Improving GOES-R/ABI rainfall estimates with ground-based radar observations using machine learning

Abstract:

The satellite-based precipitation products are essentially derived using either geostationary (GEO) satellite infrared (IR) data or low earth orbit (LEO) satellite passive microwave (PMW) measurements or a combination of both. Compared to the legacy GOES Imager, the Advanced Baseline Imager (ABI) on the GOES-R series provides measurements at much higher resolution in both spatial and temporal dimensions, which benefits many applications such as rainfall rate estimation. The operational GOES-R/ABI rainfall rate algorithm uses information from multiple IR bands of the ABI and is calibrated against microwave-derived rain rates to optimize the accuracy. In particular, the precipitation estimation algorithm estimates the precipitation occurrence and quantify its intensity using the cloud top brightness temperature information through a regressed model linking the brightness temperature to rainfall rate. However, it is challenging to characterize the particle size distribution in the clouds and resulting rainfall intensity with only the IR radiation and converted brightness temperatures. Multi-year analysis and comparison with ground-based radar observations show that the uncertainty associated with operational ABI rainfall rate products is large. Therefore, this paper proposes an innovative deep learning system to improve satellite rainfall retrievals based on the multi-mode, multi-channel observations from the ABI on the GOES-R series. The brightness temperature information observed by channels 8/10/11/14/15 of the ABI are used as input to this machine learning framework. The rainfall estimates from a well calibrated ground radar network are used as target labels in the training of the designed deep learning model. Independent verification shows that the deep learning-based rainfall estimation system performs very well at estimating precipitation time, intensity, and amount over a range of rainfall conditions, suggesting that machine learning‐based approaches should be considered in the future development of rainfall algorithm for the GOES-R/ABI.