Can Integrated Machine Learning Models Improve Rainfall Forecasting? A Case Study from the Australian Tropics

This study investigates the integration of wavelet preprocessing with machine learning models to enhance wet-period rainfall prediction in the Northern Territory of Australia. Lagged large-scale climate indices, including the El Niño Southern Oscillation (ENSO), Dipole Mode Index (DMI), Madden-Julian Oscillation (MJO), and Interdecadal Pacific Oscillation (IPO), were employed as predictors for Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) models, both with and without wavelet transformation. Results indicate that hybrid models consistently outperformed their standalone counterparts, with the Wavelet-LSTM (W-LSTM) using Daubechies (db) wavelet emerging as the most effective configuration. Across three stations (Warruwi, Waterloo, and Avon Downs), the W-LSTM (db) achieved high correlations with observed rainfall (0.86–0.95) and substantially reduced RMSE, by up to 89.6% at Warruwi. It also demonstrated a strong ability to capture rainfall peaks and troughs, with correlations ranging from 0.76 to 0.91. Performance metrics further confirmed its superiority, with R² values of 0.80–0.89 and NSE values of 0.78–0.87, compared to the relatively lower performance of individual models (R² ranges 0.74–0.81). Although some challenges remain in accurately predicting extreme rainfall events, the overall findings present the advantages of combining wavelet preprocessing with machine learning. The W-LSTM (db) model provides a more robust and reliable approach to rainfall forecasting by improving the representation of variability and extremes. These results indicate the potential of hybrid wavelet-machine learning frameworks for enhancing predictive capability in regions influenced by high climate variability, with broader applicability to other tropical and monsoonal environments. The study develops ML based hybrid models for wet-period rainfall forecasting across multiple stations in Australia’s tropics using lagged climate indices as predictors. Wavelet preprocessing is integrated with LSTM and SVM models to enhance prediction accuracy by capturing multi-scale rainfall patterns. The W-LSTM model with Daubechies wavelet achieves the highest correlations (0.86–0.95) and reduces RMSE by up to 89.6%, outperforming standalone models. Model performance is rigorously evaluated using RMSE, R², NSE, Taylor diagrams, and violin plots, confirming the robustness of the proposed hybrid approach. The findings offer practical applications for water resource management and flood forecasting in tropical regions with highly variable rainfall patterns.