Abstract: The rainfall threshold method is widely used in flash flood warning and forecasting, but its traditional applications usually rely on static indexes and fail to identify disaster risk levels. In this study, a grading warning method is developed based on dynamic rainfall thresholds and applied to a small watershed in the downstream of the Minjiang River in Fujian Province. It uses an artificial neural network to construct a statistical relationship between runoff thresholds and risk factors to obtain the spatial distribution of runoff thresholds in the ungauged areas. By defining four risk levels of flash floods, it determines the corresponding runoff thresholds and rainfall thresholds through a frequency analysis based on long-term hydrological simulations. Results show that by considering the effect of antecedent soil moisture on flash floods, this method can prolong the forecast period and improve the monitoring capability significantly. Compared with traditional methods, its adoption of flash flood risk levels enhances its applicability to real flash flood warning.

Key words: flash flood warning, dynamic rainfall threshold, flash flood risk level, small watershed