Machine learning-based integration of large-scale climate drivers can improve the forecast of seasonal rainfall probability in Australia

Publication Type:
Journal Article
Environmental Research Letters, 2020, 15, (8)
Issue Date:
Full metadata record
© 2020 The Author(s). Published by IOP Publishing Ltd. Probabilistic seasonal rainfall forecasting is of great importance for stakeholders such as farmers and policymakers to assist in developing risk management strategies and to inform decisions. In practice, there are two kinds of commonly used tools, dynamical models and statistical models, to provide probabilistic seasonal rainfall forecasts. Dynamical models are based on physical processes but are usually expensive to operate and implement, and rely overly on initial conditions. Statistical models are easy to implement but are usually based on simple or linear relationships between observed variables. Recently, machine learning techniques have been widely used in climate projection and perform well in reproducing historical climate. For these reasons, we conducted a case study in Australia by developing a machine learning-based probabilistic seasonal rainfall forecasting model using multiple large-scale climate indices from the Pacific, Indian and Southern Oceans. Rainfall probabilities of exceeding the climatological median for upcoming seasons from 2011 to 2018 were successively forecasted using multiple climate indices of precedent six months. The performance of the model was evaluated by comparing it with an officially used forecasting model, the SOI (Southern Oscillation Index) phase model (SP) operated by Queensland government in Australia. Results indicated that the random forest (RF) model outperformed the SP model in terms of both distinct forecasts and forecasting accuracy. The RF model increased the percentages of distinct forecasts to 64.9% for spring, to 71.5% for summer, to 65.8% for autumn, and to 63.9% for winter, 1.4 ∼ 3.2 times of the values from the SP model. Forecasting accuracy was also greatly increased by 28%, 167%, 219%, and 76% for four seasons respectively, compared to the SP model. The proposed rainfall forecasting model is based on readily available data, and we believe it can be easily extended to other regions to provide seasonal rainfall outlooks.
Please use this identifier to cite or link to this item: