Short-Term Rainfall Forecast Based On Doppler Weather Radar Images

Rainfall factors often induce meteorological disasters,which will have adverse effects on production and life.In particular,the severe convective weather such as short-term heavy rainfall has great destructive power due to its rapid evolution and high rainfall intensity,which seriously threatens the safety of human life and property.Therefore,meticulous short-term rainfall prediction is of great practical significance to prevent meteorological disasters.At the same time,with the development of meteorological detection technology,Doppler weather radar is widely deployed,which can provide a large number of high spatial and temporal resolution data.Using meteorological radar data for short-term rainfall prediction has become a very important topic.At present,the main prediction method is based on the optical flow method for echo extrapolation,but the accuracy still does not satisfy the requirements of meteorological operational.Therefore,this thesis introduces the deep learning method to study the two types of problems of radar echo extrapolation and rainfall prediction in multiple regression regions,and constructs the prediction models respectively to improve the accuracy of short-term rainfall prediction.The specific work is as follows :(1)Aiming at the problem of short-term rainfall prediction based on radar echo extrapolation,the traditional optical flow method is used to realize the radar echo extrapolation model.Although the optical flow method can retain the texture features of the radar image,the constraints of the optical flow method are strict and the radar data cannot be fully utilized.In this thesis,the deep learning network model Unet is used to construct the radar echo extrapolation model SA-Unet.The test is carried out on the collected and preprocessed radar data set.The experimental results show that SA-Unet is superior to the optical flow method in meteorological indicators such as hit rate and critical success index,and the extrapolated radar image is closer to the real radar image,which proves the effectiveness of deep learning method in echo extrapolation.This thesis further discusses the influence of different loss functions on the model extrapolation.Experiments show that the effect of using multi-scale structural similarity index as the loss function is better than the average absolute error,and it is more suitable for radar echo extrapolation.(2)The regional rainfall prediction of multiple regression is a more refined rainfall prediction problem.Since this problem belongs to the spatio-temporal prediction problem,a CNN-Bi LSTM rainfall prediction model with end-to-end output results is proposed by referring to the Long-term Recurrent Convolutional Networks(LRCN)and using the characteristics of long-term and short-term memory neural networks that are good at dealing with long-term sequence prediction,combining it with convolution neural networks and attention mechanism.This model first extracts the spatial information of each radar echo image in the sequence through the convolution neural network,and then uses the bidirectional long-short-term neural memory network to analyze the relationship in the time dimension of the radar echo sequence,and extracts the temporal and spatial characteristics to predict the rainfall.The model uses efficient channel attention mechanism in the convolutional neural network part,so that the model can adaptively assign weights to spatial features of different importance,thereby increasing the impact of useful information on the model.The experimental results show that the proposed model is superior to other models in terms of root mean square error,mean absolute error and correlation coefficient.