Development of Deep Learning Based Data Fusion Approach for Accurate Rainfall Estimation Using Ground Radar and Satellite Precipitation Products

Rainfall estimation based on onboard satellite measurements has been an important topic in satellite meteorology for decades. A number of precipitation products at multiple time and space scales have been developed based upon satellite observations. For example, NOAA Climate Prediction Center has developed a morphing technique (i.e., CMORPH) to produce global precipitation products by combining existing space based rainfall estimates. The CMORPH products are essentially derived based on geostationary satellite IR brightness temperature information and retrievals from passive microwave measurements (Joyce et al. 2004). Although the space-based precipitation products provide an excellent tool for regional and global hydrologic and climate studies as well as improved situational awareness for operational forecasts, its accuracy is limited due to the sampling limitations, particularly for extreme events such as very light and/or heavy rain. On the other hand, ground-based radar is more mature science for quantitative precipitation estimation (QPE), especially after the implementation of dual-polarization technique and further enhanced by urban scale radar networks. Therefore, ground radars are often critical for providing local scale rainfall estimation and a "heads-up" for operational forecasters to issue watches and warnings as well as validation of various space measurements and products. The CASA DFW QPE system, which is based on dual-polarization X-band CASA radars and a local S-band WSR-88DP radar, has demonstrated its excellent performance during several years of operation in a variety of precipitation regimes. The real-time CASA DFW QPE products are used extensively for localized hydrometeorological applications such as urban flash flood forecasting. In this paper, a neural network based data fusion mechanism is introduced to improve the satellite-based CMORPH precipitation product by taking into account the ground radar measurements. A deep learning system is designed and trained using the satellite IR data and passive microwave rainfall retrievals as the input and the CASA DFW rainfall products as the target output. The detailed prototype architecture will be presented. The deep learning based rainfall products are evaluated using surface rainfall measurements from gauges.