Machine learning algorithms to forecast wet-period rainfall using climate indices in Northern Territory of Australia

Abstract

Accurate rainfall prediction is crucial for understanding and managing a region's social and agricultural environment. As a key indicator of climate change, natural disasters, and local geography, rainfall data empowers us to make informed decisions for various beneficial purposes. Machine learning offers powerful tools for improving rainfall prediction accuracy and estimation capabilities. This study examines how multiple climate indices simultaneously influence wet-period rainfall patterns at two Northern Territory (NT) stations, Hermannsburg and Undoolya. We selected two machine learning models, Random Forest (RF) for its robustness and Long Short-Term Memory (LSTM) for its ability to capture temporal patterns, to investigate these relationships. For this purpose, a variety of input sets, including lagged Indian Ocean Dipole (IOD), El Nino Southern Oscillation (ENSO), and Madden Julian Oscillation (MJO), were proposed and utilized to calibrate and validate, RF and LSTM Models. Our analysis revealed that large-scale climate factors like IOD, Nino 3.4, and MJO significantly influence wet-period rainfall predictions of the NT. Furthermore, the LSTM model outperformed the RF model to predict the wet-period rainfall at the selected stations. For instance, the LSTM achieved higher R² i.e., 0.86 and lower values for both RMSE (ranging from 0.63 to 0.72) and MAE (ranging from 0.43 to 0.64) during the testing phase, indicating a closer fit between predicted and actual wet-period rainfall values.