Can direct CMIP6 model simulations reproduce mean annual historical streamflow change?

Direct streamflow outputs from CMIP6 models are important for studying future water supply under climate change. However, the hydrological community generally ignores such data as its ability to predict streamflow change is largely unquantified and presumably considered biased. Here, we examined the ability of the 15 CMIP6 models to directly reproduce the observed mean annual streamflow change (Delta Q) between 1999-2014 and 1982-1998 for 115 large basins across the globe. The Delta Q and error of Delta Q in CMIP6 models (Delta Q_error) were then attributed to nine independent climate and vegetation variables using a stepwise regression model. Compared to the observations, almost all of the CMIP6 models overestimate Delta Q. In variable attribution, the observed Delta Q in the 115 basins can be well attributed (R2 = 0.59) to changes in precipitation (66 %), temperature (15 %), and leaf area index (-18 %). However, these variables could not explain Delta Q obtained from CMIP6 models (R2 = 0.13). In error attribution, errors in precipitation change contributes the most (69 +/- 12 %) to Delta Q_error in CMIP6 models, followed by location-related variables, changes in vegetation-related variables, and changes in temperature-related variables. Although the current performance of CMIP6 model outputs varies in simulating observed streamflow change, they may become more reliable with improved model structures for better representations of climate and land surface processes.

Automated summary

Direct streamflow outputs from CMIP6 models are important for studying future water supply under climate change. However, the accuracy of these data is still debated, and the relevant variables cannot fully explain the observed streamflow change.