Capability of Artificial Neural Networks for predicting long-term seasonal rainfalls in east Australia

Rainfall in southeast Australia is known to be affected by large scale climate modes variability. This study focused on investigating the use of lagged El Nino Southern Oscillation (ENSO) as potential predictors of spring rainfall in Victoria and Queensland in east Australia. Six rainfall stations including Bruthen, Buchan and Orbost in Victoria and Barcaldine, Kalamia and Augathella in Queensland were chosen as case study. Artificial Neural Network (ANN) approach was used as a nonlinear technique to capture this complex relationship. The Pearson correlation coefficients of past values of ENSO with spring rainfalls were calculated; it was discovered that the three months of June, July and August of Nino3.4, have significant correlation with spring rainfall. These correlations are very weak for Victoria and relatively higher for Queensland. These lag months of ENSO were incorporated into ANN models; i.e. the set of Nino3.4 (Jun-July- Aug) was used as inputs for developing ANN models for the stations in Victoria and Queensland. Multilayer Perceptron (MLP) architecture was chosen for this purpose. The models were trained based on Levenberg- Marquardt algorithm. ANN models showed higher correlation for Queensland compared to Victoria indicating that ANN is more capable of finding the pattern and trend of the observations in Queensland. After calibrating and validating the models, in order to evaluate the generalization ability of the developed ANN models, out-of-sample tests were carried out. It was discovered that ANN models are showing very poor generalization ability for east Victoria regarding finding the pattern of the series (r = -0.97, 0.23 and - 0.67 for Bruthen, Buchan and Orbost respectively) compared to Queensland with correlation coefficients of 0.74, 0.100 and 0.98 for Barcaldine, Kalamia and Augathella respectively. This study shows the ability of ANN in finding nonlinear relationships between complex large scale climate models and rainfalls in southeast Australia