Stationary-based Cross Validation Support Vector Regression for Drought Prediction in the Colorado River Basin

Recent above-normal temperatures, which exacerbated the impacts of precipitation deficits, are recognized as the primary driver of droughts in the Colorado River Basin (CRB). Moreover, climate change/variability has caused non-stationarity in temperature time series and changes in spatiotemporal patterns of droughts. The aim of this research is to improve the performances of drought prediction models by capturing these structural changes in non-stationary timeseries in a novel stationary-based cross validation support vector regression (SCV-SVR) model. For tuning hyperparameters, a k-fold cross validation based on the assumption of stationarity in time series was developed. To divide the non-stationary time series into k stationary folds, a change point detection technique was used. Among a number of combinations of hyperparameters that go through the cross-validation, the value with the lowest error is selected and set for the testing period of the SCV-SVR model. Three types of drought indices in monthly, seasonal, and semi-annual time scales including multivariate, bivariate standardized drought indices, and univariate standardized drought indices were used as the target variables. The SCV-SVR and SVR models were tested and compared in CRB. To setup the models, 1979-2010 and 2011-2016 data from the North American Land Data Assimilation System (NLDAS) were respectively used for training and testing of the SCV-SVR and SVR models. The results indicated that SCV-SVR outperformed SVR since variable hyperparameters (regularization constant and tube size) were used in SCV-SVR to deal with structural changes in the data. The SCV-SVR can predict drought more accurately than traditional SVR in a warming climate.