Ensemble stationary-based support vector regression for drought prediction under changing climate

Non-stationarity due to climate change and/or variability can reduce the capabilities of drought prediction models. The objective of this study is to improve drought prediction by eliminating non-stationarity from temperature time series, a key factor in development and propagation of droughts in a changing climate. In order to relax the assumption of stationarity, an ensemble stationary-based support vector regression (ESSVR) method was developed and compared with the traditional support vector regression (SVR). Three types of drought indices in three time scales (monthly, seasonal, and semiannual), including multivariate, bivariate standardized drought indices, and univariate standardized drought indices were used as the target variables. In an application to the Red River of the North Basin (RRB), the North American Land Data Assimilation System (NLDAS) data from 1979 to 2016 were used for the training and testing of the prediction model. The Pearson correlation, mot mean square error (RMSE), and Taylor diagram were used to evaluate the performances of the ESSVR. Remarkably, the distribution of identified change points varies by climate divisions. The results of the SVR and ESSVR in the RRB were compared, demonstrating the better performances of the ESSVR for most of the drought indices, particularly those with higher sensitivity to temperature. It was found that the extreme (high and low) values of hyperparameters mostly assigned by SVR cause a higher risk of overfitting for SVR. In contrast, ESSVR improves the drought prediction by removing the non-stationarity, thus providing more accurate drought predictions, especially for a warming climate.