Exhaustive Searching and LASSO for Reliable Drought Forecasting over South Korea

The spring drought over South Korea has been extensive damage recent years and its forecasting can be important in water management and agricultural industries. However, the drought forecasting is not an easy task because of the difficulty to find predictors to the precipitation predictand. Also, limited hydrological records for applying to complex models such as nonlinear or deep learning models do not produce reliable forecasting results. In the current study, we proposed the drought forecasting approach by exhaustive searching for explanatory variables and a regression model for limited record lengths. At first, the target drought index was set with the accumulated spring precipitation (ASP) obtained by the median of the 93 available weather stations over South Korea. Then, exhaustive searching for predictors was performed with association between the ASP and the differences of two pair combination of the global winter MSLP, say Df4m, for the time lag of the spring seasonal drought. The 37 Df4m predictors were found with high correlation over 0.55. The detected 37 variables were categorized into three subregions. The predictors in the same region contain highly similar to each other. Subsequently, the multicollinearity problem cannot be avoidable. To solve the multicollinearity problem, the Least Absolute Shrinkage and Selection Operator (LASSO) model was applied resulting five Df4m predictors and the good agreement of the forecasting value with the observed value as R²=0.72. Therefore, we concluded that the proposed LASSO model with the exhaustive searching of the global MSLP can be a good alternative to forecast the spring drought over South Korea. The spring drought forecasting with the LASSO model and the Df4m predictors can be extensively used for water managers and water industry.