Does high-resolution HadGEM3 overestimate precipitation over land?

Abstract:

Previous studies showed that high-resolution global climate models (GCMs) overestimate land precipitation when compared against observation-based data. In particular, grid point models (e.g. HadGEM3) show a significant increase of precipitation on mountain regions, where the scarcity of gauge stations increase the uncertainty of gridded observations and reanalysis. The goal of this work is to assess the effect of such differences in precipitation on river discharge, an integrator of the water balance at catchment scale.

We diagnosed the river flow by routing the runoff generated by CMIP6-HighResMIP HadGEM3 simulations at low and high resolution (~1° and ~0.25° respectively). The river flow is evaluated using a set of 344 monitored catchments distributed around the world. We also infer the global discharge by constraining the simulations with observations following a novel approach that involves: a) a bias correction in monitored rivers with two methods (linear scaling and cumulative density function mapping), and b) an extension of the correction to the river mouth, and along the coast, considering the catchment’s orography as a key parameter to fill gaps in non-monitored lands.

Our global discharge estimate is 47.4±1.6×10³km³yr^-1, which is closer to the original high-resolution estimate (50.5×10³km³yr^-1) than to the low-resolution (39.6×10³km³yr^-1). At high resolution, (1) precipitation occurs in the correct place leading to a more realistic spatial distribution of runoff (e.g. Southeast Asia, Alps, Alaska Range in Fig. 1), and (2) the orographic precipitation rises reducing the dry bias of runoff observed at coarse resolution probably due to a better resolved coastline and orography (e.g. Maritime Continent, Southern Andes in in Fig. 1). However, high-resolution simulations slightly increase the wet bias in catchments dominated by flat terrain (e.g. Congo and La Plata). The improvement of model parameterizations and tuning may reduce the remaining errors in future HR simulations.