Projection of climate change impacts on precipitation using soft-computing techniques: A case study in Zayandeh-rud Basin, Iran

• Three statistical transformation function-based methods are used to correct biases in large-scale atmospheric attributes.
• A kernel version of supervised PCA is presented to reduce the impact of the high dimensionality in atmospheric predictors.
• The Bayesian learning algorithm shows superiority in addressing the nonlinear relationships in the downscaling modeling.

Due to the complexity of climate-related processes, accurate projection of the future behavior of hydro-climate variables is one of the main challenges in climate change impact assessment studies. In regression-based statistical downscaling processes, there are different sources of uncertainty arising from high-dimensionality of atmospheric predictors, nonlinearity of empirical and quantitative models, and the biases exist in climate model simulations. To reduce the influence of these sources of uncertainty, the current study presents a comprehensive methodology to improve projection of precipitation in the Zayandeh-Rud basin in Iran as an illustrative study. To reduce dimensionality of atmospheric predictors and capture nonlinearity between the target variable and predictors in each station, a supervised-PCA method is combined with two soft-computing machine-learning methods, Support Vector Regression (SVR) and Relevance Vector Machine (RVM). Three statistical transformation methods are also employed to correct biases in atmospheric large-scale predictors. The developed models are then employed on outputs of the Coupled Model Intercomparison Project Phase 5 (CMIP5) multimodal dataset to project future behavior of precipitation under three climate changes scenarios. The results indicate reduction of precipitation in the majority of the sites in this basin threatening the availability of surface water resources in future decades.