Research On Quantitative Precipitation Estimation Algorithms Based On Sequential Regression Of Spatio-temporal Data

The mining and analysis of spatio-temporal data is an important research problem,which has a wide range of application scenarios in many fields.For example,the meteorological field contains a large number of multi-source heterogeneous spatio-temporal data.At present,most of the problems in this field are still solved through traditional methods and statistical learning methods,lack the ability to model these data essentially,which makes it difficult for the degree of intelligence of weather services to adapt to large-scale data accumulation.Therefore,it is very important and full of challenges to design suitable algorithms to learn the relationship between temporal and spatial data in the meteorological field and solve practical problems.Aiming at the problem of quantitative precipitation estimation in the field of meteorology,this dissertation establishes a variety of sequential regression algorithms based on spatio-temporal data,and carries out research on quantitative precipitation estimation algorithms.Based on two types of spatio-temporal data,radar echo data and automatic rain gauge data,this dissertation studies and establishes a quantitative precipitation estimation algorithm based on sequential regression.At present,traditional methods and machine learning methods are mainly used for quantitative precipitation estimation.There are many shortcomings,especially the estimation of heavy rainfall.Therefore,this dissertation proposes a multitask recurrent convolutional network quantitative precipitation estimation algorithm.Designed a scaling recurrent convolutional neural network unit,extracting multi-scale original image information,fully and effectively considering the context characteristics of spatio-temporal data.We cascade the unit based on the sample construction strategy to model the complex relationship between the regional sequential characteristics of radar echo data and the accumulated rainfall at the central point.The introduction of classification task to assist regression prediction to improve the accuracy of the regression of heavy rainfall.In the regression and classification tasks,different loss functions are also studied and compared.A large number of experiments have been carried out on the processes of heavy rainfall in Guangdong Province,and the results show that this method is better than the baseline methods,especially in the estimation of heavy rainfall.Further analysis found that while the multitask recurrent convolutional network quantitative precipitation estimation algorithm improves the effect of heavy rainfall estimation,there will be a problem of high estimation errors of some medium and low grades,which will reduce the global index.The method of Z-R relationship has relatively small estimation errors in the middle and low levels.In order to combine the advantages of the method of Z-R relationship and the recurrent convolutional neural network,this dissertation proposes a quantitative precipitation estimation algorithm based on Wide & Deep framework.A joint training function is designed to learn the parameters of the two parts simultaneously.A large number of experiments have been carried out on the processes of heavy rainfall in Guangdong Province,and the results show that the algorithm not only maintains a good effect in the estimation of heavy rainfall,but also has been further improved in the global estimation,which proves the effectiveness of Wide & Deep joint learning and provides another way of thinking for quantitative precipitation estimation.