An explainable and causally informed machine learning framework for hydrological drought transitions, propagation,...

Hydrological droughts (HDs) pose increasing challenges to water-system management, yet the processes governing their wet-dry transitions, propagation delays, and resulting vulnerability remain poorly resolved. Here, we developed an explainable and causally informed machine-learning (ECML) framework to distinguish basin-wide attribution from region-specific relationships and applied it to the monsoon-influenced Yuan-Red River Basin. Meteorological predictors accounted for 63.97% of the total predictive attribution to HD variability, led by precipitation of 36.81%, which exhibited a nonlinear predictive response with a model-derived breakpoint at 38.89 mm month⁻¹. However, its basin-wide importance was not accompanied by spatially coherent lagged conditional relationships. Five response regimes further revealed how topographic and land-surface contrasts modulated the hydroclimatic signal, whereas human-related influences were weaker and localized. Wet-dry transitions were most closely associated with monsoon-related variations in moisture supply and atmospheric demand. By contrast, drought propagation lags ranged from 1 to 7 months and varied regionally with climatic conditions, vegetation, topography, and human-related indicators, consistent with differences in catchment buffering and runoff transmission. The Poly-Adaptable Hydrological Drought Vulnerability Index (PHDVI) summarized this heterogeneity, revealing distinct vulnerability hotspots and strong seasonality, with the highest values in winter and the lowest in summer. These findings demonstrate that basin-scale predictor importance does not imply spatially uniform local relevance and that drought transition, propagation, and vulnerability reflect distinct relationships. The framework provides a process-informed basis for understanding why similar climatic anomalies can produce contrasting hydrological outcomes and supports stage- and region-specific drought early warning and adaptive water-resource management.