Weather Radar Echo Extrapolation Based On Convolutional Neural Networks

Radar echo extrapolation technology possesses a widely application prospects in short-term nowcast.Accurate and efficient extrapolated radar echos are crucial for improving the accuracy of short-term nowcast.Based on Doppler weather radar reflectivity factor data,this thesis applies Convolutional Neural Networks(CNNs)to weather radar echo extrapolation.Trained by a variety of sequential weather radar images,CNNs learn the regular of radar echos change,and predict the future radar image.The main contents of this thesis include:Firstly,the preprocessing methods of Doppler weather radar raw data is studied.The radar data in three-dimensional polar coordinates are converted into CAPPI(Constant Altitude Plan Position Indicator)data by coordinate transformation,interpolation,horizontal slice and so on.Then after cutting,compressing and sampling images,the Radar Echo Dataset is constructed,which provides Train-set and Test-set for the training and testing of CNNs respectively.Secondly,to adapt the strong correlation between the predicted radar image and input radar images,a new radar echo extrapolation method based on Dynamic Convolutional Neural Networks(DCNN)is proposed.Through a function that maps the convolution kernels to the input,convolution kernels of DCNN will vary from input to input during testing,which enhances the association between the input radar images and the future one,and improves the extrapolation accuracy.Finally,to adapt the strong correlation between the predicted radar images and the historical input image sequence,a new radar echo extrapolation method based on Recurrent Dynamic Convolutional Neural Networks(RDCNN)is further proposed.Learning from recurrent neural networks(RNNs),RDCNN adds cycle structures in convolution layers,which establishes the association between the predicted image and the input image sequence over a period of time,and enhances the ability of the network to process the timing-related images.The results of comparative experiments show that the methods based on CNNs have achieved higher accurate of extrapolation compared with traditional ones,which meets the application requirments of short-term nowcast.