Coupling a Markov Chain and Support Vector Machine for At-Site Downscaling of Daily Precipitation

Abstract

Abstract Statistical downscaling is useful for managing scale and resolution problems in outputs from global climate models (GCMs) for climate change impact studies. To improve downscaling of precipitation occurrence, this study proposes a revised regression-based statistical downscaling method that couples a support vector classifier (SVC) and first-order two-state Markov chain to generate the occurrence and a support vector regression (SVR) to simulate the amount. The proposed method is compared to the Statistical Downscaling Model (SDSM) for reproducing the temporal and quantitative distribution of observed precipitation using 10 meteorological indicators. Two types of calibration and validation methods were compared. The first method used sequential split sampling of calibration and validation periods, while the second used odd years for calibration and even years for validation. The proposed coupled approach outperformed the other methods in downscaling daily precipitation in all study periods using both calibration methods. Using odd years for calibration and even years for validation can reduce the influence of possible climate change–induced nonstationary data series. The study shows that it is necessary to combine different types of precipitation state classifiers with a method of regression or distribution to improve the performance of traditional statistical downscaling. These methods were applied to simulate future precipitation change from 2031 to 2100 with the CMIP5 climate variables. The results indicated increasing tendencies in both mean and maximum future precipitation predicted using all the downscaling methods evaluated. However, the proposed method is an at-site statistical downscaling method, and therefore this method will need to be modified for extension into a multisite domain.