Synoptic circulation classification and downscaling for the Midwestern United States

The human impact of climate change and climate variability generally occurs on smaller spatial scales than those currently resolved by General Circulation Models. Hence, impact studies of climate scenarios and regional-scale manifestations of climate variability require downscaling of large-scale climate realizations. In this study, the results of a synoptic circulation classification are subjected to statistical downscaling procedures to simulate daily maximum and minimum surface air temperature, total daily precipitation, and total monthly precipitation at Indianapolis, IN.

Variables used in the synoptic classification are 500 and 700 mb height, 850 mb temperature, and 850 mb dew point temperature, plus a single derived variable: precipitable water content, at seven rawinsonde locations in the Midwestern USA. Varimax orthogonally rotated principal components analysis (PCA) of these variables yields five significant components, which explain 78% and 83% of the variance in the original data for the growing and non-growing seasons (based on variance of surface temperature). The component scores are then clustered to yield synoptic circulation types and are used as input variables for regression models and feed-forward backpropagation artificial neural networks (ANNs) for surface air temperature and precipitation.

In accord with apriori expectations, the study finds that ANNs outperform regression models in terms of prediction accuracy, and the accuracy of the downscaling models for temperature is superior to that for precipitation. In general, the temperature models performed better during the growing season, while the precipitation models performed better during the non-growing season. The accuracy of all downscaling models for surface air temperature is improved by adding an auto-regressive term. Predicted and observed daily maximum and minimum air temperature for the independent model evaluation data sets have RMSE as low as 2.8°C and correlation coefficients as high as 0.87. The downscaling model formation and application is discussed in terms of model performance, comparison of downscaling methodologies, and possible model improvements.