Short-Term Predictions of Hydrological Events on an Urbanized Watershed Using Supervised Classification

Many tasks of operational watershed management at the local level require stream flow predictions delivered to decision makers in a timely manner. In highly urbanized watersheds with an impermeable surface, stormwater runoff can cause rapid increases in water levels in streams leading to flood and even flash flood events. Usually, such rapid increases in water flow characteristics are predicted by process-based models with high levels of uncertainty. In this study, the prediction of magnitudes of the stream hydrological characteristics is replaced by the forecasting of an event (i.e., flood or no-flood) using data collected by stream and rain gauges at the watershed. The proposed approach is based on a black box model developed as an ensemble of classifiers generated by independent inducers to predict the class of a future hydrological event in a small highly urbanized watershed. Eight inducers were investigated in the phase space reconstructed from observation data using time-delay embedding extended to multiple observation sites. Five inducers were selected for the ensemble, where the final decision is made by majority vote. The developed model generates 45-minute and hourly predictions of high-flow events with more than 80 % precision – a threshold used in operational flood management. Model site-specific parameterization is replaced by the training step using observation data on water levels and precipitation which are collected at 15-minute intervals and are readily available. The proposed approach to developing a prediction tool can be used by local authorities as one of the methods for flood management.