Accurate and real-time monitoring of extreme precipitation is of great significance for improving flood forecasting， however， the current geostationary satellite-based precipitation products are generally characterized by low precipitation recognition accuracy， which seriously affects their application in flood warning. Based on the infrared band brightness temperature data observed by Himawari-8 geostationary meteorological satellite， ERA5 reanalysis atmospheric profile data and ground rain gauge observation data， this study developed a set of real-time precipitation recognition method suitable for geostationary meteorological satellites by establishing a real-time and dynamic precipitation recognition model based on random forest. On the one hand， this method solves the problem that the precipitation recognition accuracy of the static training machine learning models decays with time by introducing real-time hourly precipitation data of the surface rain gauge for real-time training of the precipitation recognition model. On the other hand， it effectively improves the accuracy of precipitation recognition based on the infrared data of the geostationary meteorological satellite by adding the atmospheric environmental condition data closely related to the formation and development of precipitation. The hourly precipitation observation data of 2 157 surface rain gauge stations in Chinese Mainland are used for verification. The proposed precipitation recognition algorithm has a POD of 0.73， a FAR of 0.49， and a CSI of 0.43 on an hourly scale. All indicators are better than the real-time precipitation product GSMaP_NOW and the official real-time precipitation estimation product QPE of FY4A.