Towards Operational Flood Early Warning: A Modular Evaluation of Physics-Informed Constraints, Conformal Prediction, and Missing-Data Robustness

An effective flood early warning system depends on accurate forecasting, calibrated uncertainty estimates, robustness to missing data, and effective risk assessment. This study evaluates and compares different techniques using the Fish River catchment and a complementary panel of 15 CAMELS catchments that includes Fish River, as case studies. The results indicate that, under matched model settings, the tested precipitation-monotonicity constraint provides no stable improvement in forecast accuracy: the mean paired difference in macro-averaged Nash–Sutcliffe efficiency (NSE) is $+0.0017$. Conformal prediction provides overall empirical coverage close to the nominal 90% level. Adjusting the intervals according to forecast values increases coverage under high-discharge conditions, but also widens the intervals; coverage nevertheless remains below the target. Experiments involving missing precipitation data show that model performance varies across missing-data scenarios and imputation methods, while predictive accuracy declines with increasing lead time. Using a threshold set at the 99th percentile of daily runoff depth during the training period to identify extreme events, the six-day lead time produces a critical success index of 0.0432 and a Brier skill score of $-0.1800$, with no detected event onsets. Cross-catchment analysis shows pronounced spatial heterogeneity, and the tested physical constraints offer no consistent advantage in accuracy. These findings support separate evaluation of forecast accuracy, interval calibration, precipitation-gap sensitivity, and event-level warning performance. The modular comparison identifies trade-offs under historical conditions and provides a basis for further validation before operational deployment.