Research On Doppler Radar Map Estimation And Forecasting Based On Variational Autoencoder

Doppler Radar is the main tool for human to study precipitation system,and plays an important role in the observation on whether.Radar Echo Extrapolation includes a series of algorithms that extract information from the current radar echo maps and use them in the prediction of the future radar echo maps.Traditional Radar Echo Extrapolation includes many algorithms,such as cell centroid tracking algorithm,cross-correlation algorithm and optical flow algorithm.These traditional algorithms can well model the translation of the radar echo maps,but it is difficult for them to capture the transformation like dissipation and enhancement.Deep Learning is a rapidly developing branch of machine learning since the 21 st century.Deep Learning methods have made outstanding contributions to image processing and sequence processing.The research object of radar echo extrapolation task are series of successive radar echo maps,which means that Deep Learning has the potential to be applied to Radar Echo Extrapolation task.In fact,previous researchers have tried to apply Convolutional Long Short Term Memory model(CLSTM)to radar echo extrapolation task,and achieved the state-of-the-art results.However,CLSTM still has its drawbacks.The radar echo maps generated by CLSTM has obvious ambiguity effect,and the radar maps generated in the further future become much more blurry.In this paper,the characteristics of CLSTM are analyzed in detail,and assumptions are made on its defects.This paper considers that the main reason why CLSTM is ineffective in radar echo extrapolation is that it is easy to fall into the local optimum point.In order to predict the result of the next time step,the model needs to carry out a large-scale search in the feature space.Based on the above analysis,this thesis considers to introduce Variational Autoencoder(VAE)to radar echo maps extrapolation.Variational Autoencoder is a generation algorithm.First,it generates a new representation of the data points through a recognition model.This representation is called latent variable.Thereafter,the model generates new data points according to the given latent variable.The characteristics of VAE make the distance of latent variables corresponding to similar data points close.In the same radar echo sequence,two consecutive radar echo maps are usually very similar.If VAE is used in radar echo maps,then the distance of their corresponding latent variables is also similar.Therefore,it is easy to predict the latent variable of the next time step by using the latent variables of the current time step,and there is no need for a large-scale searching,thus our method reduces the risk of falling into a local optimal point.Basing on the analysis above,a novel extrapolation method is introduced.To describe our method,firstly,the latent variables of the radar echo maps are generated by the recognition model.Second,given these latent variables,LSTM is used to predict the latent variables corresponding to the future radar echo maps.Last,given the latent variables corresponding to the future radar echo maps,the generation model is used to generate the prediction of the future radar echo maps.The experiments show that the radar echo maps generated by our method is less effective than CLSTM in prediction of the near future,but better than CLSTM in prediction of the far future.