D-PRA: A Dynamic Two-Step Real-Time Precipitation Retrieval Algorithm Based on Geostationary Satellite Observation

Real-time precipitation retrieval is crucial for timely warnings of extreme weather events such as heavy rain or floods. Geostationary satellite observations combined with machine learning methods provide an effective way to achieve real-time precipitation retrieval yet are hampered by two issues. The traditional stable model, pretrained with historical data, lacks precipitation information from adjacent time periods, and its accuracy declines over time. Geostationary satellites can only capture cloud top information, leading to inaccurate localization of precipitation areas in estimations based solely on such data. To solve these problems, a novel dynamic two-step real-time precipitation retrieval algorithm (D-PRA) has been developed. It conducts real-time retrieval in two steps through dynamic data and variables’ selection: identifying precipitation and then retrieving precipitation intensity. It further incorporates atmospheric profile information to enhance the details of in-cloud and under-cloud conditions. D-PRA was applied to Himwari-8 observations and was validated with the rain gauge observations in Chinese regions at the hourly scale. The results showed that in comparison with Global Satellite Mapping of Precipitation (GSMaP)_NOW, D-PRA exhibits significant enhancements. The probability of detection (POD) attained 0.72, representing a nearly fivefold increment, and the root mean square error (RMSE) was 0.99 mm, a 25.5% reduction. Moreover, D-PRA is stable across seasons and time periods, demonstrating its good reliability and robustness. D-PRA has great potential to improve the accuracy of precipitation retrieval for extreme event monitoring and disaster management.