Statistical downscaling of GCM using kernel support vector regression for rainfall prediction in Bireuen district

Abstract

Global climate phenomena, including El-Nino and Southern Oscillation, Indian Ocean Dipole and Madden Julian Oscillation, have affected rainfall behavior in Indonesia. This study attempts to model a statistical downscaling of the General Circulation Model (GCM), which is usually used as learning data for the prediction of climate change, using Support Vector Regression (SVR) to predict rainfall during the dry season in Bireuen, Aceh Province. Data consists of observational rainfall data at 7 (seven) weather stations and 10 models of 7x7grid-scale hindcasts of GCM data collected within 24 years (1990-2013). The RBF kernel was found to produce the best performance (correlation value: 0.828; RMSE: 24.035) compared to the linear (correlation value: 0.538; RMSE: 27.207) and polynomial kernel (correlation value: 0.639; RMSE: 25.584). Among GCM downscaling output, CMC1-CanCM3 was considered to be the best model for rainfall prediction in Bireun District, particularly in Peudada and Gandapura regions. This model is recommended to be used as a rainfall forecasting tool during drought and would help in optimizing the arrangement of cropping strategy in Bireun area to prevent possible crop failure.