Performance Assessment of ANN and SVR for downscaling of daily rainfall in dry regions

Studies of climate change impacts on water resources need to conversion of projection of climate variable pattern from coarser scales to a suitable scale. Downscaling processes improves projection of General Circulation Models (GCMs) significantly. In this study is assessed performance of Artificial Neural Network and Support Vector Regression. Observation rainfall was collected for 1961-1990 from Birjand synoptic station. Also value of 26 predictors from CanESM2 output in Assessment Report Five (AR5) was extracted. Predictor selection was performed by Stepwise regression. Model skill was evaluated using indices R2, RMSE and NSE. Also for better analysis using of various tests such as uncertainty assessment, reproduction of descriptive statistics, dry and wet spells and monthly rainfall. Results of assessment indices showed that estimation performances of both methods in daily rainfall are relatively suitable. Value of R2, RMSE and NSE are achieved 0.48, 1.5 mm and 0.47 for SVR with Polynomial kernel function in best case respectively. Results of skill model in estimation of descriptive statistics indicated that SVR with Polynomial kernel function outperforming others. Comparison of observed and downscaled monthly rainfall illustrates that SVR performs better than ANN in winter season. Also both methods have overestimation in summer months. Results of identification of dry days sowed that performance of both models is well and same. In this case ANN identified 96% dry days correctly. Results of identification of wet days showed that SVR outperforms ANN. Assessment of skill method for estimation of dry spells length indicated that both methods have more efficiently in long spells than short