Research On Precipitation Nowcasting In Guangxi Based On Radar Echo Extrapolation

Guangxi is located in low latitude,and meteorological disasters such as rainstorm and typhoon are common.Therefore,carrying out precipitation nowcasting work in Guangxi has important practical significance.At present,the important means of precipitation nowcasting is the extrapolation of radar echo maps.How to improve the accuracy of radar echo extrapolation prediction is currently a major challenge in the field of meteorological research.In this study,the optical flow,deep learning and their dynamical blending algorithm are used to extrapolate the radar echo.The research results are expected to further improve the radar echo extrapolation prediction effect.The research work of this article mainly includes the following aspects:(1)Research on precipitation nowcasting in Guangxi based on sub-pixel with the pyramid Lucas-Kanade optical flow technique(SPLK-CCI).Firstly,the LK optical flow method and pyramid technology are used to track the movement of radar echoes,then the linear extrapolation method is used for extrapolation prediction,and finally the extrapolation prediction map of radar echoes is modified by the cubic convolution interpolation method.Based on SPLK,this study changed the redistribution method to cubic convolution interpolation method,and proposed SPLK-CCI,which effectively improved the image edge blur phenomenon caused by the bilinear interpolation method in the original SPLK.The critical success index(CSI),probability of detection(POD)and false alarm rate(FAR)were used to compare the prediction results of the sub-pixel pyramid based optical flow method with the pixel pyramid based optical flow method.The POD and CSI increased by 39.5% and 22.9% respectively,and the FAR decreased by22.7%.(2)A study on precipitation nowcasting in Guangxi based on SmaAt-UNet-Ⅰneural network.Due to the widespread issue of data imbalance in experimental data,rainfall data accounts for a relatively small proportion of the experimental data.In the experiment of this chapter,based on the SmaAt-UNet model,data was filtered,and training was conducted only when the proportion of radar echo reflectivity exceeding 25 in the first 10 images was greater than 5%.The SmaAt-UNet-Ⅰmodel was proposed.Comparing the forecast results of SmaAt-UNet-Ⅰwith the original SmaAt-UNet,POD and CSI increased by 27.3% and 23.0% respectively,while FAR decreased by 25.1%.(3)A research on precipitation nowcasting in Guangxi based on the blend of sub-pixel with the pyramid Lucas-Kanade optical flow method(SPLK-CCI)and SmaAt-UNet-Ⅰ.Due to the fact that both SPLK-CCI and SmaAt-UNet-Ⅰuse a single method for experiments,there are limitations: SPLK-CCI performs poorly in predicting the evolution of radar echoes,while SmaAt-UNet-Ⅰcannot accurately grasp the details of radar echo evolution during the extrapolation process.The blend method not only considers the accuracy of SPLK-CCI in the location of precipitation prediction systems,but also considers the ability of SmaAt-UNet-Ⅰ to cope with changes in precipitation systems,avoiding the limitations of a single method.This chapter adopts a dynamical weight blending algorithm to dynamically fuse the results of SPLK-CCI and SmaAt-UNet-Ⅰ.Comparing the prediction results of the blending algorithm with SPLK-CCI and SmaAt-UNet-Ⅰ,POD increased by 1.8% and 25.6%,CSI increased by27.2% and 16.1%,and FAR decreased by 64.3% and 8.9%,respectively.