EGU2020-18909
https://doi.org/10.5194/egusphere-egu2020-18909
EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Can a hydrological model be efficient and robust at the same time?

Paul Royer-Gaspard, Vazken Andréassian, and Guillaume Thirel
Paul Royer-Gaspard et al.
  • Université Paris-Saclay, INRAE, UR HYCAR, 92761, Antony, France (paul.royer-gaspard@inrae.fr)

It has been shown in various experiments that many conceptual rainfall-runoff models experience difficulties to simulate annual or longer-term variations of the streamflow (e.g. Coron et al., 2014). Whether this problem is inherent to the structure of the model in question or could be solved by a change of the calibration procedure is still a matter of debate: for example, the work of Coron (2013) tended to show that no parameter set able to solve the issue can be found, while Fowler et al. (2018) argued that such parameter sets exist, and should be identifiable by a change of objective function.

The aim of this study is to explore further the existence of such a parameter set in the case of the GR4J model (Perrin et al., 2003). Parameters sets were in particular tested against their ability to provide efficient (i.e. with good performance) and robust (i.e. transposable in time) discharge simulations over three flow ranges (low, mean and high flows). To this purpose, a large number of parameters sets of GR4J were sampled in 545 French and Australian catchments. The obtained performances were confronted to those obtained with automatic calibration with a range of objective functions focusing on diverse streamflow ranges.

Because of our large catchment set, we were able to identify a variety of cases: catchments for which highly robust parameter sets exist, catchments for which relatively robust parameter sets exist, and catchments for which no robust parameter sets can be found. Compared to the best sampled parameters sets, those derived through automatic calibration often yielded poorer performances regarding at the same time efficiency and robustness of the discharge simulations over the three flow ranges. We discuss the link between model failures and catchments characteristics, as well as the ability of the GR4J model to adequately simulate streamflow on different timescales and flow regimes.

How to cite: Royer-Gaspard, P., Andréassian, V., and Thirel, G.: Can a hydrological model be efficient and robust at the same time?, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-18909, https://doi.org/10.5194/egusphere-egu2020-18909, 2020

Display materials

Display file

Comments on the display material

AC: Author Comment | CC: Community Comment | Report abuse

Display material version 1 – uploaded on 01 May 2020
  • AC1: Answer to Flora Branger, Paul Royer-Gaspard, 04 May 2020

    @ Flora Branger "Can you say more about your database for meteorological forcings?"
    We used SAFRAN reanalysis (Vidal et al., 2010) for daily precipitation and temperature observation, and computed potential evaporation with the Oudin model (Oudin et al., 2005). Do not hesitate if you have further questions.

  • CC1: split-KGE for low vs. high flows, Michael Stoelzle, 04 May 2020

    Regarding your comment in the chat: 

    Paul Royer-Gaspard, INRAE, author (14:29)

    @ Micheal Stoelzle: Yes, but not for this work. We tested Fowler SKGE on French and Australien cathments and realized it gave more robust simulations of low flows but seemingly at the expense of higher flows

    Have you published your work with split-KGE? Could you comment a bit more on the trade-off between robust low flows and robust high flows with SKGE?

    Thanks, Michael

     

    • AC4: Answer to Michael Stoelzle, Paul Royer-Gaspard, 04 May 2020

      @Michael
      Nothing has been published about this for the moment but I had a small discussion about this with Keirnan Fowler. I think that SKGE tends to favor dryer years because KGE usually reaches lower values in these conditions due to its structure (dividing by average flows for bias computation, particularly). It is also quite sensitive to years with high data errors in peak flows.

      • AC5: Answer to Michael Stoelzle (suite), Paul Royer-Gaspard, 04 May 2020

        I would really like to know what you think about this ? Have you also tested SKGE ? In what kind of contexts ? What kind of results did you obtain ? Thanks a lot for answer.

        • CC2: Reply to AC5, Michael Stoelzle, 04 May 2020

          Hi Paul,

          Thanks for your answer. Yes, we used SKGE in a modelling framework with recharge stress tests to quantify the sensitivity of different catchments to changed pre-conditions before drought events. More details here:

          We used a combination of 50% MARE (Mean absolute relative error) and 50% logKGE as the objective function and calculated both metrics split-wise, i.e. average from all years. For MARE we used a weighting to focus on low flow periods during calibration. For SKGE there is now weighting as the literature suggests that SKGE increases performance during dry years. 

          Fowler, K., Peel, M., Western, A., & Zhang, L. (2018). Improved rainfall‐runoff calibration for drying climate: Choice of objective function. Water Resources Research54(5), 3392-3408. (Figure 5)

          • CC3: Reply to CC2, Michael Stoelzle, 04 May 2020

            Sorry the URL in the post was removed somehow... See our EGU2020 contribution: D112, Session HS2.4.5  

             

          • AC7: Reply to CC2, Paul Royer-Gaspard, 04 May 2020

            Many thanks for you answer Michael.

            So if I understand well, you used MARE and logKGE for parameter calibration, split-wise, and with stronger weights on dryer years. I actually also used MARE in my work on Split KGE calibration, to evaluate model performance on other metrics than the one used in calibration. I think it was known to match expert judgement to evaluate model performance on low-flows [1]. Since logKGE also seems to focus on low flows, I would be interested to know why you used both to calibrate your model.

            Guillaume Thirel, one of the co-author of the work I presented, also co-published a paper to warn against the use of log-transform with KGE [2]. It may lead to really strange results if average observed flows match specific conditions. I think it might interest you if you do not already know about it.

            I will definitely take a look at your display. We might maybe continue the discussion on your abstract page.

            Paul

             

            [1] Crochemore, L., Perrin, C., Andréassian, V., Ehret, U., Seibert, S. P., Grimaldi, S., ... & Paturel, J. E. (2015). Comparing expert judgement and numerical criteria for hydrograph evaluation. Hydrological Sciences Journal, 60(3), 402-423.
            [2] Santos, L., Thirel, G., & Perrin, C. (2018). Pitfalls in using log-transformed flows within the KGE criterion.
            • CC4: logKGE, Michael Stoelzle, 04 May 2020

              We used both metrics as MARE focuses on volumetric errors but timing in baseflow simulation is also important. Yes, we are aware of the logKGE paper but as we adjusted the long-term model input (here: recharge) to match the long-term model output (here: baseflow) the water balance should be an issue during calibration.

  • AC2: Comment on EGU2020-18909, Paul Royer-Gaspard, 04 May 2020

    @Luis Samaniego "How do you define robutness within the framework of uncertainty of HM?"
    I hope that I understand well your question. Robustness is here referred to as climatic robustness, i.e. the possibility for the same model (with the same parameter set) to correctly simulate
    heterogeneous climatic conditions in a catchment. We think that adressing the question of parameter transferability should be accompanied by a verification that such parameter sets exist (e.g. Fowler et al., 2016)

  • AC3: Answer to Rafael Pimentel, Paul Royer-Gaspard, 04 May 2020

    @Rafael "Thanks Paul, is there any connection between your conclusions, performance wise, and the location of the selected catchments across France?"

    We found a link between model versatility, i.e. possibility to simultaneously provide robust and unbiased simulations of low, mid and high flows, and some catchments characteritics. In particular,  higher seasonality in streamflow regime and stronger interannual variations of runoff ratio were significantly are significantly associated to lower versatility and worse performances. This is in agreement with other studies on RR model robustness

  • AC6: Answer to Luis Samaniego, Paul Royer-Gaspard, 04 May 2020

    @Luis Samaniego "How do you define robutness within the framework of uncertainty of HM?"

    I hope that I understand well your question. Robustness is here referred to as climatic robustness, i.e. the possibility for the same model (with the same parameter set) to correctly simulate heterogeneous climatic conditions in a catchment. We think that adressing the question of parameter transferability should be accompanied by a verification that such parameter sets exist (e.g. Fowler et al., 2016)