A progressive early warning system for flash drought based on meteorological precursors: A case study of the Dongjiang River Basin, China

Study region The Dongjiang River Basin in southeast China, a critical water source region for the Pearl River Delta with a subtropical monsoon climate and frequent dry–wet transitions. Study focus Flash droughts threaten water and food security due to rapid onset, yet operational warnings remain largely single-indicator based. This study developed a progressive early warning system using pentad-scale ERA5-Land root-zone soil moisture (1950–2025). Precursors were screened through collinearity diagnosis, time-lag correlation, and random forest importance, followed by Receiver Operating Characteristic (ROC) analysis, with predictive skill quantified by the area under the ROC curve (AUC) to determine warning thresholds within a two-stage regression-to-classification framework. New hydrological insights for the region Five precursors spanning three pentads prior to onset were identified. A two-pentad-ahead precipitation anomaly ( P lag2 , AUC = 0.74) and a one-pentad-ahead anomaly ( P lag1 , AUC = 0.92) formed the main warning chain, issuing L1 and L2 alerts 2 and 1 pentads ahead, respectively. Three weaker precursors ( P lag3 , ET / PET lag3 , T lag2 ; all AUC ≈ 0.5) were integrated into a three-category supporting-condition score (LOW/MEDIUM/HIGH). The system achieved a false positive rate of 6.3% (training) and 8.5% (validation), substantially lower than single-indicator (16.5%–18.6%) and logistic regression (16.3%–22.3%) baselines, with a higher Heidke Skill Score (0.370 vs. 0.262–0.269). Results were robust across 10 flash drought definition parameter combinations, and application to the 2019 event demonstrated timely, coherent warnings with no false alarms.