Metrics that Matter: Calibration Choices and Their Impact on Signature Representation in Conceptual Hydrological Models

Hydrologists are generally aware that the choice of calibration metric will affect how their model reproduces catchment runoff. However, scientific literature mainly provides a theoretical or case study specific discussion of the topic and no general guidelines. This study thus aims to develop a broader picture by evaluating the influence of 8 different objective functions on the representation of 15 hydrologic signatures for 45 lumped conceptual models in 10 climatically diverse catchments.

The 10 selected catchments are a subset of the CARAVAN dataset (Kratzert et al. [2023]) chosen by using a k-means clustering algorithm based on climate indices (Willmott and Feddema [1992]). The 45 models are taken from the MARRMoT toolbox (Knoben et al. [2019], Trotter et al. [2022]) and only models performing over a specified benchmark are used for the analysis. The signatures that will be analysed represent different processes and aspects of the hydrological regime and the following 8 calibration metrics are investigated: KGE, NSE, log KGE, log NSE, NP-KGE (Pool et al. [2018], Split KGE (Fowler et al. [2018], SHE (Kiraz et al. [2023], DE (Schwemmle et al. [2021]).

Preliminary results show that the ability to reproduce specific signatures is clearly influenced by the chosen metric and therefore this choice should always be based on the specific goal of the prospective modelling study. Each metric has specific strengths and weaknesses that may be used to make a decision. However, the results vary based on climate conditions, the applied model structure and the investigated signature. It is therefore difficult to disentangle all interdependencies and develop more general guidelines with the limited catchment set used in this study. We speculate that very dominant processes shaping the general runoff generation in a catchment (such as snow melt) reduces the impact of the choice of calibration metric, and that more complex models typically are more consistent in process representation.

References:
Fowler et al. (2018): doi: 10.1029/2017WR022466
Gupta et al. (2009): doi: 10.1016/j.jhydrol.2009.08.003
Kiraz et al. (2023): doi: 10.1029/2023WR035321
Knoben et al. (2019): doi: 10.5194/gmd-12-2463-2019
Kratzert et al. (2023): doi: 10.1038/s41597-023-01975-w
Nash and Sutcliffe (1970): doi: 10.1016/0022-1694(70)90255-6
Pool et al. (2018): doi: 10.1080/02626667.2018.1552002
Schwemmle et al. (2021): doi: 10.5194/hess-25-2187-2021
Trotter et al. (2022): doi: 10.5194/gmd-15-6359-2022
Willmott and Feddema (1992): doi: 10.1111/j.0033-0124.1992.00084.x

Disclaimer: The first author conducted the presented research at TUD, now a PhD student at UofC