Thank you for your comments.

While we do not expect to provide a definitive ranking of products under different conditions (new products and versions are frequently released and such a ranking would rapidly be obsolete), we have collected the newer SEBS dataset and will replace the older SEBS data in our analysis for resubmission.

Thank you for your comments and recommendation. Our respnses are provdied below in bold.

General comments

C1.I believe the authors could bring their analysis one step further by better discussing why some RS product combinations perform better than others for specific areas / catchment characteristics.

In order to understand why certain products perform better in certain catchments, one needs to know what are the general differences between the products for the same variable (i.e. measurement principle, data sources, idea of the algorithm, etc.). A brief explanation of the differences between each product could be added to paragraphs 2.1.1, 2.1.2 and 2.1.3. In chapter 3, the authors could refer to these differences when explaining why for example one ET product performs better than another ET product, in combination with the same P and dS/dT products, in certain types of catchments. This would be very valuable information for those who have to choose products to estimate discharge.

We will add some detail on the differences between the products as we agree they could give useful insight into the different model performances.

C2. Consider adding a Figure/Table which shows the best performing product combinations for certain common combinations of catchment characteristics. This would be most informative for water managers of ungauged basins.

Some of the products we used in the study are no longer produced in the versions used and there are continuous improvements to the algorithms used to produce the data. We therefore do not aim to produce a final classification of best/worst products which would likely be obsolete by the time of publication, but rather provide a consideration of the factors which influence the closure of the water balance. We believe that this is more valuable to users than a ranking.

C3. Most of the data sets used are products from missions that are not in orbit any more. I think it is relevant for the target users to describe what are the current / future missions and products, which are comparable to the missions used for the current analysis. Some of this information is given in a footnote in Table 1 (GRACE-FO, TMPA), but given its relevance for the users of this work, I think this deserves a paragraph in the text.

We will add more detail on this in the text.

C4. Did you consider adding dominant soil type and/or subsurface bedrock depth to the catchment characteristics? These are important factors affecting the discharge of a catchment, particularly when looking at different continents.

We did not add these, but agree they could be of interest. We will add these as suggestions, in particular for the soil type – we are however not aware of a globally available subsurface bedrock depth product which could be used in this study.

Line-specific comments

C5. In line 77-103, a clear overview of previous related studies is given. However, I find the novelty of the current study a bit underexposed in lines 104-107. Both aims that are mentioned (i.e. to investigate both the ability of different combinations of RS products to reproduce in situ measurements of discharge, and to identify catchment characteristics that affect how well the closure of the water balance can be achieved) have been studies before. Here, it is not yet clear what the current study adds to the previous studies, and why this study is necessary. I recommend to formulate the gap in knowledge and added value of the current study more clearly in this section.

We will make the novelty clearer.

C6. L54 / L70: Equation 1 and 2 are exactly the same, while Eq. 2 should be a rewritten version of Eq.1.

Yes, thank you for spotting this error. Eq. 2 should be Q = P – ETa – dS/dt

C7. L66 “This is equivalent to the assumption that subsurface fluxes in and out of the basin are negligible”.  Consider referring to Bouaziz et al. (2018), who disprove this assumption.

Thank you for the reference, we will add a comment on the assumption and the reference.

C8. L440: “100,000 km2”. Based on previous text, I assume this should be 10,000 km2.

Yes, thank you for spotting this error.

C9. L233: “We selected 11 RS derived catchment characteristics… . These are summarized in Table 2 …” Table 2 only shows 10 RS catchment characteristics.

Indeed, we will fix this error throughout the paper.

C10: L395: “This was unexpected as the GRACE data in particular is expected to perform better for larger catchments” Given the next sentence, I think this sentence is redundant.

We agree, the text will be edited accordingly.

C11. Fig. 5: The colors of SSEBOP and and GLEAM are too similar. It is difficult to distinguish the two. Please adjust the colors.

We have changed the colormap and removed the grey background for improved legibility.

C12. Fig. 6: Consider revising the color scheme used. Make sure the colors for a positive correlation can be clearly distinguished from the colors for a negative correlation. Now, the green color representing a Pearson correlation of 0-0.25 is similar to the green/blue colors representing a negative correlation.

We had chosen a color scheme that could work in black and white but agree a diverging color scheme adds to legibility and we will adjust the color scheme.

References

Bouaziz, L., Weerts, A., Schellekens, J., Sprokkereef, E., Stam, J., Savenije, H., & Hrachowitz, M. (2018). Redressing the balance: quantifying net intercatchment groundwater flows. Hydrology and Earth System Sciences, 22(12), 6415-6434.

Thank you very much for your comments and recommendation. Our respsonses are in bold below.

Main issues:

* Relevant context on the GRACE spatial resolution and signal leakage is missing for the particular products used. In the study, gridded GRACE products have been used. This is in principle ok, but it the inherent spatial resolution of the GRACE derived total water storage is important and should be mentioned. The truncation (spherical harmonic) degree of the input gravity field solutions in combination with the strength of the applied spatial filtering will result in highly smoothed fields and signal attenuation (and contamination from nearby sources). I suspect that the smoothing may also contribute to the lacking correlation found between basins area and the obtainable NSE: below the typical GRACE resolution the results for the sub-catchments will result in essentially the same GRACE times series.

I see 3 ways out of this conundrum: (1) filter the P-ET data with a spatial filter which has similar smoothing as that was is applied GRACE, (2) try to restore the signal in GRACE (see e.g. Vishawakarma et al. 2016 below or similar references), or (3) leave it as is but write a disclaimer in the discussion on how this can effect the results.

Way (2) (then (1)) would be the best but it would essentially constitute a lot of work. I would find (3) also acceptable in light of the, likely larger, errors introduced through the P and ET estimates.

The inherent GRACE resolution will be added to the manuscript in the result section. We do note the issue of signal location and leakage, though not in detail, when discussing GRACE’s ability to locate punctual storage changes from dams (l.252-253)

Regarding the suggested options, our goal in this paper is to apply the simplest approach to solving the water balance – following the approach that a standard application user of these products would – and because of this we will be sticking to the use of the gridded GRACE data as is. However, we will make the issues brought up by the reviewer more explicit in the paper.

* How are biases treated in R_estimated and R_observed?  If biases are large, some of them may strongly influence the NSE, in particular for arid catchments (bias is larger compared to the standard deviation). Maybe a NSE variant can be computed where the biases are removed?

In the paper, we aim to assess the performance of the products as is, including potential biases. We therefore did not perform any bias corrections on inputs or outputs. We also chose NSE as we wanted the results to be understandable to practitioners and NSE is the most commonly used indicator for hydrological model performance. For this reason, we would prefer to retain the NSE, but will add a sentence explaining this, and the reason why biases are not considered in the study.

Note that at higher latitudes, errors in the Glacial isostatic adjustment trend-correction for GRACE may also manifest themselves as biases in the storage changes, although I suspect that this is not a big issue as the GIA models generally perform well over the Northern Hemisphere (where most of the considered catchments are)

Thank you for this note, we will look up and add some reference to this in the discussion on the latitude dependency of the results.

* North American nested-catchments may be disproportionally present in the catchment characteristics comparison. The authors mention that several catchments are nested-catchments of their larger parent groups. Is this somehow compensated when computing the correlations of the catchment characteristics? If not I wonder whether e.g. the Missisippi basin, which is well gauged may be overrepresented in the statistics. The authors may consider limiting the amount of subcatchments which are fed into the statistics, if this is an issue.

It is indeed likely that this has an influence on the general statistics – we include this limitation in lines 305-308, as well as the fact that these catchments will typically be in (for example) similar climate zones. We chose to use all available data throughout the paper, otherwise questions of which catchments to keep and remove become intractable.

* Suggestion: Use histograms (cathment nr. on y versus NSE on x) to visualize and thus understand the distribution of the obtained NSE's. It would also help justify the computation of the correlation of the NSE (essentially a correlation of a correlation), which implicitly assumes that the computed NSE's are normally distributed.

Thank you for this suggestion, we will include a histogram for better visualization of results, and you are right, the NSE values are not normally distributed. We therefore decided to compute the Spearman rank correlation as it is non-parametric and does not rely on the assumption of normality. We found that for 6 of the 8 tested continuous catchment characteristics, the general pattern remains similar though with higher correlations and more significant correlations found (see figures in supplement). For the remaining 2 characteristics (Ro_yearly/Sdam and Ro_yearly/P) the pattern is inverted for one, and it is hard to tell for the other (Ro_yearly/P) as Pearson mostly returned non-significant values. We will revise our paper to include the Spearman rather than Pearson results.

Minor issues

* Please mention how aggregate catchment values are obtained from the gridded values. Are grid areas taken constant, or are they latitude weighted?

Thank you for this comment, we will recompute the time-series to include latitude weighing and detail the process in the resubmission.

* I would like to ask the authors to consider releasing their analysis code, to improve reproducibility for others

We will organize our code and release it.

References:

Dutt Vishwakarma, B., Devaraju, B., Sneeuw, N., 2016. Minimizing the effects of filtering on catchment scale GRACE solutions. Water Resources Research 52, 5868–5890. https://doi.org/10.1002/2016WR018960

Thank you for your comments, we have responded to them in bold below.

This reviewer appreciates the effort of the authors to investigate sources of variability in closing the terrestrial water balance with remote sensing. The novelty of this study is to link the catchment characteristics (and LCC, LU, and climate class) to the performance of different product combinations. Although the manuscript collects multi-source RS water cycle variable datasets, it is unclear why certain product combinations perform better than others. For example, it is unclear the impact of different spatial resolutions of different products on the final results.

This reviewer suggests the authors put more effort into clarifying the above perspective, for example:

Clarifying and quantifying the potential impacts of using RS data from different spatial/temporal resolutions;

In terms of quantifying such impacts, the triple-collocation type of approach could be applied here among multi-source RS data products to understand the relative errors between each other, which will provide further information for identifying the sources of variability in closing the terrestrial water balance using different product combinations.

In terms of the varying spatio-temporal resolutions of the products we think that as we are running the comparisons on a basin-wide and monthly basis, it would be difficult to identify these issues directly.

Further, as far the triple-collocation method is concerned, one challenge in application is that it requires independent datasets to be used. This is not the case for most of the products as many of them use the same input datasets and/or equations and parameterizations (as noted e.g. on line 153 of the discussion paper for precipitation products – this will also be clarified for other products).

Minor comments:

Eq.2 looks exactly the same as Eq.1

Thank you for pointing this out, Eq. 2 should read: Q = P – ETa – dS/dt

Line 36 "the results point to the importance ..." is confusing

We will adjust the line for clarity.