Hierarchical Flood Alert Frameworkto Curb Underwarnings

Abstract Based on multisource urban flood data from Yangpu District during 2025, this study develops a hierarchical sequential decision-making framework that decouples risk detection and severity assessment through a dual-output long short-term memory (LSTM) architecture. Compared to conventional multiclass LSTMs, the framework achieves superior overall accuracy (99.77% versus 99.51%) and F1-score (0.8818 versus 0.4503), with a significantly lower false alarm rate (0.02% versus 0.06%). The framework demonstrates strong warning detection capability (62.92% exact level matching) and perfect performance on high-level warnings (Levels II and III). Compared to 79.78%, the underwarning rate of 37.08% indicates the framework’s effectiveness in alleviating the underwarning issue.