Modelling seasonal rainfall forecasts forced with improved predictive ocean surface temperature

Abstract

Seasonal rainfall forecasts are in high demand for users such as irrigators and water managers in decision making and risk management. Both statistical and dynamical models are widely used to generate probabilistic rainfall forecasts in advance for a season. Statistical prediction systems establish a stationary relationship between the predictor and the predictand variables, for example, sea surface temperature anomaly (SSTA) patterns over the Indian and Pacific Oceans are found to provide probabilistic seasonal rainfall forecast throughout Australia. On the other hand, dynamical models are based on the laws of physics and thus they can capture non-linear interactions of the atmosphere, land and ocean. In the case of seasonal forecasts, these models use a two-tiered process by predicting global Sea Surface Temperatures (SSTs) first; an atmospheric general circulation model (GCM) is subsequently forced by the pre-forecast SST to make a seasonal prediction. Improvement in predicted SST is therefore significant for issuing better concurrent seasonal rainfall forecasts. Consequently the motivation of this research is to improve seasonal SST forecasting on a global scale and accordingly provide concurrent seasonal rainfall prediction by applying statistical techniques and similarly forcing a climate model. For the purpose of improving seasonal SST forecasting, this thesis presents a multimodel combination approach which considers intermodel dependency between the multiple participating GCMs. The algorithm provides globally gridded SSTA for a season ahead based on the degree of correlation between the forecast errors and the relative size of each model s error variance. This methodology demonstrates an attractive way of improving seasonal SSTA forecasts over the majority of grid cells in the globe compared to the recent multimodel approach wherein the correlations are ignored. The standard practice of multimodel approach by pooling the models over a common time period can, however, cause loss of information if some models have a longer period of data. Hence this thesis also presents another simple approach of combining models which have variable data lengths. In the second part of this thesis, concurrent seasonal rainfall forecasts are issued from the improved SST, and single model SST predictions separately. The statistical techniques Bayesian Joint Probability (BJP) and Bayesian Model Averaging (BMA) are applied to translate seasonal rainfall forecasts using six predicted SSTA indices over the Indian and Pacific Oceans. The BJP-BMA approach shows encouraging results derived from improved SSTA indices, although no upper atmosphere predictor variable is considered. In addition, the global climate model ACCESS is used to issue concurrent seasonal rainfall prediction on a global scale from the improved forecast SST. The results indicate that there is merit in formulating global seasonal rainfall forecasts from the predictive uncertainty reduced SST, rather than relying on a single model predicted SST.