Rainfall Prediction Using Statistical Downscaling Based on Support Vector Machine in Selangor

Global climate change gains notoriety in literature discussions for potentially triggering extreme change intensity and regularity, like floods and droughts. In this study, the amount of daily rainfall in Selangor was predicted using a downscaling model based on the machine learning technique of the Support Vector Machine (SVM) approach. The collected atmospheric data (predictor) and daily rainfall data (predictand) between 2008 and 2018 used incorporate five imputation methods: mean imputation, K-nearest neighbor, Nonlinear Iterative Partial Least Squares (NIPALS) algorithm, Markov Chain Monte Carlo (MCMC) multiple imputation algorithm, and Expectation Maximization (EM) algorithm. The predictor selection was obtained using Principal Component Analysis (PCA). Primarily, gamma, cost, and epsilon were determined using K-fold cross-validation. Once the parameter value was identified, varying kernel types (linear, RBF, polynomial, and sigmoid) allowed the SVM performance as a regression model to be measured. The SVM model was developed by first handling missing data using imputation methods. The model generating the lowest RMSE value performs best because the difference of the estimated and observed value is minor. PCA efficiently reduced data dimension while retaining key variabilities. The SVM model with a Radial Basis Function (RBF) kernel outperformed others in predicting daily rainfall by displaying the lowest RMSE during calibration (13.95071) and validation (12.60423). The most fitting parameter set for the SVM model is C set to 4.00, y set to 1.935, and e set to 0.2. Based on the study, the SVM model performance is limited when applied to this dataset. For future studies, exploring advanced imputation techniques and broadening the methodology to other tropical climates for broader applicability are recommended.