General comments

This study examines the impact of four gridded daily precipitation products on daily runoff simulations in ungauged sites. The evaluation is tested using a large sample of near-natural catchments in Chile. The regionalisation of model parameters includes three methods (spatial proximity, similarity and regression). The results are evaluated by various runoff efficiency criteria and are compared over diverse hydrological regimes across the study region. The results show any noticeable impact of spatial resolution of precipitation product on runoff model efficiency. Also, the precipitation products corrected with daily gauge observations did not necessarily translate into improved hydrological model performance. The precipitation products with the best performance during calibration and/or verification did not necessarily provide the best simulations in ungauged sites. The best regionalisation approach is the similarity, regardless of the choice of gridded precipitation product or the calibration criteria. The results indicate that the performance of regionalisation depends on the hydrological regime.

The study presents an interesting analysis that adds some new understanding of the impact of using different precipitation inputs to predict daily runoff hydrographs in data-sparse and ungauged catchments. The manuscript has a good structure and reads well. I like the comparative focus of the analysis. The use of a large sample of catchments in very diverse climate and runoff generation conditions brings interesting and new findings. I have only a few remarks or suggestions which might be considered for extending some of the results and supporting interpretations made. These include:

1. The main focus is to evaluate the impact of selected precipitation products on regionalisation model performance. Still, the description and demonstration of the similarity and differences between the products and eventually at site measurements are very brief. Comparing of the long-term averages gives some impression about the consistency between the products, but it will be very interesting to see (and evaluate) some more detailed analysis of the differences between these products, such as the difference in seasonal distribution, intermittency/frequency of rainfall, days without rain, daily extremes etc. Can such additional differences/similarities explain more the differences between the results (i.e. calibration/validation/regionalisation performance, methods, time periods)?
2. The testing of regionalisation performance for the full period of observations (1990-2018) somewhat mixes the interpretations. I like the idea of splitting the period into calibration and two validation periods showing different climate conditions. So, testing the regionalisation model performance for the same periods can bring some additional information about the performance of different methods in different climate conditions and also show the loss in regionalisation compared to calibration and validation efficiencies.
3. The results indicate that the calibration procedure can compensate for some differences in precipitation products. It will be interesting to see which model parameters/runoff generation processes are affected and whether and how the parameter values change between different hydrological regimes. Are the other simulated components of water balance similar as well (between the products)?
4. It will be interesting to see some disentangling difference between similarity and spatial proximity methods for linking the results with previous studies. Is it the averaging of ten simulations or the similarity between the catchment attributes that impact the regionalisation performance? What is the impact of averaging ten simulations compared to simulations based on the most similar catchment only?

Specific comments

Table 1: Leave-one-out and jackknife cross-validation are likely the same procedure. Please consider using, for consistency, only one term for the same procedure.

HESS

Review of Manuscript # hess-2021-156



Title: On the selection of precipitation products for the regionalisation of hydrological model parameters



Baez-Villanueva et al.



Summary

The study investigates the impact of four different gridded precipitation products (ERA5, MSWEPv2.8, RF-MEPv2, and CR2MET) having different spatial resolutions on the relative performance of three regionalization techniques (spatial proximity, feature similarity and model parameter regression) over 100 near-natural catchments in Chile, characterized by varying topography, climate and hydrologic regimes.  TUWmodel, a semi-distributed HBV-like model, was driven separately by four precipitation products and calibrated/evaluated for each catchment using two performance metrics (KGE’ and AOF) utilizing daily streamflow. The authors compared the performance of the regionalisation techniques through a leave-one-out cross-validation exercise, which consists of leaving out each one of the 100 catchments, transferring model parameters, conducting flow simulations and computing performance metrics. They concluded that the calibration procedure is able to compensate for the differences in precipitation products and the spatial resolution of the precipitation product does not largely affect the regionalization performance. Overall, feature similarity provided the best regionalization performance followed closely by spatial proximity, while parameter regression performed the worst. They also reported that the hydrologic regime impacts the performance of regionalization.

Overview Comment:

I would like to thank the authors for conducting this interesting work. The topic of the manuscript fits well to the journal scope and readership. The use of language and structure is well. The authors put significant effort to summarize the outcome of substantial number of cases (four precipitation products, three regionalization techniques, 2 objective functions as well as hydrological regimes) into a single manuscript. However, it is still difficult for the reader to follow through and understand the full reasoning behind some of the study outcomes. My suggestion for improving the manuscript would be to explain the link between different precipitation products and calibration/verification and regionalization methods through differences they make in model components (parameters/fluxes) and components of the efficiency metrics (KGE and flow duration curve segments). I listed my main and minor suggestions for improvement/revision in the sections below.

Main Comments:

1) Discussion of how the model components compensated for differences in precipitation products: There are significant differences in precipitation products (for example ERA5 reported four times higher precipitation in the dry north), however most of the calibrated model parameters have similar distribution for ERA5 and other products (Figures S4 and S5), except SCF, Lprat, FC, beta (I am surprised to see that parameter distributions barely touch the limits). An analysis of how the model compensated for differences in precipitation products through parameter values and fluxes other than streamflow will significantly improve the reasoning behind study outcomes. For example, did the large bias in ERA5 in dry regions compensated by more evapotranspiration/groundwater loss etc? The authors also mentioned this topic in Lines 64-67: “Although hydrological model calibration can partly compensate for errors in the representation of P, this may lead to unrealistic model behavior, thus affecting the quality of parameter regionalisation results” and in Lines 457-459: “The equifinality of model parameters may also impact the relative performance of the regionalisation techniques by producing unrealisitic parameter sets, particularly for the case of parameter regression.”. Therefore, providing answers to these statements will significantly improve the manuscript.

2) The authors utilized KGE’ and signature-based objective functions for calibration, both of which can be decomposed into hydrologically meaningful components to understand differences in model performance driven by different precipitation products as well as in different hydrological regimes. An analysis of how well different hydrograph characteristics (variability, bias, low flows, high flows etc) are reproduced through regionalization will significantly improve the manuscript.

3) Line 226: More information on the optimization algorithm parameters should be provided, such as termination criteria, maximum number of iterations permitted etc. to better understand the calibration procedure and its impact on calibrated parameters.

4) Lines 408-410: ERA5 re-analysis product is also a merged product and uses ground-based measurements (after 2009). Please discuss and revise where necessary. (see  Hersbach et al., 2020; https://rmets.onlinelibrary.wiley.com/doi/10.1002/qj.3803). Possibly focus on differences in gauge density used in correction of different precipitation products.

Minor Comments:



Line 123: replace “perform well over small catchments” with “perform well even over small catchments”

Line 124: replace “tend perform” with “tend to perform”

Line 127: The reader would be interested to see how well CR2MET performs compared to ERA5. A few sentences based on the reference will be helpful.

Line 128: CR2MET: Boisier et al (2018) states that CR2MET is produced from ERA-Interim but not ERA5. Please check.

Lines 149-154: It will be interesting to report the control of elevation to the precipitation differences in the products. Representative West-East elevation and precipitation cross sections for three zones may help in this regard.

Lines 166-171: There are over 20 references in 6 lines. Reducing this number to 2-3 for each point the authors are targeting may help brevity. Same in line 225.

Figure 4: Caption: Replace “Leave-out-out” with “Leave-one-out.”

Figure 5: The ECDF color-coding description should be corrected in the caption based on the legend (calibration-green, Ver-1 blue). I also suggest drawing the country outline thinner and with less intensity to improve visualization of the markers. If possible, a light gray shaded relief map will add elevation information.

Line 325: CR2METv2 – please use consistent acronym throughout the manuscript (earlier use was CR2MET).

Lines 335-337: “Choice of objective function does not impact the spatial performance of regionalization methods”: This sentence seems to be an overstatement because different objective functions are sensitized to different components of the hydrograph (e.g log transformed flows will put more emphasis on low flows) and hence put more emphasis on fitting different model parameters. Instead of using a general statement, please use a statement valid for the present study. In Figure 5, comparison of ECDFs for P product columns between KGE’ and AOF metrics indicates some differences. For example, compare feature similarity ECDF for ERA5 between KGE’ and AOF metrics.

Lines 347-348: Please discuss which catchment characteristics are lacking that represent the hydrological behavior in this region.

Lines 354-355: “For feature similarity, all P products yielded a performance similar to that obtained for individual basin verification during the dry Verification 2 period”. This statement somehow contradicts with the information provided in Figure 6. In Figure 5, ECDFs for regionalization and calibration/verification represent different time periods and hence their comparison may yield to misinterpretation. In Figure 6, however, regionalization and calibration/verification periods overlap and it can be seen that feature similarity performance is lower than Verification-2 performance.

Line 376: Indicate that Figure 8 is for AOF.

Figure 9 and Lines 401-402: “The performance of parameter regression did not change substantially when evaluated with and without nested catchments.” Actually, the performances with and without nested catchments look precisely the same (median values, distribution etc.). Is this related to using one RF model for each TUWmodel parameter? (see Lines 283-287). Please clarify here (Lines 400-401) and in Lines 283-287). I see some explanation in Lines 499-502, but more direct evidence should be provided.

Lines 416-417: MSWEP is also a merged product as stated in Line 411? Please clarify.

Lines 422-423: See earlier Comment on ERA5. Also, ERA5 reported 4 times more precipitation in the dry north (Figure 2) that may lead to low regionalization performance.

Line 424: Please revise the language: “The use of ground observations as they:”

Line 465: Please clarify “opposite” here. Does it refer to the regime?

Line 477: Please move the left parenthesis before the year of the reference.

Line 489: Please indicate “snowmelt” recession curves.

Line 494: Typo “Figure” 9.

Line 512: Remove “model” after TUWModel.

Boxplots: Please describe box limits and vertical line (median) in first boxplot figure caption.

Lines 518-519: See earlier comment on ERA5 (merged product).

The manuscript presents the results of a certainly massive amount of work, aiming at analysing the performances of a set of gridded rainfall products when regionalising the parameters of a rainfall-runoff model, to be used at ungauged river sections in a data-scarce region. It addresses therefore a very interesting, and novel topic, with important practical impacts and potential for improving predictions in ungauged basins (PUB) in many regions of the world.

On the other hand, the topic is complex and the aim of the work is perhaps a bit too ambitious for a single paper.

Given the main final focus of the work, the paper requires first of all a more thorough analysis and comparison of the rainfall products, that should be carried out at catchment scale and over the different seasons (see, for example, Tarek et al, HESS 2020). The interpretation of such comparison should be based on the knowledge of the differences among the rainfall products, considering and explaining in particular the differences among their sources, first and foremost the availability and use of raingauges (that I personally, as a hydrologist, still find the most reliable information on actual rainfall). Such analysis is currently only partial in the paper and it is probably its weakest point.

Following such analysis, a more detailed presentation of the performances of the streamflow simulations when the model is “regularly” calibrated (not regionalised) may allow to understand if the products are indeed equally suitable to reproduce the rainfall fields over the catchment in a “reasonable” way, through the simulation of the rainfall-runoff transformation processes. I totally agree with the authors that the parameters may compensate for the differences in the rainfall fields, up to a given extent, and such analysis is therefore very interesting, especially for a data-scarce region. Some of the interpretations in the discussion and conclusions (see reference to Fig. 5 in Sect 5.1) on the compensation are referred to the regionalisation results, while they should be demonstrated through the analysis of the results of the “calibrated” simulations.

Such first two steps would be enough for a paper of its own, without considering the regionalisation that is built on top of them. The importance of the rainfall estimates is certainly crucial also for model regionalisation purposes, but I think this is mainly due to an indirect impact of their reliability on the different climatic areas, extremely diversified in the Chilean region, and not directly to the regionalisation approaches: for this reason, a deeper understanding of such reliability is mandatory.

In an already very complex framework, the authors have added a second, complicated objective function (AOF), which would be justifiable only if the study focussed much more on the analysis of the simulation of the different parts of the hydrograph (low flows, flow duration curves, baseflow…), which is not instead the focus of the present work, and the issue is only partially addressed, and only late, in the discussion section on the hydrological signatures. As it is, the second OF lengthens and complicates the paper, and I would suggest to remove it.

Another analysis that I think may be removed or restructured is the comparison of the regionalisation performances over the different hydrological regimes. This is very interesting issue to study per se, but i) the rationale for the separation of the basins into such groups (“visual screening”?) has not been explained (and the hydrological regimes seem to overlap only partially with the identified climatic regions, which I find a bit puzzling) and ii) the results of section 4.2 are not particularly significant: the rainfall products do not show any clear pattern and the differences are probably more due to the more pronounced rainfall errors or difficulties of the model in reproducing the streamflow over the different regimes, independently of the regionalisation procedure. In order for such section to be useful, it should be preceded by an analysis - over the catchments belonging to such regimes - of the reliability of the rainfall maps and of the model performances when the model is not regionalised but calibrated (other analyses to be added in the first two steps…)

SPECIFIC COMMENTS

ll 68-74 and Section 3.1.1:

As I wrote above, the main limitation of the study is the lack of a detailed description of the rainfall products: the main difference is in their sources, whereas it is not a matter of being gridded and probably also the spatial scale is less relevant than how much they rely on ground measurements. The ERA5 is probably the product with less dependence on raingages, since reanalyses assimilate a number of both measured and remotely sensed information within the numerical models, but mainly atmospheric and ocean measures and not ground data, I think. But all the other products you use are, on the basis of your description, based on a merging of reanalysis (ERA5) and ground-based data (CR2MET and RF-MEP), and for MSWEP including also use of satellite data. For all such products some information on the location and temporal coverage of the raingages data is needed, in order to understand their differences. It is written that RF-MEP uses 331 gauges, but not if CR2MET uses the same data. And MSWEP is based on 77000 gauges globally, but how many are in Chile? (and are they the same used in the other products?). Adding a map with the locations of such gauges would also help to interpret the reliability of the products over the different parts of the country.

In addition, it would be useful to know if one of these products (CR2MET?) is considered, or it has been demonstrated in previous studies, to be more reliable or if it is the ‘reference’ product for the Chilean meteorological or hydrological offices (which one is included as meteorological driver into CAMELS-CL?)

As written above, I would suggest adding more information on the comparison of the rainfall fields, that should be carried not only over the mean annual values (Fig. 2), but over the different years (so to differentiate also the “near normal”, and “dry” periods cited when identifying the verification periods in Sect 3.3), over the different seasons, and, especially important, at catchment scale, in order to preserve the consistency among the rainfall forcing and the corresponding streamflow (see, for example, Tarek et al, HESS 2020). For each product, it may be shown, for example the yearly time-series of the box-plots representing the Mean Areal Precipitation for the catchments)

Section 3.3: the fact that the first 10 years (1990-1999) are used both as warm-up period and as verification period is a bit confusing: ten years of warm-up period is not necessary, and in this way there seems to be no warm-up for Verification 1 period? At least one year of warm-up would be needed… This point may be clarified.

pag. 11: as written above, I would remove the AOF objective function.

Section 4.1.1: more information should be given on the performances of the models when not regionalised: first of all distinguishing the performances over the different climatic regions and interpreting the efficiencies in relation to the possible lack of reliability of the rainfall forcing in specific catchments/regions. In addition, looking at the entire boxes and whiskers (and not at the median values only) in Fig. 3, I do not agree that ERA5 is equally (or better) performing than the other products: perhaps more details on local performances may help to better understand.

Section 4.2.2: as above written, if this section is preserved, there should be a similar analysis for the ‘at-site parameterised’, not regionalised simulations (and more information on how the regimes have been identified).

TECHNICAL COMMENTS

l.1: I suggest to clarify in the abstract that what is regionalised are the parameters of a rainfall-runoff model.

Fig. 2: I would use a linear scale for the legenda, since the current colours do not clearly distinguish the regional differences in the rainfall values among the products.

l.323: I do not agree that the spatial proximity and feature similarity results are so close: adding an x-axis grid in Fig. 4 would highlight that Feature similarity performs better, looking at the entire boxes.

ll 340-344: actually, not only RF-MEP but all products (a part from ERA5) are a merging of reanalysis and ground-based observations, from what I understood.

lines 349-354 (and bottom panel of Fig. 5) are more related to the results presented in Section 4.2.1 (and in Fig. 6) and may be moved there?

Section 4.2.1: maybe at least a first comment on the differences among the rainfall products shown by Fig, 6 should be added/moved here.