https://doi.org/10.1016/j.jhydrol.2021.127113

Infrared (IR) is an important data source for satellite quantitative precipitation estimation, and has been widely applied in the fields of meteorology, hydrology, and agriculture. In the past decades, a series of IR retrieval algorithms have been developed to support the production of IR-based and IR-microwave merged precipitation products. Recently, deep learning techniques such as the convolutional neural network (CNN) show great potential in obtaining IR precipitation estimates with higher accuracy than traditional retrieval algorithms. In this study, we present an upgraded version of IR Precipitation Estimation using CNN (IPEC), i.e., IPEC version 2 (IPEC-V2), with a new end-to-end manner. IPEC-V2 is used to generate a precipitation dataset based on IR data from China’s Fengyun (FY) geostationary satellites (FY-2F, 2G, and 4A). To overcome the difficulty of model training in regions with sparse observations, IPEC-V2 models are firstly pre-trained over the data-rich Continental US (CONUS) using IR data from the Geostationary Operational Environment Satellite (GOES). The models are then transferred to China through re-training with multi-band IR signals from FY satellites. Finally, a long-term record of high-resolution FY IR precipitation estimates is produced during the period from November 9, 2012 to February 28, 2021, which is named as IPEC-FY. IPEC-FY shows better performance than the baseline Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System (PERSIANN-CCS) with 76.2% gain in Pearson’s correlation coefficient and 18.4% gain in root mean squared error in China. This study shows that transfer learning is an effective way to build CNN models in regions without enough observations, and the high-quality IPEC-FY can act as an alternative precipitation dataset for research in China.

https://doi.org/10.1016/j.jhydrol.2021.127113