High-Resolution Quantitative Precipitation Estimation Using Geostationary Satellite Observations and Deep Learning

Geostationary satellites have long been used for precipitation retrievals. However, the performance of geostationary satellite-based precipitation is still limited due to the limitations of conventional parametric algorithms. In this paper, we propose a deep learning framework for precipitation retrievals using multi-channel observations from geostationary satellites, including GOES-16 and GOES-18 for the western hemisphere, and Himawari-8/9 in Asia. In particular, a hybrid convolutional neural network (CNN) model is designed for precipitation detection and estimation using cloud-top Infrared brightness temperatures, the lightning flash rate (if available), and the digital elevation model information. The precipitation estimates from ground-based weather radars are used as target labels in the training phase and references in evaluations. The experimental results show that the proposed deep learning framework can achieve promising precipitation retrieval in their respective domains and can be further extended to the nearby regions where ground-based radars are not available.