Nowcasting Multicell Short-Term Intense Precipitation Using Graph Models and Random Forests

Abstract

Short-term intense precipitation (SIP; i.e., convective precipitation exceeding 20 mm per hour) nowcasting is important for urban flood warning and natural hazards management. This paper presents an algorithm for coupling automatic weather station data and single-polarization S-band radar data with graph model and random forest for the nowcasting of SIP. Different from the pixel by pixel precipitation nowcasting algorithm, this algorithm takes the convective cells as the basic units to consider their interactions and focuses on multi-cell convective systems. In particular, the following question could be addressed: will a multi-cell convective system cause SIP events in the next hour? First, a method based on spatiotemporal superposition between cells is proposed for multi-cell systems identification. Then, the graph model is used to represent cell physical attributes and the spatial distribution of the entire system. For each graph model, a fusion operation is used to form a 42-dimensional graph feature vector. Finally, combined with the machine learning approaches, a random forest classifier is trained with the graph feature vector to predict the precipitation. In the experiment, this algorithm achieves the probability of detection (POD) 79.2% and critical success index (CSI) 68.3% with the data between 2015 and 2016 in North China. Compared with other precipitation nowcasting algorithms, the graph model and random forest could predict SIP events more accurately and produce fewer false alarms.