Real-time anticipatory urban flood warning using CCTV and Page–Hinkley change detection

Effective flood risk mitigation in urban areas increasingly relies on the integration of real-time information and proactive decision-making. This study presents a lightweight and operationally practical framework for anticipatory urban flood early warning using existing CCTV infrastructure. Rather than performing long-horizon hydrological forecasting, the proposed system focuses on detecting incipient flood-risk signals from surveillance video and translating them into interpretable short-term warning indicators. The framework extracts visual water-occupancy signals from CCTV footage and applies Page–Hinkley-based change-point detection to identify statistically meaningful rising trends under noisy urban conditions. These trends are further characterized using short-term regression to estimate expected time-to-threshold (ETA) and probabilistic flood risk, which are integrated into a hysteresis-based dual-alert mechanism for reliable warning issuance. Using the best-performing preprocessing pipeline, the flood segmentation model achieved an mAP of 85.1%. Empirical validation across multiple urban flood scenarios further confirmed zero false positives under normal conditions (FPR = 0) and secured an average early-warning lead time of 32.9 ± 5 s during the pre-flood stage. The results indicate that the framework provides a robust and economically scalable alternative for real-time urban flood early warning, particularly in environments where dense sensor networks or computationally intensive hydrological models are impractical. • A lightweight and operational AI framework is developed for short-term urban flood early warning using existing CCTV infrastructure. • Flood-related visual information is transformed into interpretable water-occupancy time series for robust real-time analysis. • Page–Hinkley change detection enables early identification of rising flood risk while suppressing false alarms under normal conditions. • Short-term trend analysis provides estimated time-to-threshold to support anticipatory warning decisions. • Scenario-based validation demonstrates reliable real-time alerting performance across different urban flood stages.