Drought Forecasting with ML-based Regionalized Climate Indices

Abstract

Drought forecasting in South Korea has become imperative due to the increased frequency of occurrence leading various damages such as property loss and casualties. Precipitation in South Korea is distributed with high deviation, substantially concentrated in summer. Other seasons have a comparatively low amount of precipitation resulting unbalanced water resources of each season. To overcome the skewed seasonal precipitation, numerous dams and reservoirs have been constructed and operated. The management of those water-related structures should be carried out carefully to meet seasonal requests of water resources, and the precipitation prediction for each season has become critical. However, the seasonal precipitation forecasting has been a challenging task due to complex weather systems and climate patterns. The current study proposes a novel procedure for forecasting seasonal precipitation as: (1) regionalization of climate variables; (2) extraction of features with PCA, ICA and Autoencoder; and (3) finally regression model applications. Two globally gridded climate variables, Mean Sea Level Pressure (MSLP) and Sea Surface Temperature (SST) were teleconnected with the Accumulated Seasonal Precipitation (ASP) of South Korea. The results indicate that the k-means clustering successfully regionalizes the highly correlated climate variables with the ASP and all three feature selection algorithms, PCA, ICA, and Autoencoder present their superiority in different seasons combining GLM and SVM models. Especially, the PCA performs better with the linear GLM model and the Autoencoder shows better performance with the nonlinear SVM model. Overall, it can be concluded that the proposed seasonal precipitation forecasting procedure combining ML-based algorithms can be a good alternative.