Predicting hydrological drought indices using a hybrid data-driven model incorporating hydrological, geomorphological, and human activity impacts

This study presents a hybrid data-driven model to predict hydrological drought indices by integrating geomorphological, hydrological, and human activity factors. The model is trained using streamflow data simulated by the SWAT (Soil and Water Assessment Tool) and incorporates spatial zoning via Self-Organizing Map (SOM) networks to account for spatial variability across different zones. Each zone is trained independently using a ConvLSTM (Convolutional Long Short-Term Memory) model, which captures spatial and temporal information critical to hydrological time series data. Key input factors include geomorphological features such as drainage area, stream order, land cover, and hydrological and meteorological conditions like precipitation and evapotranspiration. Human activity factors, such as groundwater abstraction and industrial water consumption, are also integrated to reflect their impact on drought conditions. The trained model outputs two key hydrological drought indices, the standardized runoff index (SRI) and drought deficit volume, which are used to assess drought severity and further employed to calculate more metrics concerning drought termination. The hybrid model enhances drought prediction accuracy by leveraging the spatial and temporal dynamics of the watershed system without the additional use of a hydrological model. With a 30-day (1-month) prediction window, the model effectively captures temporal drought patterns while maintaining a balance between accuracy and computational efficiency. Furthermore, key evaluation metrics confirm the model’s accuracy and robustness. The Mean Relative Error (MRE) is less than 0.058, indicating minimal prediction error, while the Nash-Sutcliffe Efficiency (NSE) is greater than 0.905, demonstrating strong agreement with observed values. Additionally, the Pearson Correlation Coefficient (PCC) exceeds 0.976, highlighting a near-perfect correlation between predictions and actual data. These findings confirm the model’s reliability and effectiveness in drought prediction. These improvements provide valuable insights for efficient water resource management and drought impact mitigation.