Statistical prediction of 20th century European summer temperatures based on ERA20c reanalysis data

Abstract

An alternative to dynamical seasonal prediction of European climate is statistical modeling. Statistical modeling is an appealing and computationally effective approach for producing seasonal forecasts by exploiting the physical connections between the predictand variable and the predictors. We assess the seasonal predictability of summer European 2m temperature (T2m) using canonical correlation analysis. Seasonal means of spring Soil Moisture (SM), Sea Level Pressure (SLP) and Sea Surface Temperature (SST) are used as predictors of mean summer T2m. For SSTs, we test the potential predictability of T2m using three different regions. These regions include what we define as: Extratropical North Atlantic (ENA), Tropical North Atlantic (TNA), and North Atlantic (NA). The predictability is explored in the ERA20c reanalysis and in comprehensive Earth System Model (ESM) fields. The results are provided for the European domain on a horizontal grid of 1°x1° degrees. In order to identify the local T2m predictability related to the different predictor variables, we first built Univariate Linear Regression models, one for every predictor. The regression models are calibrated and validated during 1902-1950 and a prediction is provided for the periods 1951-1998, 1951-2004, and 1951-2008, respectively. The resulting correlation maps between the original and the predicted T2m anomalies showed that for the predictor variables SLP, SM, and SSTENA the results of the experiments using ESM data share similar T2m predictability patterns with the results of the experiments using reanalysis data. Most prominent disagreements between the predictability patterns resulting from ESMs and from ERA20c refers to the T2m prediction that utilizes tropical SSTs. SM is identified as the most important predictor for the summer European temperature predictability. The ERA20c data show that the SM predictor field can be used for the T2m prediction over most of our study region west of 15° E and that the ENA SSTs can be used for the prediction over Europe east of 15° E. The resulting gridded correlation coefficients vary between 0.3 and 0.5. These results are not sensitive to the prediction period and to the number of Canonical Coefficients used in the regression model. Our approach complements existing numerical seasonal forecast frameworks and can be implemented for ensemble prediction studies.