Rainfall Prediction in Semi-Arid Regions of Australia using Gradient Boosted Trees

Rainfall prediction in semi-arid regions remains a challenging yet critical task for effective water resource management, agricultural planning, and climate risk mitigation. This study applies three state-of-the-art gradient boosted tree algorithms—XGBoost, LightGBM, and CatBoost—to the Australian WeatherAUS dataset, which contains ۱۴۵,۴۶۰ daily observations recorded from ۲۰۰۷ to ۲۰۱۷. After extensive preprocessing, including imputation of missing values, encoding of categorical features, and temporal partitioning of the dataset into training and testing subsets, the models were systematically evaluated using both threshold-independent metrics (ROC-AUC, PR-AUC) and threshold-dependent measures (precision, recall, and F۱-score). The results demonstrate that all three algorithms achieved strong predictive skill, with CatBoost outperforming XGBoost and LightGBM by a small but consistent margin (ROC-AUC = ۰.۸۸۸, PR-AUC = ۰.۷۴۵). Feature importance analysis further revealed that atmospheric humidity, pressure, and wind dynamics are the dominant drivers of rainfall occurrence, while temperature and spatial variability provide complementary signals. The probabilistic forecasts generated by the models not only deliver accurate predictions but also provide actionable insights for managing water resources in semi-arid Australia. These findings confirm the utility of ensemble machine learning methods in addressing rainfall variability and strengthen their potential role in climate adaptation strategies.

Keywords:

Rainfall prediction , Gradient boosted trees , XGBoost , CatBoost