Calibration/validation time-period selection in hydrologic models leads to uncertainty in water balance simulations

Published: 09-11-2020| Version 4 | DOI: 10.17632/hrdycfbm4m.4
Contributors:
Daniel Myers,
Darren Ficklin,
Scott Robeson,
Ram Neupane,
Alejandra Botero-Acosta,
Pedro Avellaneda

Description

The temporal split-sample approach is the most common method to allocate observed data into calibration and validation groups for hydrologic model calibration. Often, calibration and validation data are split 50:50, where a hydrologic model is calibrated using the first half of the observed data and the second half is used for model validation. However, there is no standard strategy for how to split the data. This may result in different distributions in the observed hydrologic variable (e.g., wetter conditions in one half compared to the other) that could affect simulation results. We investigated this uncertainty by calibrating Soil and Water Assessment Tool hydrologic models with observed streamflow for three watersheds within the United States. We used six temporal data calibration/validation splitting strategies for each watershed (33:67, 50:50, and 67:33 with the calibration period occurring first, then the same three with the validation period occurring first). We found that the choice of split could have a large enough impact to alter conclusions about model performance. Through different calibrations of parameter sets, the choice of data splitting strategy also led to different simulations of streamflow, snowmelt, evapotranspiration, soil water storage, surface runoff, and groundwater flow. The impact of this research is an improved understanding of uncertainties caused by the temporal split-sample approach and the need to carefully consider calibration and validation periods for hydrologic modeling to minimize uncertainties during its use. The file "Research_Data_for_Myers_et_al.zip" includes the water balances and observed data from the study.

Files