences quantitatively and critically describe what such errors mean for the resulting projections.

The authors report that the performance of their MLR model falls off at predicting the highest extremes, which is consistent with previous findings. I therefore wonder whether it would be useful to test how weighting their multi-linear regression would affect the performance at predicting extremes? This could be an interesting test that would hopefully not be very difficult to implement and would nicely complement our recent study on weighting the loss function for training neural networks. Thanks!
Citation: https://doi.org/10.5194/egusphere-2025-3558-CC1
- AC3: 'Reply on CC1', Maialen Irazoqui Apecechea, 03 Mar 2026
  
  We sincerely thank you for your comments.
  Regarding the weighted regression proposed, we agree that this would be a very interesting model to test, as our regression currently optimizes for the average conditions and not extremes. This is highlighted in the revised manuscript (to be uploaded) in the Discussion as a possible way to improve the statistical model. In our study, the objective was not to develop a new methodology, but adopt a method previously shown to be successfull for the reconstruction of storm-surges, tailor it to Europe, and evaluate its capability to project extreme storm surge changes, which is enabled through benchmarking against dynamical simulations. We have now better highlighted this objective in the revised manuscript, and we have better highlighted the limitations of the model throughout the paper. In the revision, we have also focused statistical projections on the 10-year event, given larger biases now shown for the 100 year event.
  Finally, regarding the comment on the validation of statistical projections of ESS changes against the numerical model projections, we agree that results deserved a more detailed description and a quantitative assessment. The reviewers also highlight that these results and associated conclusions needed to be better supported. Therefore, we have now added quantitative skill metrics to the comparison between the statistical vs. dynamically downscaled projections of changes in the 10-year storm surge level (pattern correlation coefficient,mean absolute bias, % of points where the sign of projected changes in reproduced) and we have considerably expanded this section to describe these metrics and properly justify the different claims in performance.
  
  Citation: https://doi.org/10.5194/egusphere-2025-3558-AC3
RC1:
'Comment on egusphere-2025-3558', Marta Marcos, 23 Oct 2025

Please, see the attached file with the revision

Citation: https://doi.org/10.5194/egusphere-2025-3558-RC1
- AC1: 'Reply on RC1', Maialen Irazoqui Apecechea, 03 Mar 2026
  
  We thank the reviewer for their thorough and constructive review. We provide a detailed reply in the attached document.
  
  Citation: https://doi.org/10.5194/egusphere-2025-3558-AC1
RC2:
'Comment on egusphere-2025-3558', Anonymous Referee #2, 30 Jan 2026

I congratulate the authors for the substantial effort invested in this study. The manuscript addresses an important and timely topic and provides a valuable contribution by exploring future projections of daily maximum storm surges along European coasts using statistical downscaling applied to a large CMIP6 ensemble.
The numerical and statistical workload behind this study is impressive, and the authors make a huge effort to assess the validity of their statistical downscaling approach under climate change conditions. The pan-European scope and ensemble-based perspective are clear strengths.
That said, I believe the manuscript would benefit significantly from improvements in structure, clarity, and methodological rigor, particularly regarding hypothesis formulation, inference, and the separation between methods and results. In its current form, methods and results are often interwoven, making the paper difficult to follow. In addition, while many validation steps are presented, the confidence in the final projections remains limited, partly due to the lack of formal hypothesis testing and inference in several key analyses.
My comments attached are intended to be constructive. Some are necessarily subjective or based on my interpretation (comprehensive summary attached in the same document); please feel free to disregard them where they are not useful or where I may have misunderstood aspects of the work.

Citation: https://doi.org/10.5194/egusphere-2025-3558-RC2
- AC2: 'Reply on RC2', Maialen Irazoqui Apecechea, 03 Mar 2026
  
  We thank the reviewer for their thorough and constructive review. We provide a detailed reply in the attached document.
  
  Citation: https://doi.org/10.5194/egusphere-2025-3558-AC2

Peer review completion

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

ED: Reconsider after major revisions (further review by editor and referees) (04 Mar 2026) by Philip Ward

AR by Maialen Irazoqui Apecechea on behalf of the Authors (10 Apr 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (13 Apr 2026) by Philip Ward

RR by Anonymous Referee #2 (13 Apr 2026)

Suggestions for revision or reasons for rejection

I would like to congratulate authors on the thorough and thoughtful response to the review comments. The level of clarity, completeness, and care with which you have addressed the points raised truly sets a high standard, and I very much appreciate the effort that has gone into improving the manuscript.

I have only one minor comment that I believe could further enhance clarity, although it does not affect my overall positive assessment.

In your response, you clarify that the objective is not to introduce a novel statistical downscaling methodology, but rather to apply and evaluate an established PCA feed multilinear regression based SDM framework across a pan-European domain. This is well justified. However, in the manuscript you note that coefficient level significance filtering is not applied, and robustness is instead assessed through testing a large set of configurations.

In this context, I was wondering whether an additional step, i.e. testing the significance of the PCs included in each model configuration (T1–T7), could help reduce unnecessary complexity. My expectation is that some of the more complex models (in terms of atmospheric predictors) might achieve similar skill with fewer statistically significant covariates. This would also be consistent with the multiple linear regression approach adopted in Tadese et al. (2020).
I believe this step is particularly relevant given the choice to apply PCA individually to each variable. While PCA ensures orthogonality among EOF–PC pairs within each variable, this orthogonality does not extend across variables. As a result, some degree of redundancy or interdependence between predictors may remain. An alternative could be to apply PCA jointly across all fields to enforce orthogonality between components; however, I understand that this may introduce practical challenges, as the resulting covariance matrix could become close to singular due to dependencies among variables.

Related to this, Table 2 could benefit from indicating how many principal components are required to explain 99% of the variance for each predictor variable. If I understand correctly, this effectively corresponds to the number of covariates entering each model (with PCA applied separately per variable), and including this information would improve transparency on model complexity.

Additionally, a few very minor points:
In Table 2 (T6), I believe “Dmax” is intended for the squared wind components.
In Figure 2, one of the summations starts at i = 0 and the other at i = 1.

The figure also highlights that PCA is applied per individual component. While this is perfectly valid, it might be helpful to clarify that index j refers to the EOF (I guess ordered by explained variance) of each variable (indexed by i).

Overall, I find the manuscript to be significantly improved and very clear in its scope and contribution.

Hide

RR by Marta Marcos (16 Apr 2026)

ED: Publish subject to technical corrections (17 Apr 2026) by Philip Ward

AR by Maialen Irazoqui Apecechea on behalf of the Authors (28 Apr 2026) Author's response Manuscript

Download

Article (7305 KB)
Full-text XML

Short summary

We applied a fast statistical model to estimate future extreme storm surges in Europe using data from 17 climate models – about twice as many as in past studies. Results show robust regional patterns – decreases in the Mediterranean and Moroccan coast, increases in the Irish Sea and Gulf of Finland – with high uncertainty in other areas. Out results increase our knowledge on future storm surge uncertainties, needed for informed coastal planning.