Forecasting climate variables using a mixed-effect state-space model

This paper demonstrates the potential advantage of using a linear, mixed-effect state-space model for statistical downscaling of climate variables compared to the frequently used approach of linear regression. This comparison leads to the development of a method for estimation of model parameters using the EM algorithm approach. The model is applied to the prediction of temperature and rainfall statistics at both a sub-tropical and temperate location in Australia. The results indicate that for lead times of 1-10 years this state-space approach is able to predict observed seasonal temperature and rainfall means with substantially greater precision than climatology, multivariate linear regression (MLR) or a standard linear state-space (LSS) approach. The model is seen as a first step in the development of a short-term climate change projection system that will utilise both historical climate data as well as dynamically derived mean climate change projection information obtained from global climate models (GCMs).