A Comparative Analysis of Downscaled Multi-model Decadal Climate Predictions over the Iberian Peninsula

Decadal climate predictions are a source of climate information to anticipate the evolution of the climate system from 1 to 10 years ahead. Whereas both climate projections and decadal predictions contain information about external forcings, their main difference is that decadal predictions also include information on the phase of the internal climate variability. To achieve this, decadal climate models are initialised once per year with observation-based initial conditions. 

 

This study assesses the effectiveness of various statistical downscaling methods applied to multi-model decadal predictions of mean near-surface air temperature and precipitation for the forecast years 1-5 over the Iberian Peninsula. The multi-model ensemble combines predictions from 13 forecast systems contributing to the Decadal Climate Prediction Project (DCPP) component of the Coupled Model Intercomparison Project Phase 6 (CMIP6). The performance of the different downscaling methods is determined by comparing their forecast quality against the raw, coarse-resolution predictions using four deterministic or probabilistic metrics: the Anomaly Correlation Coefficient (ACC), Root Mean Square error Skill Score (RMSSS), Ranked Probability Skill Score (RPSS) and Continuous Ranked Probability Skill Score (CRPSS). The downscaling and forecast quality assessment are carried out using the high-resolution ERA5Land reanalysis as the reference dataset, and it is performed in leave-one-out cross-validation mode in order to emulate real-time conditions and not to overestimate the actual skill.

 

Three kinds of downscaling methods have been examined. The first type is  based on calibrating the interpolated raw predictions (i.e. correcting biases in the mean value or variance, among others). The second involves building linear regressions using different predictors: (i) large-scale decadal indices as (e.g. the Atlantic Multi-decadal Variability, AMV; or the North Atlantic Oscillation, NAO) (ii) interpolated model data (basic linear regression) or (iii) a combination of the 9 nearest neighbours of model data. Finally, the third approach involves the search for past analog days in the high-resolution reference dataset.

 

The results show that the skill estimates primarily depend on the calibration or linear regression approaches, with small differences from the interpolation method used during the downscaling. While the first type of methods maintains the spatial distribution of the skill compared to the raw predictions, the second and third types can change it. For temperature, the raw predictions show high skill, which is maintained after applying calibration, basic linear regressions or 9 nearest neighbours linear regression. However, the skill is reduced after calculating the linear regressions with external predictors. For precipitation, the skill of the raw predictions is rather low, and the calibration methods do not generally increase such skill. On the other hand, the linear regression method using the AMV index as a predictor is the one that shows the most improvement in skill compared to the raw predictions in some regions.