Research On Precipitation Inversion From Brightness Temperature Data Of Sunflower-8 Satellite Based On Deep Learnin

Precipitation at various scales in the atmosphere is crucial for the water and energy cycle,and it also have an important impact on human activities,particularly the productivity of societies.Extreme weather and secondary geological hazards triggered by precipitation can lead to significant losses in terms of life,property,and economy.Therefore,it is essential to monitor precipitation activity comprehensively,accurately,and in a timely manner.Precipitation observation and identification,including the intensity and extent,can not only provide critical information for studying precipitation development mechanisms but also aid in disaster mitigation and prevention,ensuring stable social development.Currently,precipitation retrieval algorithms based on the infrared channel of geostationary satellites has a wide range of applications in operational meteorology.With the continuous development of artificial intelligence technology,new precipitation intensity retrieval methods and quantitative precipitation estimation models are proposed based on machine learning(ML)or deep learning(DL).However,many studies have shown that although ML/DL technologies have achieved good results,several problems still exist,such as underestimation of precipitation intensity in the retrieval,incorrect prediction of precipitation zones and blurring of the output contours.In this study,Himawari-8 geostationary imager observations and GPM multi-satellite precipitation estimate(IMERG)information for the summer season in southern China are used.The target dataset of GPM IMERG needs to wait for 3.5 months to be obtained,while the data of Himawari-8 can be obtained in real-time.GPM data is essentially a 30 minute precipitation intensity,which is the average precipitation within 30 minutes.Himawari-8 inversion of precipitation advantages can achieve precipitation intensity of 30 minutes per 10 minutes,filling the gap in surface rainfall gauge data to some extent.The Himawari-8 satellite data has higher horizontal resolution and can provide a more refined description of precipitation activities.Although GPM IMERG precipitation products are close to real precipitation,the timeliness and resolution of the data are difficult to meet the needs of existing precipitation observation operations and research.Therefore,GPM IMERG is used as the baseline data of precipitation in this study,and Himawari-8 data is used for grid to grid point reflection to generate high-resolution,timely,fast and effective satellite precipitation data.Based on the advanced deep learning algorithm,a precipitation inversion model for the infrared channel observations of the Sunflower 8geostationary satellite was constructed.Sensitivity studies for different observation channels,times and periods of input satellites and deep learning models were carried out.And a new depth model structure is established to effectively improve the inversion accuracy of precipitation intensity and extent.In this paper,a machine learning training set of 9 infrared channels of bright temperature data and GPM IMERG half-hourly average precipitation on the Himawari-8 geostationary satellite from 2016 to 2019 was constructed and then divided into a training set(5352 samples in total),a validation set(1784)and a test set(1784).The U-Net model,the most widely used and stable deep learning technique,is chosen to analyze the contribution of different IR channels of the Himawari-8 satellite to the precipitation estimation.The retrieval results of single and multiple channels are compared,and it is found that the Channel 13 yields better results.Subsequently,the accuracy of model outputs with the input of single-moment observation and the input of multiple-moment combinations is compared,and it is found that the latter did not improve the accuracy of the inversions.Finally,to further improve the precipitation prediction,the accuracy of three different DL models is compared,including U-Net,pix2 pix GAN and ConvMixer.The results show that pix2 pix GAN is the best model,but its output still suffers from omission and underestimation.In order to further optimize the inversion effect of pix2 pix GAN,in this paper,we construct a B-pix2 pix GAN model based on the pix2 pix GAN model of the two-step forecast method to improve the accuracy of precipitation inversion.