The study on precipitation products corrected by optimal selection FY-4B AGRI channel based on random forest algorithm

To improve the accuracy of quantitative precipitation products derived from regional satellite observations, the level 1 radiation observation products, level 2 cloud detection products, and level 2 quantitative precipitation estimation operational products from the FY-4B geostationary radiation imager AGRI (Advanced Geosynchronous Radiation Imager), and national rain gauge observations collected between June 1 and July 31, 2022, in China were utilized in this study. The random forest algorithm was employed to establish separate models for predicting the presence or absence of rainfall and estimating rainfall amounts during the day and at night. The importance of 15 channels from the AGRI level 1 radiation observations was evaluated, and suitable channels for determining and estimating ground rainfall during the day and night were selected. Judgment and retrieval models based on these optimal channels were then created. Finally, using the models based on their respective optimal channels, precipitation for August 2022 was estimated and compared with the FY-4B AGRI level 2 quantitative precipitation estimation operational products. The results are as follows: (1) Channels 7 and 8 in the mid-wave infrared range were of lower importance during the day, and the daytime judgment and retrieval models did not use these channels. However, the nighttime models employed all infrared channels, and the retrieval precipitation closely matched the national gauge observations. (2) The optimal channel judgment models for both day and night outperformed the FY-4B precipitation operational products in identifying whether precipitation occurred, particularly showing a significant improvement in hit rates. Nevertheless, rainfall retrieval from both optimal channel retrieval models tended to overestimate rainfall amounts, with daytime retrieval precipitation being more accurate than nighttime. (3) The errors of rainfall retrieval from both daytime and nighttime optimal channel models were both lower than those from the FY-4B precipitation operational products. Both the precipitation operational products and the two optimal channel model estimates tended to overestimate light rainfall and underestimate heavy rainfall, with the underestimation becoming more severe as precipitation intensity increased. (4) For the retrieval of different precipitation magnitudes using optimal channel models, improvements in root mean square errors were observed for nearly all levels of rainfall except for nighttime heavy rain, with the most significant accuracy enhancement seen in the small to moderate rainfall range.