Retrieval of Radar-Based Convection Intensity Parameters from Passive Microwave Observations via Machine Learning

Abstract Spaceborne precipitation radar (PR) observations have strongly promoted the understanding of global precipitation. Nevertheless, only two spaceborne PRs are currently in routine operation, which limits the applications of PR data in precipitation monitoring and research. In this work, a machine learning (ML) algorithm is applied to derive radar-based convection intensity parameters with multiple-channel passive microwave observations from the GPM Microwave Imager (GMI). The algorithm first determines whether the precipitation features (PFs) reach a certain intensity, for example, a maximum radar reflectivity exceeding 20 dB Z . The classification process achieves a high score, with an accuracy rate of 97% and a recall rate of 100%. Then, the intensity parameters of the PFs meeting a certain intensity, such as the maximum 20-dB Z echo top height (MAXHT20) and the temperature at MAXHT20 (T-MAXHT20), are fitted with the optimal ML algorithm, a random forest (RF) model. The RF model is applied to derive MAXHT20 with global PFs, and the root-mean-square error and the coefficient of determination are 1.07 km and 0.79, respectively. Nevertheless, there are significant regional discrepancies in biases between the ML model results and the observations. Therefore, investigations are conducted in different regions and for different PF intensities. The results imply that applying the algorithm in a specific region can improve model performance and that the algorithm performs best in cases with considerable amounts of ice particles. These results confirm the feasibility of applying passive microwave observations to derive radar-based parameters in the assessment of precipitation systems. Significance Statement Despite the significant advancements in global precipitation research made possible by spaceborne precipitation radar (PR), the limited operation of only two PRs currently restricts their application in precipitation monitoring. We propose a novel method for deriving radar-based parameters using machine learning to retrieve these parameters from passive microwave observations. Our results show R 2 and RMSE values of 0.79 and 1.07 km for the retrieval of the maximum 20-dB Z echo top height, demonstrating the feasibility of this approach. These results could help us make up for the lack of spaceborne precipitation observation with passive microwave observation. Future research could further improve the algorithm by focusing on pixel-level data analysis, using observation of geostationary passive microwave sensors, and incorporating additional microwave bands to enhance the retrieval capabilities in precipitation systems.