| dc.description.abstract | Rainfall as a part of the highest climate fluctuation and characterize the most dominant climate in Indonesia is strongly influenced by global climate change, such as extreme rainfall. Extreme rainfall can cause flood and various disadvantages like crop failure in agriculture. The analysis that examines the extreme events is needed to minimize the bad impact due to the extreme rainfall events. Extreme rainfall can be analyzed such as using statistical downscaling (SD). SD is the process of transforming information from large-scale (global) as an explanatory variables to small-scale (local) as a respon variable. In this research, monthly raifall data from Indramayu district used as respon variable and GCM output data used as explanatory variables. The precipitation of GCM output is high dimension and there are multicolinear between adjacent grids. To reduce the high dimension and solve the multicolinearity problem usually uses principal component analysis (PCA), functional principal component (FPCA) and partial least squares (PLS). PCA and FPCA focus on variety in the explanatory variables, whereas PLS focuses on the variety between the explanatory variables and the response variable. SD model requires a strong correlation between the precipitation of GCM output and the rainfall to obtain a more accurate estimate. Strong correlation delivers the same pattern of the two variables. A method to estimate the extreme rainfall in SD modeling is quantile regression, which is an expansion from median regression on various quantile value. The model that has been formed in quantile regression can be used to measure the effect of explanatory variables at the centre, right or left tail of the data distribution. The aim of this research is to estimate the extreme rainfall using quantile regression and PLS as a dimension reduction method with time lag precipitation of GCM output. The monthly rainfall of Indramayu and the precipitation of GCM output from 1979 to 2008. Data from 1979 to 2007 for developing the model and data 2008 for model validation. Dimension reduction using PLS produces one component i.e the first score of and the first score of . Then, the extreme was estimated by linier, quadratic and cubic quantile regresssion with score as explanatory variable and rainfall data as respon variable at the 75th, 90th and 95th quantiles. Model validation and consistency were implemented at the last step of this research. The results show that the rainfall estimate using linier, quadratic and cubic quantile regression model were not following the pattern of actual rainfall. The model was modified by adding dummy variables. Using dummy variables resulted better estimate. The best model is cubic quantile regression with dummy variables. The pattern of extreme rainfall was similar to the pattern actual rainfall. In February, the actual rainfall is about 439.33 mm and well predicted by the 95 th quantile. The cubic quantile regression with dummy variables can also result the consistent prediction for next two years . | id |
