Editor Comments on:
"Design flood estimation for global river networks based on machine learning models"

By Gang Zhao et al.

This MS attempts to derive a DF for global river basins using a large sample (n=11793) using state-of-the-art ML techniques. I welcome this effort and consider that that the HESS readership will do so too.

Three reviewers have provided numerous comments that need to be addressed by the Authors in the revised manuscript. There are many comments regarding the notation, Methodology and the statistics that are not clear.

In my opinion there are major shortcomings that need to be addressed before this MS is accepted in HESS:

1. The authors should put in the appendix all statistical test that are standard, e.g., Mann-Kendall test and any test that has been already published. Writing once again text-book equations only adds bulk to the MS but not insight. Same for K-means et. If need, write all tests in an appendix for reference. Table 3 to appendix. The same with Anderson-Darling (AD) test

2. I am missing the coefficients of the potential models for the regionalization of the design floods (DF) in eq. 25 and their confidence intervals. See how we reported in Samaniego and Bárdossy, 2005 JoH (SB2005). In this regard too, how did you select the optimal number of predictors? Which methods was used? Are the number of predictors the same for all regions? How the parameters (\beta_i) vary from region-to-region? Is the change significative?

3. In the K-means phase: did you test another metric? For example Mahalanobis ? There is not reason why the Euclidian metric describe the manifold of the predictors in the best way. Even if the Euclidean distance is the best, perhaps there is an embedding space (u) that betters describe the relationships between predictors (x). See other possibilities in Bárdossy, Pegram, & Samaniego WRR 2005 or in Samaniego, Bárdossy, and Schulz IEEE TGRS 2008.

4. Minor: use color as suggested in https://colorbrewer2.org for the map and other graphics.

5. Math notations should be consistent. Regression models must have an error term. (e.g., in eq. 25). By the way, are the errors of the DF models homoscedastic? Please comments and show tests. For the regression models, which estimators were used? Please provide Jackknifed (leave one out) bias andRMSE (see paper SB2005). this is important to understand if the coefficients are robust of just the result of one outlier, which is very likely in this global data set. L1 estimators are recommended in this case.

6. Descriptions of RF, SVM, MCMC, etc. are out of the scope of this manuscript. Here the authors should indicate how they make assumptions, set thresholds, etc. Then show results, and compare the various methods. Show which is the best in cross-validation experiments.

7. Provide uncertainty bounds of the statistics in Table 4. The same for figure6. A table with the factors used for the regressions should be included. I can't find them? in Table 4, which are the Meteorological, Physiographical, Hydrological and Anthropological factors? See in SB2005 how we described them. How the optimal nr. was found? The data set should be made available too. I see a description in L370, but a table with basic descriptors (max, min, mean ) is useful in the appendix

8. Use more pleasing colours in Fig. 2. (suggestion above)

The authors dis a great job addressing all Rev comments and mine. This MS can be published as it is.