A Hybrid Approach to Physical and Deep Learning Models for Radar-Based Precipitation Nowcasting

Abstract

This study proposes a novel approach to improving radar-based precipitation nowcasting using a boosting algorithm to blend traditional physics-based extrapolation and data-driven deep learning (DL). Here, a semi-Lagrangian approach (PySTEPS) and a data-driven DL model (RainNet) are considered as two representative types of nowcasting models of precipitation. The light gradient boosting machine (LightGBM) model is adopted as a boosting algorithm due to its efficiency and ability to correct for biases sequentially. The boosting model significantly outperforms both RainNet and PySTEPS with up to 90 min of lead time for critical success index (CSI), probability of detection (POD), false alarm ratio (FAR), root-mean-square error (RMSE), and fractions skill score (FSS) at the 0.1-, 1-, and 5-mm/h thresholds (TSs). The CSI at the 0.1- and 1-mm/h TSs for the boosting model is approximately 10% higher than that of both the RainNet and PySTEPS models. In addition to high CSI, the boosting approach can also achieve superior results in POD and RMSE compared with the RainNet and PySTEPS models across various TSs and lead times. The proposed modeling framework significantly outperforms the individual models in predicting rainfall intensity and spatial distribution, highlighting the potential of blending precipitation nowcasting models.