Bayesian learning and relevance vector machines approach for downscaling of monthly precipitation

Abstract

In this study, statistical downscaling of large-scale general circulation model (GCM) simulations to monthly precipitation of Kemer Dam, in Turkey, has been performed through relevance vector machines (RVMs). All possible regression methods along with statistical measures have been used to select potential predictors through reanalysis data providing air850, hgt850, and prate variables as the optimal. The determined explanatory variables are then used for training RVM-based statistical downscaling model. A least-squares support vector machine (LSSVM)-based downscaling model is also constructed to compare the downscaling performance of RVM through some performance evaluation measures such as R-2, AdjR(2) and RMS error (RMSE). Because RVM is able to obtain the better modeling accuracy in terms of all performance measures during the testing period, third-generation coupled climate model (CGCM3) simulations run through the trained RVM to obtain future scenario results. The effectiveness of the RVM model is illustrated through its integration to climate scenarios (20C3M and A2). The statistical significance of the probable changes obtained with used methods is examined by Mann-Whitney U (M-W) and t-tests considering scenario forecasts. According to pessimistic A2 scenario results, statistically significant decreasing trends are foreseen for both seasonal and annual precipitation in the study basin.