基于卷积光流生成网络的雷达回波图像预测算法研究  

Echo Image Prediction Algorithm Based on CNN Optical Flow Generation Network

作　　者：张毓杉 周晓[1] ZHANG Yushan;ZHOU Xiao(School of Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学机电工程学院,湖北武汉430070

出　　处：《数字制造科学》2022年第4期297-302,共6页

摘　　要：针对传统光流法预测雷达回波图像存在的问题进行了改进,提出了基于卷积光流生成网络的雷达回波图像预测算法。该算法通过向卷积光流生成网络输入连续的两帧雷达回波图像,生成对应的光流场,并结合光流场和原始图像以及半拉格朗日外推算法预测出未来十帧的雷达回波图像序列。在标准雷达回波图像数据集(standardized radar dataset,SRAD)的基础上制作了用于卷积光流生成网络训练的数据集,经过实验对比分析,所提出的算法在拥有较高的降雨预测精度下,生成光流场的计算时间仅耗时1.99 ms,而传统光流法需要耗时157 ms,该算法计算速度远远超过了传统光流法。In this paper,a radar echo image prediction algorithm based on CNN optical flow generation network is proposed,the traditional optical flow method for predicting radar echo images is improved.The algorithm generates the corresponding optical flow field by inputting two consecutive radar echo images to the CNN optical flow generation network.It combines the optical flow field and the original image to predict the radar echo image sequence of the next ten frames in a recursive manner.To fully extract and combine the deep and shallow features in the images,the CNN optical flow generation network adopts an encoder-decoder structure and a multi-scale optical flow field loss function for generating accurate optical flow field.Based on the SRAD data set,the FSRAD data set used for the network training of this paper is produced.After comparing and analyzing the experimental data,the developed algorithm has a high rainfall prediction accuracy.The calculation speed of the developed algorithm is dramatically improved if comparing with the traditional optical flow method(from 157 ms to 1.99 ms).

关 键 词：光流 雷达回波图像 预测 临近降水预报 编解码 

分 类 号：TN957.52[电子电信—信号与信息处理]