Predictive Simulation and Early Warning in Flood Digital Twins: A Systematic Review of AI Model Performance, Warning Lead Times, And Operational Deployment

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Flooding is the most economically destructive natural hazard in the United States, with annual losses exceeding $60 billion and risk concentrated unevenly across communities and regions. This systematic review synthesizes evidence on AI-enhanced digital twin architectures for U.S. flood infrastructure management. Following PRISMA 2020 guidelines, a search across six academic databases (2020–2025) identified 3,847 records, of which 94 studies met inclusion criteria. The findings are clear. Long Short-Term Memory (LSTM) networks reach Nash-Sutcliffe Efficiency (NSE) values of 0.88–0.97 in well-gauged U.S. basins, consistently outperforming the U.S. National Weather Service's physics-based operational models. Physics-informed neural networks (PINNs) work in data-scarce U.S. watersheds, improving NSE by 0.04–0.11 over standard baselines. AI-augmented systems push meaningful warning lead times from the conventional 1–3 hours out to 6–24 hours in medium-sized catchments; when paired with numerical weather prediction ensembles, operational deployments have reached 72-hour horizons. Data assimilation (DA) adds NSE gains of 0.05–0.12 over open-loop simulations, with particle filter approaches outperforming ensemble Kalman filters by up to 0.18 in non-linear inundation scenarios; satellite-based DA opens viable pathways in U.S. regions where ground gauges are sparse. The gaps are equally clear. Only 31% of reviewed studies provide calibrated probabilistic predictions, and the evidence base is geographically concentrated in well-instrumented catchments, leaving data-scarce watersheds and socially vulnerable communities systematically underrepresented. The review closes with concrete, evidence-grounded recommendations for standardizing U.S. benchmarking frameworks and accelerating equitable operational deployment.