Combing Random Forest and Least Square Support Vector Regression for Improving Extreme Rainfall Downscaling

Abstract

A statistical downscaling approach for improving extreme rainfall simulation was proposed to predict the daily rainfalls at Shih-Men Reservoir catchment in northern Taiwan. The structure of the proposed downscaling approach is composed of two parts: the rainfall-state classification and the regression for rainfall-amount prediction. Predictors of classification and regression methods were selected from the large-scale climate variables of the NCEP reanalysis data based on statistical tests. The data during 1964–1999 and 2000–2013 were used for calibration and validation, respectively. Three classification methods, including linear discriminant analysis (LDA), random forest (RF), and support vector classification (SVC), were adopted for rainfall-state classification and their performances were compared. After rainfall-state classification, the least square support vector regression (LS-SVR) was used for rainfall-amount prediction for different rainfall states. Two rainfall states (i.e., dry day and wet day) and three rainfall states (dry day, non-extreme-rainfall day, and extreme-rainfall day) were defined and compared for judging their downscaling performances. The results show that RF outperforms LDA and SVC for rainfall-state classification. Using RF for three-rainfall-states classification and LS-SVR for rainfall-amount prediction can improve the extreme rainfall downscaling.