From global to local: assessment of relative uncertainty in multiple climate variable projections

This study investigates the relative contributions of climate model selection, bias correction method, and greenhouse gas emissions scenario to uncertainties in projections of surface air temperature, precipitation, relative humidity, sea-level pressure, Wet Bulb Globe Temperature (WBGT) and labour productivity from global to local scales. Fourteen CMIP5 general circulation models were evaluated against the JRA-55 reanalysis using two bias correction methods: Linear Scaling (LS) and Quantile Mapping (QM). The bias correction coefficients, derived from historical data (1960–2005), were applied to future periods under RCP2.6, RCP4.5, and RCP8.5 scenarios. Additionally, the QuickClim machine learning framework was employed to reconstruct climate fields under a wide range of \((CO_2)\)-equivalent concentration pathways, enabling an extended sensitivity analysis. Results showed that the QM method applied to the ensemble mean outperforms LS in correcting both the mean and temporal variability for most variables, particularly precipitation and relative humidity, while LS yields slightly better results for sea-level pressure. In terms of future projections, a gradual increase in uncertainty was observed, in general, from global to local scales for all considered variables. The application of QuickClim demonstrated that higher CO2-equivalent concentrations are associated with greater uncertainty of surface air temperature. Otherwise, the uncertainty of mean rainfall increased dramatically as the spatial scale decreased, for any CO2-equivalent concentrations. WBGT analysis revealed that the differences between the BC methods gradually became more evident from global to local scales, with QM method presenting higher WBGT and higher labour productivity loss than in LS method at the local scale.