南刚强,陈明轩,秦睿,韩雷,曹伟华. 2021. 基于雷达组合反射率拼图和深度学习的中尺度对流系统识别、追踪与分类方法[J]. 气象学报, 79(6):1002-1021, doi:10.11676/qxxb2021.062
基于雷达组合反射率拼图和深度学习的中尺度对流系统识别、追踪与分类方法
Identification, tracking and classification method of mesoscale convective system based on radar composite reflectivity mosaic and deep learning
投稿时间:2021-03-29  修订日期:2021-08-23
DOI:10.11676/qxxb2021.062
中文关键词:  深度学习  中尺度对流系统  识别  追踪  分类
英文关键词:Deep learning  Mesoscale convective system  Identifying  Tracking  Classification
基金项目:国家自然科学基金项目(41575050和41801022)、北京市科技计划课题(Z171100004417008)
作者单位E-mail
南刚强 中国海洋大学青岛266100
北京城市气象研究院北京100089 
 
陈明轩 北京城市气象研究院北京100089 mxchen@ium.cn 
秦睿 北京城市气象研究院北京100089  
韩雷 中国海洋大学青岛266100
北京城市气象研究院北京100089 
 
曹伟华 北京城市气象研究院北京100089  
摘要点击次数: 194
全文下载次数: 140
中文摘要:
      中尺度对流系统(Mesoscale Convective System,MCS)是很多对流性天气的主要致灾体,可导致严重的气象和水文灾害,如雷暴大风、冰雹、龙卷风和山洪。对MCS进行准确的识别和追踪,并根据追踪轨迹及获得的MCS特征实现MCS的分类,对灾害天气的分析和预报有重要意义。基于京津冀地区2010—2019年的雷达组合反射率拼图资料,分别使用支持向量机(SVM)、随机森林(RF)、极度梯度提升决策树(XGBoost)和深度神经网络(DNN)4种机器学习方法,研发了京津冀地区MCS的自动识别算法。使用时、空重叠追踪法对识别的MCS进行追踪匹配,得到包含强度、时间和空间信息的MCS追踪数据资料。在区分线状对流系统和非线状对流系统的基础上,进一步从经典的尾随层云(Trailing Stratiform,TS)、前导层云(Leading Stratiform,LS)和平行层云(Parallel Stratiform,PS)三类准线性MCS的概念模型和结构特征出发,根据追踪轨迹计算MCS的运动方向和MCS近似长轴两侧层状云和强对流云的面积占比,建立准线性MCS的分类算法。MCS的识别属于二分分类问题,以命中率(POD)、虚警率(FAR)、临界成功指数(CSI)和准确率(ACC)为评价指标,综合对比各项指标发现DNN模型较SVM、RF和XGBoost模型对MCS的识别效果更好。使用时、空重叠追踪法对DNN模型识别的MCS进行追踪,结合对两个追踪实例的分析,发现本研究所用的算法取得了很好的追踪结果,也进一步说明了深度学习方法识别MCS的准确性和优势。根据追踪轨迹计算某时刻MCS的运动方向,结合识别的层状云和强对流云的分布位置,准确实现了TS、LS和PS型准线性MCS的分类,为准线性MCS的生命史预测及其致灾天气特别是短时强降水的强度、位置和持续时间的客观预报提供了一种技术思路。
英文摘要:
      Mesoscale convective system (MCS) is the main cause of lots of convective weather, which can lead to severe meteorological and hydrological disasters such as thunderstorms, tornadoes and flash floods. Accurate identification and tracking of MCS and the realization of MCS classification based on the tracking trajectory as well as understanding the MCS features are of great importance for the analysis and forecast of catastrophic weather. Based on the radar composite reflectivity mosaic data in the Beijing-Tianjin-Hebei region from 2010 to 2019, the support vector machines (SVM), the random forest (RF), the Extreme Gradient Boosting (XGBoost) and the deep neural network (DNN) are used to develop an automatic recognition algorithm for MCSs in Beijing-Tianjin-Hebei region. Secondly, the tracking and matching of identified MCS slices are completed according to spatiotemporal overlap tracking, and a tracking database of MCS is established, which includes MCS intensity and spatial and temporal information. Finally, on the basis of distinction between linear convection and non-linear convection and starting from three conceptual models and structural characteristics of thee classical quasi-linear MCSs, i.e., the trailing, leading, and parallel stratiform precipition, an algorithm for quasi-linear MCS classification is established based on the area ratios of stratiform and intense convection on both sides of the approximate major axis of MCS and its movement direction, which is calculated according to tracking trajectory. The recognition of MCS is subsumed under binary classification. Taking POD, FAR, CSI and ACC as evaluation indexes, the DNN model is better than the SVM, RF and XGBoost models in MCS recognition after comprehensive comparison. Spatiotemporal overlap tracking is used to track MCS slices identified by the DNN model. The analyses of two tracking examples suggest that the algorithm used in this research has achieved good tracking results, which further demonstates the accuracy and advantage of deep learning in identifying MCS. The accurate realization of MCS classification including TS, LS and PS provides a technical idea for the life cycle prediction of quasi-linear MCS and objective prediction of disasterous weather, especially the intensity, location and duration of short-term heavy precipitation by combining the movement direction of MCS at a single radar snashot with the distributions of stratiform and intense convection in MCS slices.
HTML   查看全文   查看/发表评论  下载PDF阅读器
分享按钮