Regional Flood Early Warning Calibration Under Geographic Shift and Sparse Sensors

Abstract

Flood early warning models must remain reliable when gauge histories are incomplete, and target regions differ from the regions used for training. This paper presents a leakage-checked regional benchmark for short-horizon discharge forecasting using CAMELSH/USGS data from 30 stations across six HUC02 regions. The evaluation uses station-specific q95 and q99 warning thresholds estimated from pre-2024 history, six-fold leave-HUC02-out validation, sparse-sensor stress tests, event-level warning metrics, and split conformal prediction intervals. At the 24-hour horizon, persistence is the strongest overall benchmark, with an MAE of 4.229, q5 F1 of 0.738, missed-event rate of 0.266, and absolute coverage error of 0.051. Among learned models, the transformer gives the strongest regional average, with an MAE of 5.171 and a q95 F1 of 0.671. Graph-dropout and the proposed availability-aware graph model retain more relative q95 F1 at 50% sensor availability, but neither exceeds persistence in absolute warning skill. Calibration is also region-dependent: in HUC02 17, empirical 24-hour coverage falls to 0.564 for graph-dropout and 0.551 for the proposed method, compared with the nominal 0.90 target. These results show that sparse-sensor robustness, event detection, and calibration should be jointly evaluated under geographic shifts.