the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 24 Mar 2023
140-year daily ensemble streamflow reconstructions over 661 catchments in France
Abstract. The recent development of the FYRE climate (French Hydroclimate REanalysis), a high-resolution ensemble daily reanalysis of precipitation and temperature covering the period 1871–2012 and the whole of France, offers the opportunity to derive streamflow series over the country from 1871 onwards. The FYRE Climate dataset has been used as input for hydrological modelling over a large sample of 661 near-natural French catchments using the GR6J lumped conceptual model. This approach led to the creation of the 25-member hydrological reconstructions HyDRE spanning the 1871–2012 period. Two sources of uncertainties have been taken into account: (1) the climate uncertainty by using forcings from all 25 ensemble members provided by FYRE Climate, and (2) the streamflow measurement error by perturbing observations used during the calibration. The hydrological model error based on the relative discrepancies between observed and simulated streamflow has been further added to derive the HydREM streamflow reconstructions. These two reconstructions are compared to other hydrological reconstruction with different meteorological inputs, hydrological reconstructions from machine learning algorithm and independent/dependent observations. Overall the results show the added value of the HydRE and HydREM reconstructions in terms of quality, uncertainty estimation, and representation of extremes, therefore allowing to better understand the variability of past hydrology over France.
This preprint has been withdrawn.
-
Withdrawal notice
This preprint has been withdrawn.
-
Preprint
(4027 KB)
Interactive discussion
Status: closed
-
RC1: 'Comment on hess-2023-78', Anonymous Referee #1, 02 May 2023
This paper by Devers et al presents a valuable dataset of reconstructed streamflow across France for the extended period 1871-2012 using the GR6J model. The results show the integrity of the dataset compared to other available datasets in France, as well as the added value of the HydREM deterministic set. The paper presents a novel approach to uncertainty estimation, and is within the scope of the HESS journal. The paper is well structured, and makes good use of graphics.
Please not that I did not follow the intricacies of the maths in sections 3.5.2 to 3.5.4 so I can’t comment on the validity of the approach. I have included grammatical corrections in the attached PDF.
Q1 It is a shame the dataset can’t be extended up to the present for consistency and longevity. Is there a current climate dataset that follows on consistently FYRE that could enable this?
Q2 Does fixing the CemaNeige parameters to the median of the “snowy” catchments for the non-snowy catchments make sense? Should the values not be set to something to indicate there is less snow here? Or should the module not be “switched off”?
Q3 is sampling the observational error randomly a justifiable choice? Would that not affect the variability of the timeseries? Is measurement error not likely to be systematically over and under for periods of time longer than one day? Is there any literature on this?
Q4 you chose 25 associations randomly and then compared them with all 625 for 3 catchments. What tests did you do to show that the differences were not significant. Would bootstrapping not have been a better test?
Q5 looks to me from Fig 2 that you would benefit from breaking Q<1 into more than 1 residual group, but you don’t actually use this?
Q6 Fig 8 it doesn’t look like the reconstructions follow the multidecadal variations well at all pre 1970. I think more discussion is needed on this.
Q7 Fig 10 is interesting, but can an example with reliable observations be found to better demonstrate the validity of HydRE and HydREM? Here you are stating that HydRE and HydREM are “more realistic”, but this is purely subjective based on the available information.
Q8 the reconstructions reviewed here are all using the GR6J model I believe (except GRUN, which is only reviewed in the multi-decadal variability study). Model uncertainty has not really been commented on. A completely unrelated catchment model could produce quite different results. A vast number of studies have shown the large impact of hydrology model uncertainty, which comes in addition to climate input uncertainty, parameter uncertainty and observation uncertainty. This should definitely be discussed in section 5.2 at the very least. Unless this is somewhat accounted for in the error model maths that I did not fully understand…
-
RC2: 'Comment on hess-2023-78', Anonymous Referee #2, 04 May 2023
In the manuscript entitled “140-year daily ensemble streamflow reconstructions over 661 catchments in France”, Devers and colleagues present and evaluate two new set of 25-member hydrological reconstructions over the 1871-2012 period. The added value of these new long-term reconstructions is that three types of uncertainties are taken into account: (i) the climate uncertainty, (ii) the streamflow measurement error and (iii) the hydrological model error. Overall, I found that the manuscript is clear and well written. The evaluation shows that these new reconstructions are an interesting tool to better understand the variability of the French river flows, as well as the hydrological extreme events, and the related uncertainties. I found this study of interest for the reader of this journal, as it presents an innovative method and new hydrological reconstructions. I do, however, have some comments that I hope are constructive. The minor and ortographical corrections are attached in pdf.
l.145-157: Does this procedure applied to take into account the observational uncertainties induced a decrease in the quality of the calibration results ? I also wonder how it should affect the observed variability ? As instead of the “best value” of observations, we choose random values in the estimated PDF of the observation error.
l.159 How did you choose 3 years for the warm-up period ?
l.165 If I understand well, the distribution of the subsample of 25 simulations is quite representative to the one of the 625 simulations ? How did you test this similarity in the distribution ?
Section 3.5: The error model gives you the uncertainty related to the choice of parameter within the GR6J model. However, it gives not provide you the uncertainties related to the choice of the hydrological model, which I will assume could be quite important. It would be worth it to clarify that point in the discussion section.
l.200 How often this case happen ? If it is too often, it could suggest that there is an issue regarding the estimation of the observational error.
l.209-211: Did you applied a statistical test to estimate if the distribution of the 25 simulations randomly choose are representative of the distribution of the 2500 members ? Is 25 simulations enough ?
Figure 3. Do you have an idea why the three ensembles overall fails to represent the daily streamflow at Ubaye@Barcelonnette in the late spring / beginning summer ? Is this something related to the snow melting ? Is this visible on the seasonal cycle over the reference period or is this only related to this specific year ?
l.255-257: Indeed HydRE and HydREM ensemble seem more correlated than the SCOPE Hydro ensemble, but if you talk about correlations it would be nice to have the values. Maybe some boxplots on the right side of the Figure 3 could add a nice complement to the figure and allows a direct comparison of the correlations between the reconstructions over the three catchments. Same for the analysis of Figure 4.
l.266: I suppose the period if more 1973-2006, which is your calibration period ?
l.298-300: Why you don’t take all the long-term stations available over this period ? The Loire@Monjean is not used for example (it is not indicated with a triangle on Figure 1).
l.310-315: As this dry bias is present on HydRE and HydREM but not SCOPE Hydro, which used the same hydrological model, do you think it comes from the FYRE Climate reanalysis ? Or do you think it comes from the method to derive the uncertainties in HydRE and HydREM ?
Section 4.4: Large discrepancies can be found between the observed multidecadal variability and the multidecadal variability from the 3 reconstructions. It would be nice to add some discussions about it, and to bring some potential explanations.
In Figure 8, the river flow observations of the Seine at Poses are rather short to evaluate the multidecadal variations over this catchment, although we can already see some discrepancies before 1970. In Bonnet et al. (2021), they use a long-term observational dataset of the Seine at Paris Austerlitz, would it be possible to include such long-term data in order to improve the evaluation ?
Section 4.5: It is very interesting and needed to evaluate the capacity of the reconstructions to reproduce extreme events, but I found that a comparison with direct observations is missing in order to estimate how the reconstructions are able to reproduce the intensity of high and low flows for example. As a dry bias seems to be present in the mean flow of the reconstructions, the high and low flows can also be affected.
Section 5.1: For the calibration, the period is not very long in comparison to the multidecadal variability that is observed in some river flows and that can influence high and low flows (Boé et Habets, 2014, Bonnet et al. 2021). Your Figure 8 is a good example for that, regarding the Seine@Poses for example, which is overall in a wet multidecadal phases over the calibration period. So I think this section would need a bit more discussion and caution about that point. Additionally, you mentioned in this section 2 examples of methods to quantify the sensitivity of the calibration to the period, have you tried to apply one of them ? If not, why ?
l.441-443: The reconstructions developed in Bonnet et al. 2017, 2020 are also available on a lot of small catchments, as they use the Isba land surface coupled to the Modcou hydrogeological model (Habets et al. 2008), including the stations that you used for the assessment of the multidecadal variability.
l.446: It would be great to put forward some of the potential insights about the strengths and weaknesses of the FYRE Climate reanalysis that your evaluation bring forward.
References:
Bonnet, Rémy, Julien Boé, and Florence Habets. "Influence of multidecadal variability on high and low flows: the case of the Seine basin." Hydrology and Earth System Sciences 24.4 (2020): 1611-1631.
Boé, J., and Florence Habets. "Multi-decadal river flow variations in France." Hydrology and Earth System Sciences 18.2 (2014): 691-708.
Habets, F., Boone, A., Champeaux, J. L., Etchevers, P., Franchisteguy, L., Leblois, E., ... & Viennot, P. (2008). The SAFRAN‐ISBA‐MODCOU hydrometeorological model applied over France. Journal of Geophysical Research: Atmospheres, 113(D6).
Interactive discussion
Status: closed
-
RC1: 'Comment on hess-2023-78', Anonymous Referee #1, 02 May 2023
This paper by Devers et al presents a valuable dataset of reconstructed streamflow across France for the extended period 1871-2012 using the GR6J model. The results show the integrity of the dataset compared to other available datasets in France, as well as the added value of the HydREM deterministic set. The paper presents a novel approach to uncertainty estimation, and is within the scope of the HESS journal. The paper is well structured, and makes good use of graphics.
Please not that I did not follow the intricacies of the maths in sections 3.5.2 to 3.5.4 so I can’t comment on the validity of the approach. I have included grammatical corrections in the attached PDF.
Q1 It is a shame the dataset can’t be extended up to the present for consistency and longevity. Is there a current climate dataset that follows on consistently FYRE that could enable this?
Q2 Does fixing the CemaNeige parameters to the median of the “snowy” catchments for the non-snowy catchments make sense? Should the values not be set to something to indicate there is less snow here? Or should the module not be “switched off”?
Q3 is sampling the observational error randomly a justifiable choice? Would that not affect the variability of the timeseries? Is measurement error not likely to be systematically over and under for periods of time longer than one day? Is there any literature on this?
Q4 you chose 25 associations randomly and then compared them with all 625 for 3 catchments. What tests did you do to show that the differences were not significant. Would bootstrapping not have been a better test?
Q5 looks to me from Fig 2 that you would benefit from breaking Q<1 into more than 1 residual group, but you don’t actually use this?
Q6 Fig 8 it doesn’t look like the reconstructions follow the multidecadal variations well at all pre 1970. I think more discussion is needed on this.
Q7 Fig 10 is interesting, but can an example with reliable observations be found to better demonstrate the validity of HydRE and HydREM? Here you are stating that HydRE and HydREM are “more realistic”, but this is purely subjective based on the available information.
Q8 the reconstructions reviewed here are all using the GR6J model I believe (except GRUN, which is only reviewed in the multi-decadal variability study). Model uncertainty has not really been commented on. A completely unrelated catchment model could produce quite different results. A vast number of studies have shown the large impact of hydrology model uncertainty, which comes in addition to climate input uncertainty, parameter uncertainty and observation uncertainty. This should definitely be discussed in section 5.2 at the very least. Unless this is somewhat accounted for in the error model maths that I did not fully understand…
-
RC2: 'Comment on hess-2023-78', Anonymous Referee #2, 04 May 2023
In the manuscript entitled “140-year daily ensemble streamflow reconstructions over 661 catchments in France”, Devers and colleagues present and evaluate two new set of 25-member hydrological reconstructions over the 1871-2012 period. The added value of these new long-term reconstructions is that three types of uncertainties are taken into account: (i) the climate uncertainty, (ii) the streamflow measurement error and (iii) the hydrological model error. Overall, I found that the manuscript is clear and well written. The evaluation shows that these new reconstructions are an interesting tool to better understand the variability of the French river flows, as well as the hydrological extreme events, and the related uncertainties. I found this study of interest for the reader of this journal, as it presents an innovative method and new hydrological reconstructions. I do, however, have some comments that I hope are constructive. The minor and ortographical corrections are attached in pdf.
l.145-157: Does this procedure applied to take into account the observational uncertainties induced a decrease in the quality of the calibration results ? I also wonder how it should affect the observed variability ? As instead of the “best value” of observations, we choose random values in the estimated PDF of the observation error.
l.159 How did you choose 3 years for the warm-up period ?
l.165 If I understand well, the distribution of the subsample of 25 simulations is quite representative to the one of the 625 simulations ? How did you test this similarity in the distribution ?
Section 3.5: The error model gives you the uncertainty related to the choice of parameter within the GR6J model. However, it gives not provide you the uncertainties related to the choice of the hydrological model, which I will assume could be quite important. It would be worth it to clarify that point in the discussion section.
l.200 How often this case happen ? If it is too often, it could suggest that there is an issue regarding the estimation of the observational error.
l.209-211: Did you applied a statistical test to estimate if the distribution of the 25 simulations randomly choose are representative of the distribution of the 2500 members ? Is 25 simulations enough ?
Figure 3. Do you have an idea why the three ensembles overall fails to represent the daily streamflow at Ubaye@Barcelonnette in the late spring / beginning summer ? Is this something related to the snow melting ? Is this visible on the seasonal cycle over the reference period or is this only related to this specific year ?
l.255-257: Indeed HydRE and HydREM ensemble seem more correlated than the SCOPE Hydro ensemble, but if you talk about correlations it would be nice to have the values. Maybe some boxplots on the right side of the Figure 3 could add a nice complement to the figure and allows a direct comparison of the correlations between the reconstructions over the three catchments. Same for the analysis of Figure 4.
l.266: I suppose the period if more 1973-2006, which is your calibration period ?
l.298-300: Why you don’t take all the long-term stations available over this period ? The Loire@Monjean is not used for example (it is not indicated with a triangle on Figure 1).
l.310-315: As this dry bias is present on HydRE and HydREM but not SCOPE Hydro, which used the same hydrological model, do you think it comes from the FYRE Climate reanalysis ? Or do you think it comes from the method to derive the uncertainties in HydRE and HydREM ?
Section 4.4: Large discrepancies can be found between the observed multidecadal variability and the multidecadal variability from the 3 reconstructions. It would be nice to add some discussions about it, and to bring some potential explanations.
In Figure 8, the river flow observations of the Seine at Poses are rather short to evaluate the multidecadal variations over this catchment, although we can already see some discrepancies before 1970. In Bonnet et al. (2021), they use a long-term observational dataset of the Seine at Paris Austerlitz, would it be possible to include such long-term data in order to improve the evaluation ?
Section 4.5: It is very interesting and needed to evaluate the capacity of the reconstructions to reproduce extreme events, but I found that a comparison with direct observations is missing in order to estimate how the reconstructions are able to reproduce the intensity of high and low flows for example. As a dry bias seems to be present in the mean flow of the reconstructions, the high and low flows can also be affected.
Section 5.1: For the calibration, the period is not very long in comparison to the multidecadal variability that is observed in some river flows and that can influence high and low flows (Boé et Habets, 2014, Bonnet et al. 2021). Your Figure 8 is a good example for that, regarding the Seine@Poses for example, which is overall in a wet multidecadal phases over the calibration period. So I think this section would need a bit more discussion and caution about that point. Additionally, you mentioned in this section 2 examples of methods to quantify the sensitivity of the calibration to the period, have you tried to apply one of them ? If not, why ?
l.441-443: The reconstructions developed in Bonnet et al. 2017, 2020 are also available on a lot of small catchments, as they use the Isba land surface coupled to the Modcou hydrogeological model (Habets et al. 2008), including the stations that you used for the assessment of the multidecadal variability.
l.446: It would be great to put forward some of the potential insights about the strengths and weaknesses of the FYRE Climate reanalysis that your evaluation bring forward.
References:
Bonnet, Rémy, Julien Boé, and Florence Habets. "Influence of multidecadal variability on high and low flows: the case of the Seine basin." Hydrology and Earth System Sciences 24.4 (2020): 1611-1631.
Boé, J., and Florence Habets. "Multi-decadal river flow variations in France." Hydrology and Earth System Sciences 18.2 (2014): 691-708.
Habets, F., Boone, A., Champeaux, J. L., Etchevers, P., Franchisteguy, L., Leblois, E., ... & Viennot, P. (2008). The SAFRAN‐ISBA‐MODCOU hydrometeorological model applied over France. Journal of Geophysical Research: Atmospheres, 113(D6).
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 531 | 207 | 37 | 775 | 25 | 37 |
- HTML: 531
- PDF: 207
- XML: 37
- Total: 775
- BibTeX: 25
- EndNote: 37
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1