Machine learning approach for identifying and forecasting streamflow droughts in data limited basins of South Korea using threshold levels

Artificial Intelligence (AI) has been extensively utilized for streamflow prediction, primarily in gauged watersheds using meteorological and historical streamflow data. However, its application in data-limited regions requires innovative approaches due to the reliance on extensive monitoring data. Physically based models, while comprehensive, are labor-intensive and inherently uncertain. Our study leveraged AI to address these limitations, focusing on direct streamflow drought estimation using statistical threshold levels without a physically based model. Models were developed and tested using inflow data and meteorological variables from four major South Korean dams. The Threshold Level Method (TLM) was applied to daily inflow data to define drought events, creating a time series for model training. We utilized the XGBoost algorithm, integrating comprehensive meteorological data to enhance the accuracy and reliability of the drought predictions. Our findings show that AI models can effectively identify and forecast streamflow droughts, even with limited streamflow data, by using meteorological inputs. The results demonstrated significant drought patterns and characteristics across different threshold levels and time resolutions. This application provides a robust framework for integrating advanced AI techniques in hydrological studies, offering practical insights into water resource management and drought planning, particularly in semi-gauged basins where baseline data is available but limited.