Statistical downscaling using regression-based technique in Peninsular Malaysia

• Main Author: Kho, Pui Kim
• Format: Thesis
• Language: English
• Published: 2020
• Subjects: QA Mathematics
• Online Access: http://eprints.utm.my/id/eprint/101963/1/KhoPuiKimPFS2020.pdf; http://eprints.utm.my/id/eprint/101963/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:146116

General Circulation Models (GCMs) are important in projecting future climate change. Due to its coarse spatial resolution, downscaling methods are used to obtain local climate information from GCM. This study presents an application of statistical downscaling to assess rainfall changes in Peninsular Malaysia during the months of November to February. Statistical downscaling models are developed using the reanalysis output from the National Center for Environmental Prediction/ National Center for Atmospheric Research (NCEP/NCAR) to test the ability in simulating the daily time series of local rainfall. In the pre-processing step, Principal Component Analysis (PCA) and Self-Organizing Map (SOM) are used to reduce the dimensionality of the dataset. Eight variables are considered from the NCEP/NCAR reanalysis output including sea level pressure (SLP), geopotential height at 500hPa and 850hPa (P500 and P850), relative humidity at 500hPa and 850hPa (R500 and R850), near surface relative humidity (RHUM), near surface specific humidity (SHUM) and mean temperature (TEMP). Potential predictors are selected based on the correlations of NCEP reanalysis with observed rainfall. The predictors are ranked based on the strength of correlations and the model is built with high correlated predictors until the model is optimized. The humidity appears to be the most suitable predictors with the highest correlations to the observed rainfall. Eight models are developed: four with single variable (SLP) and four with combined variables (SHUM + SLP), to form Principal Component Analysis and Regression model (PCA-REG), Principal Component Analysis and Canonical Correlation Analysis Model (PCA-CCA), Self-Organizing Map and Regression model (SOM-REG) and Self-Organizing Map and Canonical correlation Analysis model (SOM-CCA). Results show that the best downscaling model is SOM-REG with combined predictors (SHUM + SLP). The calibration and validation of the best downscaling model determined in this study has shown that the (SOM-REG) model is able to adequately capture the trend of the observed rain series. This model is also capable to project future climate with GCM outputs. The overall results have shown that the future climate is predicted to be having an increasing trend.