SVM-PGSL coupled approach for statistical downscaling to predict rainfall from GCM output

Hydrological impacts of climate change are assessed by downscaling the General Circulation Model (GCM) outputs of predictor variables to local or regional scale hydrologic variables (predictand). Support Vector Machine (SVM) is a machine learning technique which is capable of capturing highly nonlinear relationship between predictor and predictand and thus performs better than conventional linear regression in transfer function-based downscaling modeling. SVM has certain parameters the values of which need to be fixed appropriately for controlling undertraining and overtraining. In this study, an optimization model is proposed to estimate the values of these parameters. As the optimization model, for selection of parameters, contains SVM as one of its constraints, analytical solution techniques are difficult to use in solving it. Probabilistic Global Search Algorithm (PGSL), a probabilistic search technique, is used to compute the optimum parameters of SVM. With these optimum parameters, training of SVM is performed for statistical downscaling. The obtained relationship between large-scale atmospheric variables and local-scale hydrologic variables (e. g., rainfall) is used to compute the hydrologic scenarios for multiple GCMs. The uncertainty resulting from the use of multiple GCMs is further modeled with a modified reliability ensemble averaging method. The proposed methodology is demonstrated with the prediction of monsoon rainfall of Assam and Meghalaya meteorological subdivision of northeastern India. The results obtained from the proposed model are compared with earlier developed SVM-based downscaling models, and improved performance is observed.