Based On The AGRI Observation Of Fengyun-4 Satellite, The Random Forest Algorithm Is Used To Invert Ground Precipitation

In order to explore the application of the Random Forest algorithm in the imager data of the Feng Yun-4A satellite,on the basis of evaluating the business precipitation estimation product of the FY-4A AGRI(Advanced Geosynchronous Radiation Imager)with the hourly precipitation data of 2167 ground meteorological stations from May 14,2018 to December 31,2019,the Random Forest method,one of the Machine Learning algorithm,is used to retrieve the hourly precipitation based on the level 1 data of AGRI,and the retrieval result is evaluated with the hourly precipitation data of the ground meteorological station and the AGRI precipitation product in June 2020.Main results are as follows:(1)Taking the precipitation data of ground meteorological stations as the true value,the false alarm rate(FAR)of AGRI’s business precipitation product is 0.69,the missing alarm rate(MAR)is 0.60,the probability of detection(POD)is 0.40,the mean bias(MB)is-1.0131 mm,the mean absolute error(MAE)is 2.3622 mm,the root mean square error(RMSE)is 5.0235 mm.Precipitation product has high estimates of precipitation for light rain,and low estimates for other rainfall levels.As the rainfall level increases,the product has a greater error in the estimation of precipitation.(2)Based on the actual AGRI observations,the test results of the random forest precipitation event judgment models show that,compare with the AGRI business precipitation product,the random forest models have lower MAR,higher POD,but higher FAR.Overall,the random forest models have a higher accuracy in judging precipitation events.The random forest precipitation event judgment models are more difficult to distinguish the situation of clouds without precipitation,resulting in higher FAR.(3)Based on the actual AGRI observations,the test results of the random forest precipitation intensity retrieval models show that,the daytime random forest method has high estimates of precipitation,when the AGRI precipitation product has low estimates during the day.Both overestimate hourly precipitation at night,and the product overestimation is more serious.Regardless of day or night,the random forest method has higher retrieval accuracy than AGRI precipitation product.The graded test results show that in the light rain level,both the random forest method and the AGRI precipitation product overestimate the hourly precipitation;at the moderate rain level and above,both methods underestimate the precipitation,and the degree of underestimation increases as the precipitation level increases.At the level below torrential rain,the random forest method has better retrieval effects on hourly precipitation than AGRI precipitation product;in the torrential rain level,the random forest method during the day is better than the precipitation product,but the random forest method at night has a lower accuracy.(4)Comparing the daytime models with the corresponding nighttime models,regardless of the random forest precipitation event judgment models or the random forest precipitation intensity retrieval models,the errors of the nighttime models are higher than those of the daytime models.The reason is that,when the high value area of visible light albedo is not consistent with the low value area of infrared brightness temperature,the precipitation area and precipitation intensity retrieved by the daytime models that add the albedo data of the visible light channels to the explanatory variables are closer to the ground observations,compared with the nighttime models that only use brightness temperature data.