Attribution of Extreme Summer Temperatures in SE Australia

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 6 January 2015: 3:30 PM
124B (Phoenix Convention Center - West and North Buildings)
Michael Richman, CIMMS/Univ. of Oklahoma, Norman, OK; and L. Leslie and K. Flenory

In recent decades, mean monthly maximum temperatures in SE Australia have exceeded historical values with increasing frequency. Climate model projections, such as those contained in the IPCC, suggest that such events will occur more frequently and intensify in future decades. The extreme temperatures have damaged ecosystems, led to increased droughts and fires and resulted in the loss of human life. The present work examines the role of sea surface temperatures (SST) and a number of climate drivers in predicting summer mean monthly maximum temperature at selected locations in SE Australia.

Ninety-one ocean grid boxes of SST surrounding Australia were used to identify simultaneous and lagged relationships. Additionally, 14 climate drivers were considered and their low and high frequency components, derived through wavelet analysis. In total, there were 224 potential predictor attributes. To achieve stable predictions, attribute reduction was a necessity. Through multi-fold cross validated regression and bagging, approximately 90% of the attributes could be removed. The remaining 10% of the attributes served as a reduced set of predictors passed to prediction equations. Linear multiple and nonlinear support vector regression (SVR) methods were applied to predict the summer monthly mean maximum temperature using this reduced set of attributes. For SVR, several types of kernels were evaluated: linear, polynomial and radial basis function. The polynomial degree and radial basis function kernel width were optimized for sea surface temperatures and climate drivers by maximizing their multi-fold cross validated correlations with air temperatures at the various locations in SE Australia.

For the prediction of mean monthly maximum temperatures, the key findings were: (1) both climate drivers and SST had similar contributions to the prediction accuracy and, (2) the combination of the reduced sets of SSTs and climate driver predictors often accounted for 40-60% of the variance. Accounting for such a large percentage of maximum temperature variance has the potential to serve as a useful tool for climate scientists and sectors of the economy that use such climate information.