皇甫江,胡志群,郑佳锋,朱永杰,尹晓燕,左园园. 2022. 利用深度学习开展偏振雷达定量降水估测研究[J]. 气象学报, (0):-, doi:[doi]
利用深度学习开展偏振雷达定量降水估测研究
Research on quantitative precipitation estimation using polarized radar using deep learning
投稿时间:2022-02-28  修订日期:2022-03-28
DOI:
中文关键词:  深度学习,偏振雷达,定量降水估测,单参量架构,三参量架构
英文关键词:Deep learning, Polarimetric radar, Quantitative Precipitation Estimation, One-moment network, Three-moment network
基金项目:国家重点研发计划(2019YFC1510304),广东省重点领域研发计划(2020B1111200001),中国气象局大气探测重点开放实验室开放课题(U2021Z05),青年科学基金项目(42105141)。
作者单位邮编
皇甫江 成都信息工程大学大气科学学院 610225
胡志群 中国气象科学研究院灾害天气国家重点实验室 100081
郑佳锋 成都信息工程大学 610225
朱永杰 成都信息工程大学 610225
尹晓燕 成都信息工程大学 610225
左园园 成都信息工程大学 610225
摘要点击次数: 43
全文下载次数: 30
中文摘要:
      文章利用2018—2020年偏振升级改造后的广州S波段双偏振雷达(CINRAD/SAD)82,892个体扫的0.5°仰角数据,以及雷达100km探测范围内1,109个雨量站共计538,560分钟雨量数据,分别构建了单参量,三参量雷达定量降水估测(QPE)深度学习网络架构(Z-Rnet, KDP-Rnet, Pol-Rnet),并以K_DP = 0.5 °/km为阈值,分别训练得到大雨、小雨、总体等9个QPE模型。在常用的均方误差作为损失函数的基础上,本文对不同降水强度采用不同权重,提出了自定义损失函数,并利用比率偏差、相对偏差、均方差、平均绝对误差和平均相对误差作为评价指标对模型进行评估。[结果与结论]通过对以积层混合云为主、以对流云为主和以层状云为主的三次降水过程的模型验证结果表明,利用深度学习训练的模型有较好的QPE效果,区分雨强的小雨、大雨模型比不区分雨强的总体模型的效果要好。采用自定义损失函数模型效果更好,其均方差、平均绝对误差和平均相对误差分别较采用传统均方误差损失函数,提升了8.62%、12.52%、16.34%。自定义损失函数中,采用ZH/ZDR/KDP三参量网络架构训练得到的QPE模型效果最好,其均方差、平均绝对误差和平均相对误差分别较采用ZH的单参量Z-Rnet架构提升了6.82%、8.43%、7.22%;较采用KDP的单参量KDP-Rnet架构提升了12.33%、17.61%、17.26%。
英文摘要:
      By means of 82,892 volume scanning data in 0.5° elevation of Guangzhou S-band dual polarization radar (CINRAD/SAD), and 538,560 minutes rainfall data from 1,109 stations within the radar"s 100km detection range from 2018 to 2020, three deep learning networks, that is, Z-Rnet, KDP-Rnet, and Pol-Rnet, are designed for radar quantitative precipitation estimation (QPE) based on single and three radar moments, respectively. Furthermore, based on the three networks and with KDP = 0.5 °/km as the threshold to divide the training dataset as heavy, light, and all rain data, a total 9 QPE models are built. On the basis of the common mean square error as the loss function, a self-defined loss function is proposed by adjusting the weights for different precipitation intensity. And then, indicators which include ratio deviation, relative deviation, mean square error (MSE), mean absolute error (MAE), and mean relative error (MRE) are used to evaluate the models, respectively. Finally, there precipitation processes, which are dominated by cumulus-stratiform mixed, by convective, and by stratiform clouds, are used to test the effect of QPE, respectively. the results suggest that: the models fitted by deep learning have better QPE effect, and QPE accuracy of dataset divided into heavy and light rain is better than that of all data used. The MSE, MAE, and MRE with the self-defined loss function are improved by 8.62%、12.52%、16.34% than that with the traditional mean square error loss function. Among them, the QPE with Pol-Rnet, i.e., ZH, ZDR, and KDP, are as input factors, is the best, and the above indicators are increased by 6.82%、8.43%、7.22% that with Z-Rnet, and by 12.33%、17.61%、17.26% than that with KDP-Rnet.
查看全文   查看/发表评论  下载PDF阅读器
分享按钮