Research On Weather Radar Echo Image Data Quality Control And Near Extrapolation

The heavy rainfall caused by severe convective weather is easy to cause waterlogging,landslides and other disasters,which would cause serious impacts and heavy losses to social and economic development and the safety of people’s lives and property.The weather radar uses the Doppler effect of the electromagnetic wave to detect the movement and change of atmospheric water vapor particles.It has a wide detection range,high spatial and temporal resolution,and high real-time,which is conducive to quickly and accurately identifying potential factors of severe convection disasters.How to use the current radar echo data to predict the near future echo changes timely and accurately,that is,the radar near extrapolation is one of the critical challenges faced by meteorological departments.In recent years,scholars in the field of artificial intelligence have been trying to use deep learning methods to carry out radar near extrapolation,and have made some research achievements.However,there are still many problems to be solved in radar sample quality,extrapolation accuracy and timeliness.To facilitate the fusion of artificial intelligence technology and meteorological deeply,aiming at the poor quality of weather radar echo image data and the insufficient accuracy of the existing methods of echo near extrapolation,this paper uses deep learning methods to research weather radar echo image data quality control and near extrapolation methods.The specific research work is as follows:(1)Aiming at the data quality problems such as beam blocking and clutter that often occur in the weather radar images directly obtained by weather radar due to terrains,flying objects,air density and other factors,this paper proposes a weather radar beam blocking correction model based on convolutional encoding decoding model,which realizes the intelligent correction of the common small range beam blockages in weather radar data and the quality improvement of radar data.The experiment shows that the model can recognize the location of the blockages and correct the missing data effectively.At the same time,the model can also realize intelligent correction of beam clutter data on the premise that beam clutter is regarded as beam blocking.The corrected results are better than other competitive deep models in hit rate,correlation and other evaluation indicators.(2)Aiming at the problem that the accuracy of weather radar echo extrapolation is insufficient,based on the control of radar data quality,this paper focuses on the deep fusion of deep learning model and weather radar extrapolation task.A weather radar extrapolation model based on convolution and the multi-head self-attention mechanism is proposed to achieve more accurate extrapolation of time series radar image data.This model takes the encoder-decoder structure as the basic architecture,organically combines convolution and deconvolution with the multi-head self-attention mechanism,and constructs fully skipping connections between convolution and deconvolution.Experiments show that this method has better performance in long-term prediction accuracy,prediction image resolution and overall prediction accuracy than other compared deep learning radar extrapolation methods.(3)Aiming at the problem that the prediction accuracy of high-intensity echo area in weather radar echo extrapolation results is insufficient,and the variation over time has a large deviation from the real,this paper also proposes a weather radar extrapolation optimization model based on the generative adversarial network.The model improves the prediction accuracy of the radar extrapolation model based on convolution and multi-head self-attention mechanisms in high-intensity areas.In this model,the convolutional multi-head self-attention mechanism extrapolation model is used as the generator,and the multi-layer convolutional neural network is used as the discriminator.The generator is responsible for prediction,and the discriminator is responsible for providing the error of prediction results in high-intensity areas to add constraints to the generator loss.During training,the two parts are optimized for each other on their own goals,and finally optimize the extrapolation performance of the generator.Experiments show that the method can effectively improve the accuracy of the extrapolation model in the prediction of high-intensity areas.The above research concentrates on the deep fusion of deep learning methods with quality control of weather radar image data and radar near extrapolation tasks.The proposed methods and models effectively improve the radar echo image data quality and extrapolation accuracy.The research results verify the feasibility and effectiveness of the application of modern artificial intelligence technologies such as deep learning in weather radar data processing,offer theoretical and technical support for precipitation nowcasting of meteorological departments,and provide new methods and ideas for timely identification,prediction and early warning of severe convective weather disasters.