Comprehensive Evaluation of Machine Learning Techniques for Hydrological Drought Forecasting

Drought is among the most hazardous climatic disasters that significantly influence various aspects of the environment and human life. Qualitative and reliable drought forecasting is important worldwide for effective planning and decision-making in disaster-prone regions. Data-driven models have been extensively used for drought forecasting, but due to the inadequacy of information on model performance, the selection of an appropriate forecasting model remains a challenge. This study concerns a comparative analysis of six machine learning (ML) techniques widely used for hydrological drought forecasting. The standardized runoff index (SRI) was calculated at a seasonal (3-month) time scale for the period 1973 to 2016 in four selected watersheds of the Han River basin in South Korea. The ML models employed were built-ins, using precipitation, temperature, and humidity as input variables and the SRI as the target variable. The results indicated that all the ML models were able to map the relationship for seasonal SRI using the applied input vectors. The decision tree (DT) technique outperformed in all the watersheds with an average mean absolute error $(\mathrm{MAE})=0.26$, root mean square error $(\mathrm{RMSE})=0.34$, Nash-Sutcliffe efficiency $(\mathrm{NSE})=0.87$, and coefficient of determination $(R^2)=0.89$. The performances of the support vector machine (SVM) and deep learning neural network (DLNN) were similar, whereas the fuzzy rule-based system (FRBS) performed very well compared to the artificial neural network (ANN) and the adaptive neuro-fuzzy inference system (ANFIS). The overall findings of this study indicate that, considering performance criteria and computation time, the DT was the most accurate ML technique for hydrological drought forecasting in the Han River basin.