Ensemble Learning for Rainfall Extremes: Threshold-Based Forecasting in the Cauvery Delta Zone

Abstract

This study investigates rainfall and temperature extremes across six agricultural blocks in the Cauvery Delta Zone (CDZ), Tiruchirappalli district, Tamil Nadu, using daily meteorological data from 1981 to 2023. A comprehensive set of climate indices—including Consecutive Dry Days (CDD), Consecutive Wet Days (CWD), Total Precipitation (PRCPTOT), Simple Daily Intensity Index (SDII), and percentile-based metrics such as R10mm, R20mm, RX1day, RX3day, RX5day, R95p, and R99p, was applied to assess interannual variability and identify anomalous rainfall patterns. Notably, 2008 and 2011 exhibited high-intensity rainfall events, while 2002 and 2023 were marked by extended dry spells and reduced precipitation. To anticipate these extremes, five machine learning models, XGBoost, Random Forest, Support Vector Regression (SVR), Backpropagation Neural Network (BPNN), and Linear Regression—were evaluated for predictive accuracy. XGBoost consistently outperformed other models, yielding the lowest Root Mean Square Error (RMSE) and highest coefficient of determination (R²) across all indices. Its precision in classifying rainfall thresholds underscores its utility for agro-climatic forecasting. These findings support the integration of ensemble and neural-based methods into early warning systems and adaptive agricultural planning in semi-arid environments.