Research On Intelligent Methods For Satellite Precipitation Estimation And Forecasting

Accurate and efficient precipitation estimation and forecasting are of great significance for agriculture,transportation,finance,water conservancy,and other industries.Benefitting from the precision of global meteorological satellite monitoring,meteorological satellite observation data are growing explosively.Thus,it is a major challenge to effectively mine the important information hidden in massive satellite observation data for improving the accuracy of precipitation estimation and forecasting.How to model the nonlinear relationship between satellite observation information with high spatio-temporal resolution and precipitation is the key to improve the timeliness and accuracy of precipitation estimation.How to extract the spatio-temporal information of atmospheric motion evolution from satellite observation series is an urgent problem to improve the ability of precipitation forecast.In this paper,deep learning technology is used to address the above issues,and tropical cyclone precipitation and precipitation in arid areas are studied in detail.The main research contents are as follows:(1)A tropical cyclone precipitation estimation method based on multi-task convolution neural network is proposed.In order to solve the problem of how to learn robust precipitation features,the convolutional neural network is used to automatically extract the spatial features related to precipitation from satellite cloud imageries.The nonlinear relationship between satellite observations and precipitation is established by the nonlinear activation function.In addition,the idea of multi-task learning is introduced to consider the impact of tropical cyclone intensity on precipitation.In other words,another wind grade classification task is added to classify the wind grade near the center of the tropical cyclone into six categories to help improve the accuracy of the precipitation estimation task.The experimental results show that this method can effevtively describe the spatial distribution details of precipitation,and the multi-task network with wind grade classification branch can further improve the accuracy of tropical cyclone precipitation estimation.(2)A tropical cyclone precipitation forecasting method based on Transformer is proposed.In order to solve the problem of how to learn the precipitation trend from the satellite cloud imagery sequence,the 3D convolutional operation is adopted to extract local spatial features related to precipitation from satellite cloud imageries in both encoding and decoding stages.Meanwhile,the Transformer structure is added at the end of each encoding module to model the time and global context information,which is helpful to capture the precipitation trend and further improve the accuracy of precipitation forecasting.The experimental results show that the method with Transformer structure can effectively learn the spatial distribution details and trend of precipitation,and the forecasting performance is improved when the precipitation threshold is large.(3)A precipitation estimation and forecasting method in Xinjiang based on a two-stage convolutional neural network is proposed.In view of the drought and lack of rain in Xinjiang,China,and the extremely uneven distribution of precipitation samples,this method first uses convolutional neural network to automatically extract the spatial information related to precipitation from satellite cloud imageries,and then discriminates whether there is precipitation at a fixed location.Finally,the specific precipitation amount is estimated or forecasted at the location identified as precipitation.In addition,considering the influence of topography on the spatial distribution and intensity of precipitation,the proposed method introduces terrain and location elements in the stage of precipitation estimation or forecasting to further improve the accuracy of precipitation estimation and forecasting.The experimental results show that the proposed method can achieve better results when compared with other precipitation products in performance evaluation.