Research On Deep Regression Of Radar Data For Precipitation Nowcasting

Precipitation nowcasting can be defined to predict rainfall distribution in the specified area within the short-term(generally 0-6 hours).Although there is various observation equipment to monitor rainfall,considering the demand for temporal and spatial resolution in the precipitation nowcasting field,this dissertation applies Doppler weather radar data as the core data.By predicting the radar echo map sequence,the intensity of rainfall can be predicted quantitatively and accurately according to the relationship between the radar echo reflectivity and rainfall intensity.However,with the generation of a large number of historical radar echo image data,how to learn the regular motion pattern from these massive data and predict future radar echo maps is the core problem of this dissertation.To explore how to precipitation nowcasting,this article studies from the perspectives of global unbiased and local biased prediction respectively.Global unbiased prediction refers to predicting the whole future map in terms of echo shape,position,intensity,and movement speed.Local biased prediction focus on predicting specific areas based on global unbiased prediction.In global unbiased prediction,the existing algorithms mainly face two problems: rainfall intensity decay under steady fluid movement and the degraded prediction ability of the model under fast fluid movement.Furthermore,extended to the other rainfall processes with different fluid movement patterns,how to solve the performance degradation of the model caused by the significant difference in spatial distribution is also needed to be studied in this dissertation;In local biased prediction,the spatiotemporal representation containing high echo reflectivity gradually be lost.This dissertation focuses on four issues to discuss precipitation nowcasting.Specifically,the main contents and novelty are summarized as follows:(1)Considering the problem of rainfall intensity attenuation under steady fluid movement,a multi-scale convolution generative adversarial network(Multi-Scale WGAN)is proposed.The model introduces the adversarial regularizations in the loss function,which alleviates the problem of radar echo reflectivity attenuation by closing the data distribution between the prediction and future images.Moreover,experimental results show that applying the adversarial regularization in other models can enhance the accuracy of prediction.From the visualization of predictions,the problem of prediction intensity attenuation has been alleviated.(2)As for the problem of the degraded prediction ability of the model under fast fluid movement,recurrent convolution neural networks based on pseudo optical flow and reconstruction are proposed(PFST-LSTM and RST-LSTM).The pseudo optical flow and global reconstruction sub-module are introduced into the two models respectively to align the positions of previous echo representation and the current so that the convolution neural network can concentrate on learning the changes of echo representation.Besides,the local reconstruction submodule is also introduced to adaptively adjust the size of the convolution field of view and improve the efficiency of capturing spatial features.The experimental result shows that the introduction of these sub-modules can effectively improve the accuracy of prediction and proposed methods are better existing models.Explainable experiments prove that the pseudo optical flow and global reconstruction submodule can indeed adjust the position of historical representation.(3)To address the problem that significant spatial distribution differences in different rainfall processes,an online incremental learning framework(CAM-OIF)based on gradient activating mapping is proposed.In this framework,the model can adaptively learn different motion patterns from different rainfall events so that any model can achieve good performance in different scenarios.Different from the existing online incremental learning framework,the novelty of the proposed online incremental learning framework can be reflected in three points: unified learning mode,approximation training mode,and gradient activating mapping weight.By the above three improvements,models can efficiently learn the regular pattern of radar echo movement from the current rainfall event without delay,even in the rainfall events with distribution gaps.In experiments,the proposed online incremental learning framework makes the overall prediction of the model has been greatly improved.(4)To solve the problem that the spatiotemporal representation containing high echo reflective gradually loses,a recurrent convolution neural network based on the attention mechanism is proposed(Pred RANN).This method predicts the heavy rainfall area more accurately without reducing the accuracy of global unbiased prediction.In this method,temporal attention and layer attention mechanisms are introduced respectively.It preserves the long-term representation of high echo information with a relatively small proportion and reduces the loss of model spatial features layer by layer during delivery,respectively.Besides,the context interaction mechanism is also proposed and embedded in the model to improve the short-term spatiotemporal representation ability.Compared with the existing methods,the experiments show the proposed method has better performance,especially in the region of high echo.In summary,focusing on the precipitation nowcasting task,this paper proposes the four prediction models including Multi-Scale WGAN,PFST-LSTM,RST-LSTM,and Pred RANN,and a framework(CAM-OIF).Experiments were conducted on radar echo image data sets observed from Guangdong Province and Shenzhen city.The experimental results prove the effectiveness of the above methods.