Abstract: In response to the current problem of high false alarm rate and low accuracy in flash flood warning, a systematic improvement plan is proposed based on the existing flash flood warning model. This plan focuses on differentiated classification warning, and accurately matches single station warning, multi station warning, surface rainfall warning, and flow warning methods based on the characteristics of the upstream watershed of the prevention and control object and its sources of flash flood risk. It is supplemented by optimizing the association rules between villages and warning criteria, and revising the warning network value based on the flood control capacity of villages. The pilot application of this plan in Fujian Province has shown significant advantages: the evaluation of the warning effect on 620 villages shows that the false warning rate has been reduced by 11.5% to 85.7% without missing reports. Among them, the total warning volume of typical heavy rainfall events in 2024 and 2025, which are dangerous (immediate transfer) and alert (preparation for transfer) levels, has generally decreased by more than 50%. Through the case of Getou Village, it has been verified that the improved classification warning mechanism can more accurately adapt to different types of prevention and control objects and sources of flash flood risks. Practice has shown that improved warning methods can effectively improve warning accuracy, reduce warning false alarms, save emergency resources, enhance warning credibility, and provide replicable and promotable technical paths and practical experience for upgrading the national flash flood monitoring and warning system.