Heavy Precipitation Weather Recognition Based On Deep Learning

Severe convective weather has the characteristics of sudden strong,small space scale,short life history,fast moving speed,violent weather process and easy to cause casualties and property losses.Under the current observation conditions,the characteristics of heavy precipitation weather are difficult to be captured by the conventional meteorological observation network,and the traditional identification technology of heavy precipitation weather lacks the ability to perceive the temporal and spatial variability in the process of extreme weather.Since there are many features of extreme weather and the occurrence probability of extreme weather is very low,it also brings great challenges to the learning of neural network.The research on extreme weather recognition based on deep learning is still in its infancy.In this paper,the intelligent recognition model of heavy precipitation based on single-station Doppler weather radar is studied,the feature selection of heavy precipitation model is realized,the precipitation recognition model based on deep learning is constructed,and the solution to the long tail distribution of data is proposed.The main results of this work are as follows:1.Based on the basic theory of information gain and Chi-square test,a method is proposed to quickly select a set of the most effective feature combination from multiple radar features,which provides a basis for feature selection for researchers2.Based on the characteristics of radar data,RRED-NET precipitation recognition model is designed and implemented.The spatial attention module is used in the hop module,which improves the fusion effect of down-sampling feature and up-sampling feature.Using self-attention module in feature transformation module endows features with more transformation ability and enhances the fitting ability of the model.In the input module,wavelet transform is used for data denoising to remove the redundant information and part of radar data,which helps the model to learn the distribution of radar features more effectively.Experimental results show that the proposed RREDNet model is superior to the traditional business methods.3.In view of the recognition task of precipitation data of long tail distribution problem,design a variety of loss function,adopted the approach of "block" data enhancement,learning strategies and use double stage,to enhance the network model of the tail interval sample identification effect,the experimental results validate the proposed method to alleviate the long tail distribution affect the performance of the model,to improve the high precipitation index of part.