Research On Radar Echo Extrapolation Prediction With Integrated Geographic Element

In recent years,with the increasing global temperature,extreme rainfall events are also increasing,which seriously endanger people’s life and property.Therefore,it is important to carry out research on radar echo extrapolation prediction to reduce the risk of strong convective weather to human beings.Due to the successful application of deep learning techniques in various fields,researchers’ interest in using deep learning techniques for precipitation prediction has been growing.However,current deep learning methods usually consider only the historical combined reflectivity to predict future radar echo scenarios without taking into account other factors related to radar echo structure.In addition,most of the research on radar echo extrapolation prediction has been achieved by analyzing radar product data information,and the influence of surface feature data on radar echo prediction is relatively understudied.Therefore,it is of great research significance and application value to combine surface feature data and multi-source radar product data for the study of radar echo extrapolation prediction.In this study,a method of combining radar echo features with a multi-source ConvLSTM(MSConvLSTM)model is proposed in order to capture multivariate correlations and spatialtemporal shifts in precipitation patterns.This study extends the original structure of the ConvLSTM by incorporating a 3D convolutional auxiliary mechanism that combines radar echo features.This auxiliary mechanism combines the features of both auxiliary and main channels,thus enhancing interdependence and improving predictive capability.In addition,this auxiliary mechanism combines long-term channel information with short-term spatio-temporal information to generate enhanced features for the input sequence.In addition to this,the surface feature dataset is also used as auxiliary data in this study,which is input to the model along with the radar echo data for training and prediction.These surface feature data have an impact on the prediction of radar echoes,so their integration into the model can improve the prediction accuracy of the model.Finally,this study conducted a multi-factor stepwise sensitivity analysis and a single-factor sensitivity analysis using the MS-ConvLSTM model for a variety of radar product data and surface feature data.The specific conclusions are as follows.(1)To improve the accuracy of proximity weather forecasting,this study proposes a new algorithm,MS-ConvLSTM,which introduces a multi-source data input training module based on ConvLSTM,by extending the original structure and adding a 3D convolutional auxiliary mechanism incorporating radar echo features.In this auxiliary mechanism,the composite reflectance is the main feature,and the echo top,vertically integrated liquid water content data and radar inversion wind field data are used as auxiliary features.The design of this auxiliary mechanism combines the features of both auxiliary and primary channels,thus enhancing the interdependence and improving the prediction capability.It also combines long-term channel information and short-term spatio-temporal information to generate enhanced features for the input sequence.In this way,the combined reflectance is used as the main channel data,and the echo top height,vertically integrated liquid water content and wind field data are used as the auxiliary channel information,by which we can improve the fitting capability and improve the prediction accuracy.(2)In this paper,we firstly carried out a study on radar echo extrapolation prediction based on radar data based on MS-ConvLSTM model and combined with a variety of radar product data.Subsequently,a study of radar echo extrapolation prediction with integrated surface features was carried out based on the MS-ConvLSTM model and a multidimensional spatial radar echo feature dataset with integrated surface feature data.The experimental results show that the radar echo extrapolation prediction using multiple radar product data and integrated ground surface feature data performs better on the radar echoes of aggregated distribution.However,the method is difficult to predict in the case of real radar echoes with large ranges and discrete distributions.In contrast,radar echo extrapolation prediction using multidimensional spatial radar echo feature data with integrated surface features performs better.The accuracy of radar echo extrapolation prediction by integrating ground surface features reached 0.804,while the accuracy based on multiple radar product data was only 0.743.In addition,the false alarm rate was reduced to 0.193 for the false alarm case based on multiple radar product data,while the false alarm rate for radar echo extrapolation prediction by integrating ground surface features was 0.172 By organically combining the radar product data with the surface feature data,the method can more accurately simulate the real 3D spatial environment,thus greatly improving the prediction accuracy.(3)This paper analyzes the sensitivity of the model based on radar data and ground surface features,aiming to investigate the influence of different feature factors on the radar echo extrapolation prediction results.Through single-factor analysis,this paper evaluates the effects of input variables such as combined reflectance,echo top height,vertically integrated liquid water content,radar inversion wind field,digital elevation model,normalized vegetation index and land use data on the radar echo extrapolation prediction results.In the experiments,radar product data and surface feature data are introduced in this paper,and their effects on the prediction results are observed.The experimental results show that various radar product data and ground surface feature data have some influence on the prediction results.By combining ET and CR,we can obtain better prediction results of radar echo extrapolation with the probability of correct identification as high as 0.784,which indicates that ET has an important influence on the occurrence of radar echoes.In addition,the combined use of DEM and CR can also obtain a high hit rate of 0.734,which indicates that the terrain undulation factor also has an influence on the extrapolation prediction of radar echoes.Vertical liquid water content data have a relatively small effect on the prediction results.In addition,the surface feature data also have some influence on the prediction results.In particular,the digital elevation model has a greater influence on the prediction results,followed by the normalized vegetation index and land use.This indicates that the surface feature data is also one of the important influencing factors of the prediction results.These results are important for improving the accuracy and precision of radar echo extrapolation prediction.