Kernel And Tree-Based Time Series Forecasting Of Monthly Rainfall Using Large-Scale Climate Indices

Abstract

Rainfall prediction is vital for various economic activities, including agriculture, sustainable water management, and disaster risk reduction. The objective of the present research is to develop machine learning (ML) models for monthly rainfall forecasting using a combination of large-scale climate indices and local climatic variables. Predictors included the Indian Ocean Dipole (IOD), El Nino Southern Oscillation (ENSO), Pacific Decadal Oscillation (PDO), Atlantic Multidecadal Oscillation (AMO), North Atlantic Oscillation (NAO), and local maximum and minimum temperatures, while rainfall was as the target variable. Four ML techniques: Multivariate Adaptive Regression Splines (MARS), Support Vector Machine (SVM), Gaussian Process Regression (GPR), and Random Forest (RF) were used to analyse and predict rainfall in Farukh Nagar station in the province of Haryana, India. The rainfall data consisted of 468 monthly observations split into 80% training and 20% testing subsets. To capture nonlinear rainfall climate linkages, Radial Basis Function (RBF) and Pearson VII kernel functions were used in SVM and GPR. Among the models, SVM with the RBF kernel outperformed others, achieving correlation coefficient of 0.811, mean absolute error of 20.44 mm, root mean square error of 49.19 mm, and scattering index of 1.15 during training. The corresponding statistics during testing were 0.718, 24.45 mm, 46.86 mm, and 1.19. The Taylor diagram further confirmed the superiority of SVM-RBF in representing rainfall variability compared with RF, GPR, and MARS. The integration of global climate indices with local temperature variables enhanced model robustness and improved the rainfall forecasting performance. The results highlight the significance of advanced ML approaches, particularly SVM-RBF, in improved rainfall forecasting leading to better agricultural planning and climate-resilient decision-making.