AI-enabled IoT systems for urban flood resilience: a validated case study from Dubai

Short-duration, high-intensity storms in arid cities produce flash floods that routinely defeat conventional drainage and threshold-based warnings. The author presents and validates an end-to-end artificial intelligence (AI)–Internet of Things–geographic information system (GIS) framework that operationalises flood resilience as a perception–learning–decision pipeline: dense, latency-aware sensing; leakage-safe, eventwise machine-learning inference; and GIS-linked asset-level alerting. The framework is implemented and stress-tested in Dubai using multi-source IoT streams (rainfall, surface runoff, and water level) collected from 2019 to 2024. Models were trained with chronologically partitioned events, strict causal feature construction, and explicit propagation of sensing uncertainty; all preprocessing, hyperparameters, and code are archived for reproducibility. On temporally independent test events, the prototype achieved precision-recall area under the curve (PR-AUC) = 0.86 (95% CI 0.80–0.89), precision = 0.81 (0.76–0.85), and recall = 0.88 (0.83–0.92), outperforming a rule-based baseline (ΔPR-AUC ≈ 0.14, paired bootstrap p < 0.02). Effective operational lead time averaged 30–45 min, and GIS-mapped alerts yielded intersection-over-union ≈ 0.73 and asset-level trigger accuracy ≈ 0.84. Robustness experiments show bounded performance degradation under random sensor loss and identify spatially clustered outages as the principal risk. The study delivers a reproducible, engineering-grade blueprint for deploying AI-enabled flood early warning in arid urban contexts, directly supporting UN SDG 11 (Sustainable Cities and Communities).