The statistical downscaling methods’ uncertainty share as a measure for adopted strategies in downscaling studies for climate change

To study climate change we rely on global climate models (GCMs) but their resolution is coarse to investigate impacts at the local scale. Hence, a downscaling task is required for the use of these coarse-resolution outputs. In this sense, statistical downscaling methods (SDMs) are commonly applied to analyse the local impacts. Furthermore, a quantification of the uncertainty share of the SDMs is advised to complement the results. However, many choices need to be done before their application and these decisions can bias the outcome of the analysis. This work examines the SDMs’ uncertainty share to evaluate to what extent the different adopted strategies can impact the climate change signal (CCS) associated with the study. For this, eleven research indicators (six representing precipitation extremes) are used with four future scenarios, 28 state-of-art GCMs, and 15 SDMs of two different types (change factor and quantile mapping methods). The uncertainty involved is quantified by the variance decomposition procedure. Three different decisions are tested:

(i) The selection of the Coupled Model Intercomparison Project (CMIP) era. The uncertainty shares in phases five and six (CMIP5 and CMIP6, respectively) are compared.  

(ii) The selection of the SDM ensemble based on the SDMs’ methodological construction. More specifically, based on an ensemble of five methods of change factor type (including an event-based change factor weather generator) and an ensemble of ten methods of quantile mapping.

(iii) The selection of the optimal SDM ensemble number. Different unique SDMs combinations are tested from k-ensemble members in [2,n] with n as the ensemble with the largest number of members (n=15).

To complement the analysis, the outcomes of the CCSs from all the combinations in (ii) and (iii) are analysed. The results showed that the uncertainty quantification of the SDMs is not sensitive to the selection of the CMIP era. However, this choice is important if the focus is on the GCMs and future scenarios. Hence, it is preferable (but not mandatory) to perform the analysis with the most recent era. The selection of the SDMs based on a methodological construction might bias the conclusions. Therefore, it is better to include methods from all possible types since the results showed that the more methods included in the downscaling, the more reliable the estimation of the SDMs’ uncertainty share. The CCS seems to strongly depend on the choice of the SDM ensemble, and it tends to converge from different k-ensemble members in [2,n] towards the largest ensemble (n). Hence, CCSs from large SDM ensembles will be more reliable. Future work must extend the analysis into different climatological regions and include more methods from all the possible types.