| 利用深度学习填补双偏振雷达回波遮挡 |
| |
| 引用本文： | 尹晓燕,胡志群,郑佳锋,左园园,皇甫江,朱永杰.利用深度学习填补双偏振雷达回波遮挡[J].应用气象学报,2022,33(5):581-593. |
| |
| 作者姓名： | 尹晓燕 胡志群 郑佳锋 左园园 皇甫江 朱永杰 |
| |
| 作者单位： | 1.中国气象科学研究院灾害天气国家重点实验室, 北京 100081 |
| |
| 摘 要： | 广州S波段双偏振天气雷达低仰角多方位存在遮挡，高仰角也存在部分遮挡。基于卷积神经网络等深度学习方法，构建垂直填补（vertical echo-filling，VEF）和水平填补（horizontal echo-filling，HEF）网络架构，基于两种架构，利用无遮挡区的反射率因子ZH、差分反射率ZDR，差传播相移率KDP构建训练集，填补遮挡区的ZH和ZDR。针对仅0.5°仰角存在遮挡的区域，基于VEF架构，利用上层多个仰角、径向、距离库的三维数据，分距离段训练垂直填补模型。针对遮挡仰角较高的区域，则基于HEF架构，利用同一仰角左右相邻的多个径向、距离库的数据，分遮挡径向训练水平填补模型。根据解释方差、平均绝对偏差和相关系数3个指标和3个个例，对模型效果进行评估。结果表明:ZH填补模型的解释方差最大为0.92，平均绝对偏差最小为1.69 dB，相关系数最高为0.96；ZDR填补模型的解释方差最大为0.92，平均绝对偏差最小为0.12 dB，相关系数最高为0.96。利用该研究构建的深度学习填补架构，可有效填补偏振雷达遮挡区域回波，提高雷达数据质量。
|
| 关 键 词： | 深度学习 双偏振雷达 波束遮挡 回波填补架构 |
| 收稿时间： | 2022-03-06 |
Filling in the Dual Polarization Radar Echo Occlusion Based on Deep Learning |
| |
| Affiliation: | 1.State Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, Beijing 1000812.School of Atmospheric Sciences, Chengdu University of Information Technology, Chengdu 6102253.Key Laboratory of Atmosphere Sounding, China Meteorological Administration, Chengdu 610225 |
| |
| Abstract: | Radar beam blockage is an important error source that affects the quality of weather radar data. The S-band dual-polarization radar in Guangzhou has multi-azimuth occlusion at low elevation and is partially occluded at high elevation. Based on deep learning methods such as convolutional neural network, two echo filling networks, i.e., VEF(vertical echo-filling) and HEF(horizontal echo-filling) are constructed. Based on this architecture, echoes from the unblocked area are used to construct training datasets and fill the reflectivity ZH and differential reflectivity ZDR in the occlusion area. For the area with only 0.5° elevation occlusion, multi-modal modeling is carried out based on VEF architecture by using 3D data from multiple upper elevations, radial directions and gates. Considering that the radar beam broadens with distance and to avoid the influence of the melting layer, the radar beam is divided into four sections according to the oblique distance of 0.5° elevation, and the vertical echo-filling model is trained respectively. For the area with high occlusion elevation, multi-mode modeling is carried out based on HEF architecture using the data of multiple adjacent radial directions and gates with the same elevation. According to the number of occlusion radial, two types of horizontal echo-filling models, three radials echo-filling model and five radials echo-filling model are constructed respectively. Finally, the models are evaluated by three cases and three indicators:Explained variance, mean absolute error and correlation coefficient. The maximum explained variance of ZH vertical echo-filling model is 0.91, the minimum mean absolute error is 1.72 dB, and the maximum correlation coefficient is 0.96. The maximum explained variance of ZDR vertical echo-filling model is 0.87, the minimum mean absolute error is 0.12 dB, and the maximum correlation coefficient is 0.92. The maximum explained variance of ZH horizontal fill model is 0.92, the minimum mean absolute error is 1.69 dB, and the maximum correlation coefficient is 0.96. The maximum explained variance of ZDR horizontal echo-filling model is 0.92, the minimum mean absolute error is 0.12 dB, and the maximum correlation coefficient is 0.96. The deep learning echo-filling model can be used to correct the echoes of Guangzhou S-band dual-polarization radar occlusion area effectively, and the quality of weather radar data is improved. |
| |
| Keywords: | |
|
| 点击此处可从《应用气象学报》浏览原始摘要信息 |
|
点击此处可从《应用气象学报》下载全文 |
|
