Hypoxia episodes in receiving waters near estuarine outlets pose persistent challenges to water-environment management because operational early warning is often hindered by noisy observations, class imbalance, and inter-annual distribution shifts. This study proposes an externally validated Event-Window early-warning framework that bridges high-frequency monitoring data and management-oriented decision support. An explainable gated recurrent unit model with temporal attention (GRU-Attn) was developed and evaluated using a strict External-Year test in 2025. To better reflect operational needs, model performance was assessed not only at the daily classification level but also at the event-window level. The model achieved a PR-AUC of 0.9138 for day-level prediction, while Event-Window aggregation further increased PR-AUC to 0.9723, reduced false alarms by 59% (from 22 to 9), and provided a median lead time of 2.0 days for severe events. To improve deployment transparency, a governance diagnostic layer integrating population stability index analysis, threshold reliability assessment, and attention-based temporal attribution was further introduced. The results show that combining External-Year validation with event-scale evaluation and transparent diagnostics can substantially improve the robustness and practical interpretability of hypoxia early warning under real-world distribution shifts.
Gao et al. (2026) studied this question.