Artificial Intelligence for Urban Flood Early Warning: A Review of Recent Advances

Abstract

Urban flooding is among the most frequent and costly climate-related hazards, and short-term early warning is one of the few interventions that can reduce its impact without new physical infrastructure. This review examines how artificial intelligence has been applied to short-term urban flood early warning, based on 67 studies published between 2019 and 2026 and retrieved from Web of Science and CNKI. Studies were screened through a four-stage PRISMA process and coded across ten fields, including method family, data source, forecast lead time, sample size, validation protocol, and the use of social-attribute variables. Reported accuracy is high across all method families, with R² values typically above 0.95, but these figures are not comparable across studies because the metrics, sample sizes, validation protocols, and evaluation targets all differ. Four systematic limitations emerge. First, 55 of the 67 studies train on simulator-generated labels rather than observed inundation, so the reported accuracy measures fidelity to a simulator, not to a real flood. Second, 53 of the 67 studies include no social-attribute variable in the model input. Third, validation protocols and sample sizes vary by orders of magnitude and are reported inconsistently. Fourth, none of the studies reports what happened after a warning was issued, so the claim that AI-based early warning improves flood outcomes remains an assumption rather than a finding.