【作者】 李海峰；

【导师】 李旭涛；

【作者基本信息】 哈尔滨工业大学 ， 计算机技术（专业学位）， 2018， 硕士

【摘要】 雷暴大风是一种严重且常见的中小尺度气象灾害天气。雷暴大风天气一般由多个雷暴单体或超级单体构成的狭长状飑线引起,雷暴大风天气具有生成时间短、强度大、破坏力强、变化快、防御困难等特点,常对工农业的生产、人民的生活以及交通运输等行业造成严重的危害。相比于中长尺度的气象灾害现象,雷暴大风的识别和预测十分具有挑战性,气象领域的中外学者对此问题做了大量预报及预警方面的研究。利用机器学习和深度学习方法进行雷暴大风识别的研究,目前仍然处在前沿探索阶段,因此本课题的研究不仅具备较高的学术价值,同时拥有紧迫的现实意义。针对雷暴大风识别任务,本文基于深圳市气象局提供的广东省2015-2017年的雷达回波拼图数据、气象自动观测站数据,构建雷暴大风样本集,并进行雷达回波杂波滤除、噪声样本剔除等数据预处理。本文使用十种基于传统机器学习的雷暴大风识别方法,通过提取雷达组合反射率、垂直液态水含量(VIL)、雷达回波顶高等共十种雷达特征,对雷暴大风区域进行识别。为了减少雷暴大风的误识别率,本文设计基于形态学的飑线分割方法,仅对飑线区域的雷达回波进行雷暴大风识别,在提升识别速度的同时,有效提升雷暴大风的识别率。为充分利用雷暴大风识别问题的时空特性,本课题结合卷积神经网络和循环神经网络的优点,考虑雷达图像的时序因素和空间上下文的关系,设计并实现了四种雷暴大风识别网络模型。分别是简单卷积神经网络模型、空间上下文循环卷积神经网络模型、时间循环卷积神经网络模型,以及综合两种模型优点的时空循环卷积神经网络模型,设计实验对四种模型进行严格的实验对比。同时将深度学习模型,与使用手工提取特征的十种传统机器学习方法进行比较。实验结果表明:本文提出的方法在雷暴大风识别问题上具有较高的识别准确率,针对雷暴大风识别问题设计的网络模型,结合飑线分割方法,在5000个雷暴大风测试样本上的识别准确率达到83.2%,相比气象领域的传统方法有明显优势。本课题利用机器学习和深度学习方法,针对雷暴大风识别问题进行了系统的研究、实验和分析。通过搭建雷暴大风识别展示系统,使雷暴大风识别问题的研究流程化,本文搭建的系统将部署到深圳市气象局,供气象预报人员使用。更多还原

【Abstract】 Thunderstorm gale is a serious and common weather of small-medium scale meteorological disasters.Thunderstorm gale weather are generally caused by a long and narrow squall line composed of multiple thunderstorms or super-monomers.Thunderstorm gale weather have the characteristics of short generation time,high strength,strong destructive power,change fast,and difficult to defense.It causes badly harm to Production,people’s lives,transportation and other industries.Compared with the medium-long scale meteorological disaster phenomenon,the identification and prediction of thunderstorm gale is very challenging.Chinese and foreign scholars in the meteorological field have done a lot of research on forecasting and early warning.The research on thunderstorm and windy identification using machine learning and deep learning methods is still in the frontier exploration stage.Therefore,the research of this subject not only has high academic value,but also has urgent practical significance.For the thunderstorm gale identification task,this paper builds a thunderstorm gale sample set based on the radar echo puzzle data and meteorological automatic observatory data of Guangdong Province from 2015 to 2017 provided by Shenzhen Meteorological Bureau,and then performs data preprocessing such as radar echo clutter filtering,noise sample rejection.This paper proposes a thunderstorm gale identification method based on ten traditional machine learning methods.The thunderstorm gale region is identified by extracting a total of ten radar characteristics such as radar combined reflectivity,vertical liquid water content(VIL)and radar echo top height.In order to reduce the misrecognition rate of thunderstorm gale,this paper designs a morphological-based squall line segmentation method,which only uses thunderstorm gale identification for radar echoes in the squall line region,and enhances the recognition rate while effectively improving the recognition rate of thunderstorm gale.In order to make full use of the spatiotemporal characteristics of thunderstorm gale identification problem,this paper combines the advantages of convolutional neural network and recurrent neural network,considering the relationship between time series factors and spatial context of radar images,and designs and implements four thunderstorm gale identification network models.They are a simple convolutional neural network model,a spatial context circular convolutional neural network model,a time-circulating convolutional neural network model,and a spatio-temporal convolutional neural network model that combines the advantages of the two models.The four models are strictly controlled by design experiments.Experimental comparison.At the same time,the deep learning model is compared with ten traditional machine learning methods that use manual extraction features.The experimental results show that methods in this paper have high recognition accuracy in the thunderstorm gale identification problem.The network model designed for thunderstorm gale identification problem combined with the squall line segmentation method is accurate in the detection of thunderstorm gale 5000 test sets.The accuracy rate reached 83.2%.It has obvious advantages over traditional methods in the meteorological field.This paper uses machine learning and deep learning methods to systematically study,experiment and analyze the thunderstorm gale identification problem.Finally,a thunderstorm gale identification display system was set up to streamline the research on thunderstorm gale identification problems.The system in this paper will be deployed to the Shenzhen Meteorological Bureau for use by weather forecasters.更多还原