Geographically and temporally weighted regression modeling in statistical downscaling modeling for the estimation of monthly rainfall

Abstract

Rainfall estimation was carried out using various Statistical Downscaling (SD) models namely Projection Pursuit, Quantile Regression, Multiple Linear Regression, Partial Least Squares Regression, Clustered Linear Regression and Two-Stage Modeling, and Clusterwise Regression. Global regression cannot handle the relationship between response variables and predictor variables in data containing spatial and temporal variability. The Geographically and Temporally Weighted Regression (GTWR) model can be used to overcome this. This study will perform SD modeling for the estimation of monthly rainfall using the Weighted Least Squares method. The response variables are monthly rainfall data from 35 stations in West Java Province from January 1983 to December 2012 and the predictor variables are temperature and monthly precipitation from the General Circulation Model from the National Centers for Environmental Prediction in the form of a Climate Forecast System Reanalysis model. The results of the study show that the GTWR method, which employs the Exponential kernel function and a fixed bandwidth to model monthly rainfall, outperforms the variable selection method, by giving the value of R² = 70.62% and the Root Mean Square Error (RMSE) = 84.25 while the variable selection method gives the value of R² = 31.21% and RMSE = 128.91. The combination of the GTWR method and Spline interpolation method is the best method for estimating the monthly rainfall value in an unobserved location.