Downscaling temperature and precipitation: A comparison of regression-based methods and artificial neural networks

A comparison of two statistical downscaling methods for daily maximum and m inimum surface air temperature, total daily precipitation and total monthly precipitation at Indianapolis, IN, USA, is presented. The analysis is cond ucted for two seasons, the growing season and the non-growing season, defin ed based on variability of surface air temperature. The predictors used in the downscaling are indices of the synoptic scale circulation derived from rotated principal components analysis (PCA) and cluster analysis of variabl es extracted from an 18-year record from seven rawinsonde stations in the M idwest region of the United States. PCA yielded seven significant component s for the growing season and five significant components for the non-growin g season. These PCs explained 86% and 83% of the original rawinsonde data f or the growing and non-growing seasons, respectively. Cluster analysis of t he PC scores using the average linkage method resulted in eight growing sea son synoptic types and twelve non-growing synoptic types. The downscaling o f temperature and precipitation is conducted using PC scores and cluster fr equencies in regression models and artificial neural networks (ANNs). Regression models and ANNs yielded similar results, but the data for each r egression model violated at least one of the assumptions of regression anal ysis. As expected, the accuracy of the downscaling models for temperature w as superior to that for precipitation. The accuracy of all temperature mode ls was improved by adding an autoregressive term, which also changed the re lative importance of the dominant anomaly patterns as manifest in the PC sc ores. Application of the transfer functions to model daily maximum and mini mum temperature data from an independent time series resulted in correlatio n coefficients of 0.34-0.89. In accord with previous studies, the precipita tion models exhibited lesser predictive capabilities. The correlation coeff icient for predicted versus observed daily precipitation totals was less th an 0.5 for both seasons, while that for monthly total precipitation was bel ow 0.65. The downscaling techniques are discussed in terms of model perform ance, comparison of techniques and possible model improvements. Copyright ( C) 2001 Royal Meteorological Society.