Please find here after my comment on essd-2024-406  manuscript. Overall, the manuscript is well written and easy to read. The dataset presented here will be useful for a large panel of users and my recommendation would be that the manuscript is accepted after revisions.

Some questions remain regarding the possible discrepancies within the dataset not being acknowledged (intermission biases, local differences due to geoid, retracking methods) in this version, regarding the way uncertainties are taken into consideration and regarding some methodological choices. The results also should be further criticized on the light of the input data, given that they may depend on the validation method and/or validation data availability.

Hereafter are my main remarks with the corresponding lines in the initial version.

Kind regards.

Section 2.3.1

I miss a § at the end of 2.3.1 stating the differences of the aforementioned databases (open accessibility or not, timeliness, NRT availability or not, etc.), so that the reader understands why several databases were used; This could be done extending Table 2.

Has any inter-validation been performed when/where overlaps (in VSs) were found? Are there any in the literature?  In this section, the reader should understand what were the choices made in case of overlap and why such choices were made. This is crucial for people that would like to duplicate / extend the dataset.

Section 2.3.1 §1 Hydroweb.next

• Please check the satellites list (L97-98) which were used to produce Hwb-Next time series of rivers WL;
• The data sources (L99,100) and the reference apply for lakes and not rivers; For rivers, please cite Santos da Silva J., S. Calmant, O. Rotuono Filho, F. Seyler, G. Cochonneau, E. Roux, J. W. Mansour, Water Levels in the Amazon basin derived from the ERS-2 and ENVISAT Radar Altimetry Missions, Remote Sensing of the Environment, 2010, doi:10.1016/j.rse.2010.04.020

Section 3.1 Construction of VS catalog

• L138-139: please specify the dams and reservoirs database used in input
• Specify that SWORD V16 (from Fig. 3) was used. Is it version dependent, or did you created a customized one, possibly with connectivity issues corrected? Will the code for creating this VS catalog with connectivity and hydrological constraints be made available to the community together with the database?
• L142-143: put the ENVISAT series list in chronological order

Section 3.2 WL time series

• Since SAEM includes L3 WL time series, I consider a dedicated paragraph on inter-mission biases is mandatory. Even though discharges are estimated through mono-mission non parametric curves, such biases shall be taken in consideration by whoever want to use the L3 WL TS from this database or any other all together to build long TS. Moreover since different retrackers are used among the missions.
• L152-153: this masking choice implies, for past missions, to keep a very low number of "H_i" measurements from your raw altimetric data. For small rivers, this would mean keeping only the one measurement geolocated over the water. What are your statistics on this point (percentage of dates with <2 H_i measurements for example), and do this have an influence on the final median that you process (and which uncertainty do you provide in this case)? I believe this should be further explained.
• L171: Is XGM2019 used in one of the L3 databases used in this study? Differences between gravity field models can lead to lat/lon dependent biases between the series from external databases (e.g. Hydroweb.next) that are on EGM2008 and the L3 processed time series. In a matter of uniformity, using the same GGM would be recommendable.
• L180-184: it would be worth having a table with those statistics. I.e. how many TS were generated, how many (in %?) passed the QC check (total and by mission), and why.

3.3 Non parametric rating curve

• L187-193: I suggest discriminating the advantages as a function of the approach. Indeed,
  
  -  the "does not necessitate simultaneous ..." is not due to the non-parametric and could be also the case for linear regression;
  
  - "it assumes no specific predefined ..." is only for the non-parametric, and it should be stated why this is an advantage comparing to other formulations;
  
  - realistic uncertainty: is not dependent on the formulation but comes from the MC simulations and the creation of a stack of WL and Q.
• L206-208: This means that the uncertainty raisen from previous steps is not being used. Why this choice? The “10% multiplicative uncertainty” should be better explained. Also, a sensitivity analysis of the MC algorithm to the uncertainty (and consequently, the sensitivity of the non-parametric RCs derived) shall be investigated (testing different bounds for uncertainty, and even informative uncertainty such as the one coming from WL processing).
• L208: Convergence of MC algorithms is an important point. What convergence criteria was considered, and did all the Q/H Ts reached convergence ? If not, how many did not passed, and on which reaches, rivers, etc?

5 Results

• L275-279: A statistics on the step where data was rejected could be very important: was it due to a lack of raw (gdr) data, during the WL processing, during the NPQM, quality test? etc. The visualization could be done maybe with a geographical (lat/lon) view?
• L300: remove the “exceptionally”. The model performs indeed well in these regions, yet KGE and Corr remain in the “good” domain and not in the “exceptional”
• L305: “the overall good accuracy” instead of “high accuracy”, for the same reasons than above
• L306: is 73% the right value? Please check, this seems pretty high to me.
• L310-311: I am quite surprised by the low values show by S3B (same as T/P Jason). It is important to bring some elements of explanation in the discussion section, since S3B is expected to perform at least as good as S3A, at least in terms of WL estimates. What can explain such difference? The length of the TS? The lack of in situ data for validation on the recent period? Anythink else? Please discuss it. Regarding Topex/Jason also, are all the missions giving similar results? Or is the global quality impacted by the oldest missions? When ENVISAT is mentioned, is it only ENVISAT or is it the family (ERS/ENV/SRL) ?
• L345-348: provide other metrics to complement KGE (KGE improvement of 0.15 is somehow unease to quantify) (e.g. NRMSE in %, other)
• L361: CCI WL rely mostly on hand-processed time series using Altis software (and in particular it is the case for Niger TSs, please correct.
• Figure 10: I would rather show the comparison in terms of anomaly (or also show), in order to better evidence differences between SAEM and CCI for all the discharge amplitude (can be anomaly or normalized anomaly). Differences for small discharges do not appear clearly here. Same for WL.

5.3 Discussion

• I miss a paragraph dedicated to the discussion around uncertainties. For example, does the CCI RC fall inside the uncertainty bound for SAEM RCs? What is the contribution of providing uncertainty in discharge estimate. Does this uncertainty is useful, and in line with other (e.g. CCI) provided uncertainties?
• L420-425: This matter is worth discussing a little more. See Kirchner 2006 "getting the right answers", the choice of non-parametric vs parametric RCs can be discussed here, bringing together the advantages (as already shown) but also the drawbacks of both methods (hydraulic-based vs data-fit based)
• As said above, the relative performance of SAEM discharges as a function of the mission should be discussed, since the classical “quality evolution” with time is not respected here. This can be due to several aspects, which need to be discussed so that the reader/user understand what lies in this validation statistics.

General comments - Overall quality of the preprint

This manuscript presents a novel dataset, called Satellite Altimetry-based Extension of the global-scale in situ river discharge Measurements (SAEM), that leverages multiple satellite altimetry missions to extend discharge estimates based on H-Q non parametric rating curves calibrated at in situ gauges. Median KGE is 0.48 for 8730 gauges extended. A catalog is also introduced, providing information (location, distance to gauge, discharge series, water level series, rating curves for the mapping…) on the virtual stations used to obtain good quality discharge close to the gauges.

I believe this manuscript to be of very high relevance to the hydrology and remote sensing communities as it is the very first to attempt creating a global dataset of discharge estimates from altimetry virtual stations alone (where most studies until now have been at continental scale at most, or using imagery-based discharge estimates predominantly). Whatever the accuracy achieved, this is no small feat, and it will be particularly useful in the context of the recent SWOT mission, which should deliver an official discharge product globally. Congratulations to the authors.

It is well written and organized, and I recommend acceptance after (quite a few but) minor revisions.

Best regards,

Specific comments - Individual scientific questions/issues

Intro

• l.27: “volume of water …”: rather “volumetric flow rate”? you are missing the per unit time information. (same remark in abstract)
• l.29: is confronted with?
• l.35: I would remove “directly”. Hydraulic variables can be rather approximated with complex measurements and processing.
• l.35: the seminal paper for water levels is I think Birkett, 1998 (10.1029/98WR00124), similarly to Smith, 1995, 1996 for river widths
• l.36-37: and even where no gauge exists, using algebraic flow laws or hydraulic models.
• l.37-41: for rating curves, you can cite Leopold and Maddock, 1953, who pioneered the notion of hydraulic geometry relationships, whether from depth, width or speed.
• l.42-51: I think you can make for an even stronger argument to your manuscript. There are a few studies that went global on discharge from satellites, but mostly from an imagery standpoint (i.e., using width-based rating curves), e.g., Elmi and Tourian, 2023, Elmi et al. (2024), Riggs et al., 2023, but also Lin et al., 2023 (10.1016/j.rse.2023.113489) that you don’t mention. To my knowledge, none have gone global on altimetry only!! So you are the first, that’s something.

Section 2

• Table 1: N gauges for USGS and Chinese database are the same, typo?
• l.97-98: to my knowledge Hydroweb.next does not contain ERS-1, TOPEX, Jason-1 nor GFO virtual stations. Please check? Also, CLS company is in charge of retracking for the most part I believe?
• I think you can mention that Hydroweb.next (and DAHITI?) are the only databases currently providing “live” level-3 data? The big advantage of your catalog will be for people to use your rating curves in the future as new WL data comes out.
• Overall a few details on the extent to which level-3 databases differ would be helpful (retracker etc.). What if the same VS is present in different databases, which WL series do you retain?
• Figure 2: it is kind of hard to see all colors (especially Hydrosat in orange), but I don’t have a way around that…

Section 3

• 3.1: SWORD (being Landsat-based from GRWL) has very few low-order reaches (compared to MERIT that’s DEM-based for instance). From experience, a handful of your gauges are located on reaches that are not contained in SWORD. So you do not associate VSs to these gauges?
• Figure 3: water occurrAnce typo
• Figure 3: Again, what happens if the gauge is located on a small tributary reach not contained in SWORD?
• 3.2:
• l.151-153 what about when/if GSW and SWORD don’t line up (even though they are both Landsat-based), you don’t get any h_i?
• l.160-161: Overall I am not competent enough to review retracking skills, but it would be good to have a little more information on, for instance, what retrackers are used in the level-3 databases compared to what you use for SAEM, and how they differ.
• l.171: the geoid model chosen differs from, e.g., egm2008, used for other retracking. Does that impact your WL series?
• l.180-184: the outlier removal procedure is referred to in Tourian et al., 2022, ok. But for the quality control, what does it take to “pass” the control?
• 3.3:
• l.189-190: I would say” This method overcomes several limitations, since:” The fact that it does not necessitate simultaneous measurements of gauges and WL comes from the use of quantile mapping, and is not related to linear regression.
• l.201-203: what is the 3 sigma test? Could explain a little more. From what I understand in the matlab code, we update at each iteration the perturbation/uncertainty applied to the Monte Carlo simulations to the max of the distance between the estimated/simulated and measured quantiles divided by 3.
• 3.4
• l.222-225: not a big fan of these statistical tests as I find they are usually rather easy to pass. I would just trust more KGE and its three components, NBIAS, NSTDERR and Corr.
• l.226: case 2 no simultaneous measurements. KGE between what? Monthly mean aggregates? Seems that is the case because that is what you are using for the SW test after.
• Visual inspection on so many VSs, what a tedious job!
• l.232: only 1400 cases rejected out of 43000? I would expect way more?

Section 4

• 4.3: so these rating curves are the 100 quantiles for WL and Q mapped from NPQM? So that anyone can interpolate a new WL observation into discharge using the 200 data points?
• 4.4: Are the uncertainties given by quantiles in the RCs as well (from the standard dev of the stack of MC simulations?)? So that users can obtain uncertainties for future discharge estimates they compute?

Section 5

• l.287: “mostly exceeding 0.76”: the median (25^th perc) corr is 0.65 (0.4) later in Figure 8.
• l.288 + l.289-290: correlation is important, but I wouldn’t say it indicates robust nor reliable performance alone. WL and discharge should be in phase, otherwise one of the two is just wrong. Robust performance to me is also about having very little bias and similar standard error (i.e., all three components of KGE).
• l.295-296: Is the NRMSE normalized by the mean of obs, the RMS of obs, or by the difference between min and max of obs? Values can vary immensely depending on which is used…
• l.306: median NRMSE from the CDF looks more like 18% than 73%!
• l.310-311: Sentinel-3B should perform similarly as Sentinel-3A. So it is probably an issue of WL time series length (S-3A starting 2016, S-3B 2018). Discuss it maybe?
• 5.3 Discussion:
• l.407-411: Regarding the temporal sampling limitations (10-day revisit at best for Jasons/Sentinel-6), you can say that it is impossible to capture rapid events on the order of a few days (Cerbelaud et al., 2024, found you need to sample 4 times more frequently than the event duration), even with densification algorithms… This means that small, low-order river reaches upstream of basins (which are typically the ones showing rapid changes) cannot be accurately captured. They are too narrow anyways for retracking…
• l.420-425: You could discuss the issue of potential overfitting to data in NPQM versus underfitting in parametric RC (e.g., when a two-step a(H-H₀)^b RC is needed) in light of the comparisons described in 5.1.

Conclusion

• The conclusion on which mission performs best begs for an explanation. Why does ENVISAT work better than the Jason series when ENVISAT has a 35-day revisit, and Jason a 10-day revisit…? The range window size of ENVISAT? To me Jason-2/Jason-3 should provide so much better results than ENVISAT.