Can hydrologists improve model performance by calibrating their models to discontinuous data?
Abstract
Hydrological modeling requires data splitting (i.e., split-sample test, SST) for model calibration and validation. Modelers tend to utilize a continuous calibration sub-period (CSP) followed by validation in their model building despite our recent work showing these approaches are inferior to calibrating to all available data (full-period CSP which skips validation). However, using a discontinuous calibration sub-period (DCSP), such as odd/even years, provides an alternative approach to compare to the full-period CSP.
This large-sample SST assessment study empirically assesses how different DCSP decisions influence hydrological model performance in post-validation model testing periods. We propose a comprehensive DCSP experimental design that consists of over 50 different DCSP decisions when splitting a data period to calibration and validation sub-periods. Three different DCSP splits are applied, which are 75/25 (i.e., 75% for calibration and 25% for validation), 67/33, and 50/50 as well as a full-period CSP which skips model validation entirely. Typically, the data record is split into three sub-periods: calibration, validation and testing. Model testing periods always follow calibration/validation periods and are independent periods for comparing performance of validated models under different DCSP decisions The lumped conceptual hydrological model HMETS is applied for testing this experiment in 463 catchments from the CAMELS dataset across the contiguous United States. Results show that using the full-period record for calibration is the most robust choice considering model testing period performance. It is observed that more calibration data in DCSP splitting results in better model testing performance: (1) When model is built in a given year, a DCSP decision with more data split into calibration yields better performance; and (2) A given DCSP decision perform better when it is applied in a scenario with more data available. However, it is found that the full-period CSP (skipping validation) significantly outperforms other DCSP decisions in 25% catchments on average over all model testing periods while the DCSP decisions rarely significantly outperform the full-period CSP. Such a proportion even exceeds 50% when only a short period of data is available.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H15F..03S