Estimating extreme river discharges in Europe through a Bayesian network

Paprotny, Dominik; Morales-Nápoles, Oswaldo

doi:https://doi.org/10.5194/hess-21-2615-2017

Articles | Volume 21, issue 6

https://doi.org/10.5194/hess-21-2615-2017

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

https://doi.org/10.5194/hess-21-2615-2017

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

Articles | Volume 21, issue 6

Research article

|

02 Jun 2017

Research article |

| 02 Jun 2017

Estimating extreme river discharges in Europe through a Bayesian network

Dominik Paprotny and Oswaldo Morales-Nápoles

Download

Final revised paper (published on 02 Jun 2017)
Supplement to the final revised paper
Preprint (discussion started on 20 Jun 2016)
Supplement to the preprint

Interactive discussion

Status: closed

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

- Printer-friendly version

- Supplement

RC1: 'RC: general comments', Anonymous Referee #1, 26 Jul 2016
- AC1: 'Reply to Anonymous Referee #1', Dominik Paprotny, 25 Aug 2016
RC2: 'comments on "Estimating extreme river discharges in Europe through a Bayesian Network"', Anonymous Referee #2, 09 Aug 2016
- AC2: 'Reply to Anonymous Referee #2', Dominik Paprotny, 25 Aug 2016
EC1: 'Comments on manuscript', Jan Seibert, 13 Sep 2016

Peer-review completion

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

ED: Reconsider after major revisions (14 Sep 2016) by Jan Seibert

Thanks for submitting this interesting manuscript to HESS. This manuscript presents
a new approach to discharge peak estimation, which obviously is a very timely issue.
The reviewers provide valuable comments on the manuscript. In addition I have the
following concerns:

The presentation of the BN performance in figs 5&8 is misleading as the values are
given in m3/s, where results always look much nicer than if specific discharge values
would be plotted (see figs 6&7 in Wrede et al. (2013) as example). Based on my com-
ments &reply before publication in HESS-D, I understand the aim to compare results
with previous papers (which used the m3/s comparison) and the drop in performance
is now mentioned in one sentence. However, it would be much better if the plots would
present the simulations in the fair way (i.e. specific discharge values) and the com-
parison would be given in words that vice versa. After all, we should aim at providing
examples of good practices!

Opposite to the authors’ interpretation of Rojas et al. (2012) different climate models
(GCM/RCM combinations) can provide largely varying results. Therefore, the use of
only one GCM/RCM comination is a clear limitation and largely ignoring the uncertain-
ties caused by the climate models (see conclusions in Rojas et al. 2012; and recent
papers by PhD students in my group (Teutschbein and Seibert, 2012; Addor et al.,
2014). Furthermore, given the often significant biases in climate model simulations,
the decision to not use any bias correction is surprising at least. Both decisions (only
one model, no bias correction) need to be motivated more convincingly. At least I
would recommend the authors to test the effects of using another model and/or bias
correction.

My major concern, however, is figure 7, which illustrates predicted future changes in
the 100-year floods over Europe. Such figures easily get the attention of media and
decision makers and as scientists we, thus, really have to be careful in how we com-
municate such results. Presenting such results without also quantifying uncertainties
is problematic. I, thus, strongly recommend to quantify the uncertainties of the results
shown in fig. 7. There are partly huge differences in the predicted changes in neigh-
boring streams (even in regions where snow is not important) and I am missing an
explanation of these differences. One approach to help understanding the patterns
would be to use a uniform precipitation change, so that the effects of differences in
catchment characteristics versus differences in precip-changes could be disentangled.
(In addition, as far as I can see, the climate change prediction is not mentioned in the
methods part.)

Addor, N., Rössler, O., Köplin, N., Weingartner, R., Seibert, J., 2014. Robust changes
and sources of uncertainty in the projected hydrological regimes of Swiss catchments,
Water Resources Research, 50(10), 7541–7562, doi:10.1002/2014WR015549.

Teutschbein, C., and Seibert, J., 2012.Bias Correction of Regional Climate Model Simulations for Hydrological Climate-Change Impact Studies: Review and Evaluation of Different Methods. Journal of Hydrology, 456-457:11-29.
doi:10.1016/j.jhydrol.2012.05.052.

Wrede, S., J. Seibert, and S. Uhlenbrook. 2013. Distributed conceptual modelling in a
Swedish lowland catchment: a multi-criteria model assessment. Hydrology Research
44(2):318-333. Doi: Doi 10.2166/Nh.2012.056

Hide

AR by Dominik Paprotny on behalf of the Authors (24 Oct 2016) Author's response Manuscript

ED: Referee Nomination & Report Request started (25 Oct 2016) by Jan Seibert

RR by Anna Sikorska-Senoner (08 Feb 2017)

Suggestions for revision or reasons for rejection

Reviewer’s Comments
The manuscript presents a novel approach of combining Bayesian Network (BN) and copulas to estimate extreme discharge peaks at a large scale of European catchments. Generally, the manuscript is well structured and has been improved after the first revision round. However, some issues still remain and require further improvement or clarification prior to the final publication in the HESS.

In particular, I have two general comments which should be better clarified in the revised manuscript:
1. The presented method is tested with catchments of different areas ranging from less than 100 km2 (how much exactly?) to above 100’000 km2 (how much exactly?). However, only daily discharges were used in the analysis which most likely is a limiting factor for small catchments (<100 km2), where peaks occur rather at an hourly scale. This is also confirmed by the validation results (Table 2.), where all efficiency metrics dropped for medium catchments (i.e. <500 km2) and most likely would drop even further for even smaller catchments (i.e. <100 km2). Thus, I have concerns regarding the application of the method for such small catchments. Could you provide these estimates for different size groups as indicated in table 1 or explain the reason for aggregating all catchments into only two groups? Nevertheless, based only on the presented results, in my opinion, application of this methodology for catchments of a size smaller than 500 km2 should be done with caution and for catchments smaller than 100 km2 should be evaluated prior to application. This issue should at least be stressed in the manuscript in both Discussion and Conclusions. Therefore, the conclusion saying that the method can be applied at any ungauged location (l. 7-8 p. 26) is not fully true (as it was not proven for small catchments) and thus should be corrected.
2. My second general issue concerns two important limitations of the presented method which are not mentioned in the manuscript (or without enough details): a) uncertainty in observed data of discharges used in this study, and b) assumption of stationarity of discharge data over observation period of 50 years. The first issue may be difficult to quantify but may have an impact on the analysis and results obtained and thus should be briefly mentioned. The second issue declines any climate impacts or land use changes (briefly mentioned in the manuscript) over the analysed period of 50 years which may or may not be true for some of the analysed catchments. Both of these issues should at least be mentioned in the discussion or Introduction part.

Major comments:
1. P. 10, Section 2.2. and Table 1. The information about the number of stations available per country, although important, is not that much informative due to the different coverage of the stations. For instance, 90 stations in Switzerland cover a relatively higher area than 133 stations in Germany. Thus, more informative would be a relative value of the number of stations per country area or the density of stations in each country available for analysis. I would suggest to add this information to the table.
2. Figure 5 and 6 panels b-d. The authors compare the simulated discharge peaks with return periods of 10-year, 100-year and 1000-year to “observed” values. How these 1000-yr observed values are obtained? To my understanding, these observed values are not really observed but are computed/extrapolated using 30 years of observed discharges. Please correct this notation properly.
3. Figure 7. I would expect that the current stage (i.e. observation years) is also included in this figure for a better comparison. In addition, I agree with the Editor that presenting this Figure without any uncertainty or critical consideration is a bit risky. Consider modifying the caption to make it clearer that these results are just simulations using specific emission scenarios.
4. P. 17 l. 28-30. These sentences are slightly misleading. Are 30 years of observations used for calibration or validation or for both, and how do you split 50 years of observations into two subsets of 30 years? Providing the exact years, which were used for calibration, should solve this issue.
5. P. 18 l. 13. The case when simulations perfectly fit the observed data (also the line 1:1 when plotted) does not necessarily mean that no bias is present in simulations. Please note that the observed data usually contain measurement errors. Hence, if these errors are not represented, fitting the line 1:1 by simulations may in fact result from the model (simulation) bias. Hence, I suggest to change “no bias” into “perfect fit”.
6. P. 20 l. 1-5 and table 2. What do you mean with the statement that the model performance is/remains acceptable? Could you provide quantitative values supporting this statement?
7. P. 20 l. 26 and afterwards and Fig. 7. The results that authors present do not demonstrate the real trends in extreme discharges in Europe but only the projected trends using specific emission scenarios. Please adopt the paragraph and figure caption respectively.
8. P. 22. The entire paragraph presenting the estimation of RFA analysis fits better into the method and results section that into discussion.
9. P. 24 l. 8. Daily data are not adequate to model both flash floods and short duration floods (of rainfall duration shorter than a day). Both of these floods may occur in small catchments which in fact may explain poorer efficiency of the model in these catchments obtained by the authors.
10. P. 25 l. 7. The given number of 246 years of simulations is wrong. 56 years of observations plus two different emission scenarios each of 95 simulation years do not sum up to 246 years! Because emission scenarios cover the same period of time, only 151 years of simulations are analysed.
11. P. 26 l. 13. Could you say something about the model performance when only some of variables are available at a given location when using BN? I think it would be worth checking (here or in the further work) how the model performance changes depending on the number of variables used/available and if this change is variable specific.

Minor comments, typos and suggestions:
There are also several typos which should be corrected. Below I present just some of them but please double-check your manuscript again.
1. P. 16 l. 29-30. Change into: “…i.e., the value of the node or set of nodes (other than discharges) is defined based on the new evidence…”
2. P. 17 l. 14. Change “30 years of data” into “30 years of discharge observations”
3. P. 17 l. 20. Change “other distributions” into “other tested distributions”
4. P. 17 l. 27. There are more general sources for referring to the maximum likelihood estimation approach than the cited paper of Katz et al. 2012. E.g. Gelman et al. 2013 (Gelman, A., Carlin, J., Stern, H., et al. & Rubin, D. (2013).Bayesian data analysis (3rd ed.). London: Chapman & Hall/CR)
5. P. 18 l. 6. Change into: “quality of return periods and average maxima simulations were…”
6. P. 23 l. 19. Change “were” into “where”
7. P. 23 l. 28. Change into “calculated as a fraction of ….”
8. P. 23 l. 32: change sentence into: “However, one alternative configuration worth mentioning is a BN incorporating terrain classification based on…”
9. P. 24 l. 21. Change into “Climate data from CORDEX have/are of the…”

Hide

ED: Reconsider after major revisions (further review by Editor and Referees) (09 Feb 2017) by Jan Seibert

AR by Dominik Paprotny on behalf of the Authors (01 Mar 2017) Author's response Manuscript

ED: Referee Nomination & Report Request started (13 Mar 2017) by Jan Seibert

RR by Anna Sikorska-Senoner (12 Apr 2017)

ED: Publish subject to minor revisions (further review by Editor) (24 Apr 2017) by Jan Seibert

Dear authors,
thanks for your efforts with the revisions. In general the reviewer and myself now find the manuscript acceptable for HESS. However, there are a few remaining issues as indicated by the reviewer. In addition, I have to say that the language partly could be improved and formulations could be more to the point. I list a few examples below. Please make sure to check your text once more very carefully to make sure formulationsare as clear as possible and to avoid language mistakes (perhaps seek advice from a native speaker or professional copy-editor).

Best regards,
Jan Seibert

Example of language/formulation issues:
Methods:
“Some additional neighbouring basins were added for complete coverage of Europe except for the territory of the former Soviet Union. Only the outermost regions of Madeira, the Azores 15 and the Canary Islands were omitted because they are outside the EURO-CORDEX climate model; notwithstanding, their river networks are very limited and therefore of little interest.”

No need to say that some rivers or of little interest (to which at least local people would disagree), just say that the islands in the Atlantic were excluded.
These sentences also needs to be reformulated:
1) Do you mean upstream parts of ‘some’ basins were added or entirely additional basins? The formulation with some.. added .. except … is a bit difficult to follow.
2) I guess you excluded these islands entirely, but the sentence actually says only some regions on them were excluded, please clarify.

Conclusions: “The accuracy at different ungauged locations however, as seen in Table 2, could differ.”
Please do not refer to a table here, but formulate the sentence (or two) so that it is clear what results you are referring to.
“The advantages of our approach is that it is has low computational expense, it is explicit and flexible.”
Grammatical mistake :… are that. Furthermore, please be more specific, what do you mean by explicit and flexible in the context of this method.

Hide

AR by Dominik Paprotny on behalf of the Authors (03 May 2017) Author's response Manuscript

ED: Publish as is (04 May 2017) by Jan Seibert

AR by Dominik Paprotny on behalf of the Authors (04 May 2017) Author's response Manuscript