A Dynamic Two-Step Random Forest-Based Method to Improve Real-Time Precipitation Retrieval

Accurate and timely precipitation data plays a crucial role in flood monitoring, forecasting and emergency management. The primary challenge with current real-time precipitation products is their limited accuracy in identifying precipitation. To improve the accuracy of precipitation retrieval, a dynamic two-step precipitation retrieval method was developed, i.e., first identifying precipitation and then estimating its intensity. The geostationary satellite infrared brightness temperature data, atmospheric profile data, and rain gauge observations were used to construct the precipitation identification model and precipitation intensity retrieval model based on a random forest machine learning method. This method was applied to precipitation retrieval in China and surrounding areas using the Himawari-8 observations and the ERA5 atmospheric profile data. The results of hourly precipitation identification showed good agreement with rain gauge observations, with a Probability of Detection of 0.725 and a False Alarm Ratio of 0.426. The estimated hourly precipitation intensity agrees well with the rain gauge data and shows a lower error than the GSMaP_NOW real-time precipitation product. The good performance of the developed method in this study for real-time precipitation monitoring was also confirmed in different seasons.