Extending calibration sequences to include multi-year drought is not sufficient to train models for future drier climate

Hydrological models’ parameters do not transfer well between periods with different climatic conditions; particularly when parameters calibrated on a wetter climate are used to project streamflow during a drier period. Many Mediterranean and semi-arid regions worldwide experienced multi-annual, persistent drought conditions in recent decades, which exposed these limitations of hydrological models and their most common calibration methods. Additionally, projections indicating a likely warmer and drier future in these regions render this issue particularly disquieting, especially as models tend to overestimate water availability during dry periods.

In south-eastern Australia, the Millennium drought (ca. 1997­–2009) called attention to this issue due to its effects on the hydrological behaviour of many catchments. Specifically, models calibrated on pre-drought conditions routinely overestimate streamflow when forced with drought data. However, in operational simulation, it has been theorised that models calibrated on long timeseries that include the Millennium drought will be able to perform well under a future drier climate.

We tested this hypothesis by studying the performance of five conceptual rainfall-runoff models in 155 catchments in the southern Australian state of Victoria. We calibrated the models on different periods, some including the drought and post-drought themselves, and compare the performance across parameter sets during and after the drought. We observe that naively extending calibration sequences to include the drought did not significantly improve model performance across these catchments, neither during the drought itself nor, in most cases, in the decade after the drought. Further adding the post-drought period also provided very limited improvement in this period and is unlikely to make any operational difference, despite drought and post-drought making up at least 30% of the calibration sequence.

These results highlight the importance of the choice of data and method used for calibrating hydrological models under transient climate and suggest that more sophisticated calibration methods are needed in this context. These include calibration with explicit and distinct treatment of different climate regimes (e.g. using multi-objective optimisers). However, where changes in climate triggers drastic shifts in catchment behaviour, novel and more adaptable model structures may be necessary to confidently project streamflow under future climate.