Abstract
Several studies investigated the effect of large-scale climatic modes on rainfall in tropical climatic zone of Australia. In the current study, machine learning models such as artificial neural networks (ANN) and random forest (RF) were used to forecast the wet-period rainfall (Austral summer: December-February) at six different stations of Northern Territory (NT), Australia. To examine the synchronous influence of potential predictors on wet-period rainfall of NT, multiple input sets with different combinations of lagged Indian Ocean Dipole (IOD), Southern Oscillation Index (SOI), Interdecadal Pacific Oscillation (IPO), and Madden Julian Oscillation (MJO) were investigated. To assess the superiority of machine learning models over traditional linear regression (MR) model, the linear models were also developed for the same study locations. The model performance was evaluated using five distinct statistical metrics, including the root mean square error (RMSE), coefficient of determination (R 2 ), relative root mean square error (rRMSE), relative mean bias (rBias), and Nash-Sutcliffe coefficient (NSC). Large-scale climate factors, primarily MJO, SOI, and Niño3.4, were found to have a considerable effect on NT wet-period rainfall, which can be used for future rainfall forecasting. It is found that among the studied models, the ANN model delivered the best results with the least RMSE ranging from 0.47 to 1.72, along with coefficient of determination values ranging from 0.84 to 0.91. Whereas, for the RF and MR models, produced results were having RMSE values ranging 0.83 ~ 2.28 and 2.52 ~ 4.19 respectively, and R 2 values ranging 0.79 ~ 0.86 and 0.62 ~ 0.72 respectively.