Estimation of hydrological drought using regional risk ranking and machine learning algorithms

Abstract Hydrological drought forecasting remains challenging due to non-linear propagation processes, temporal subsurface lags, and strong spatial heterogeneity across river basins. Conventional deep learning frameworks routinely train monolithic global networks across entire watersheds, which often induces parameter dilution and obscures localized recovery dynamics. To resolve these challenges, this study develops a novel, two-stage artificial intelligence framework that integrates density-based spatial risk regionalization with Gated Recurrent Unit (GRU) deep neural networks. Utilizing multi-decadal anomalies of the Standardized Runoff Index (SRI) and Drought Deficit Volume (DDV), sub-basins in the Zengwen River Basin are first quantitatively clustered into distinct drought risk tiers via DBSCAN. Zone-tailored GRU architectures are then trained on a multi-source fusion of hydro-environmental and anthropogenic drivers to forecast monthly SRI and DDV time series in Stage 1, and derive four operational event-level drought metrics—duration, cumulative deficit volume, maximum monthly deficit, and termination rate—in Stage 2. Empirical findings reveal that Stage 1 predictions achieve exceptional accuracy, with coefficient of determination (R 2 ) values exceeding 0.96, Pearson correlation coefficients (PC) above 0.97, and Mean Relative Errors (MRE) reaching as low as 0.039 for deficit volume estimation. Crucially, incorporating prior risk regionalization reduces prediction errors for second-stage event-level metrics by 65% to 82% compared to conventional non-classified GRU models, with R 2 values consistently remaining above 0.97 across all risk categories. Furthermore, the trained deep learning engine achieves a 175 times computational speedup over full SWAT physical simulations, processing 50 years of basin-wide data in just 14.2 s while maintaining high spatial fidelity. By effectively bridging continuous index projections with actionable event-scale recovery metrics, this risk-partitioned framework offers a robust, computationally efficient tool for real-time reservoir dispatch and disaster mitigation.