Precipitation retrievals for ground-based microwave radiometer using physics-informed machine learning methods

Abstract

Precipitation is complex due to its significant temporal and spatial variability, and current mainstream precipitation estimation techniques have their inherent limitations. The complementary role of ground-based microwave radiometer in precipitation monitoring to these technologies is gaining increasing attention. Based on the physical characteristics of microwave radiation signals affected by raindrops in the atmosphere, this study presented two novel machine learning based rain rate retrieval algorithms, random forest (RF) and gradient boosting decision tree (GBDT), for a ground-based microwave radiometer (MWR) over Swiss Plateau from 2008 to 2010. Both methods are trained using the rain rate observed by the remote sensing technology micro rain radar (MRR) as the target variable, and consider meteorological parameters in the feature input. For data preprocessing of the retrieval methods, outliers and noise in the MRR rain rate are removed. Cross-validation results show that both RF-based and GBDT-based methods achieve superior precipitation estimation performance, with R² values of 0.96 and 0.95 and mean square error of 0.01 mm/h and 0.02 mm/h, respectively. Comparing light gradient-boosting machine (LightGBM) and support vector machine (SVM) algorithms, rain rate retrieval based on RF and GBDT are highly competitive in terms of accuracy and model training timeliness, respectively. This study offers an accurate and reliable method for high temporal resolution (10 s) precipitation estimation from MWR under complex terrain conditions, and it also has the potential for application in other regions and with other tropospheric microwave radiometers.