基于微博的强对流天气灾情信息提取与分析：以江苏2021年“4·30”强风雹天气为例

Disaster information extraction and analysis of severe convective weather based on Weibo: A case study of severe wind-hail event on April 30, 2021 in Jiangsu

摘要

【目的】强对流天气通常具有突发性和局地性。有效获取社交媒体中的灾情信息，能够弥补强对流天气实况观测密度不足的问题，为灾害应急管理提供信息支撑。【方法】对典型强对流天气数据文本进行中文分词和统计分析，融合气象专家知识，形成适用于强对流天气的灾害主题语料库。将语义信息融入主题模型(LDA)和支持向量机(SVM)分类算法，构建了强对流灾害天气灾情信息提取模型。以江苏2021年“4·30”强风雹天气为例，收集了16 334条原创微博文本信息进行仿真试验。【结果】结果显示：(1)构建的强对流灾情信息提取模型对微博文本中的灾情信息识别与分类效果显著。一次文本主题挖掘提取出天气状况、科普防御、灾情影响、求助救援和其他信息5个主题，对“灾情影响”信息进行二次分类，提取出公共设施、电力通信、车辆交通、农业设施、人员伤亡和其他6种具体灾情信息。经过交叉验证，一次分类平均准确率为92.70%,二次分类平均准确率为90.95%。(2)将强对流天气发生发展过程划分为预警期、突发期和灾后期3个阶段，各类信息均在突发期处于高值区。灾害突发期，公共设施、电力通信信息最多，灾后期，人员伤亡信息讨论度最高。(3)灾情信息数量的空间分布基本与灾害影响严重地区一致。强对流天气发生时，公共设施暴露风险较高，受损现象最普遍。【结论】基于微博的强对流灾情信息提取模型能够有效获取微博文本隐含的灾情信息，反映灾害事件的变化特征和舆论焦点，对灾害监测预警服务和应急决策指挥具有一定参考意义。

Abstract

[Objective] Severe convective weather is typically characterized by abrupt onset and localized impact. Effectively acquiring disaster information from social media can compensate for the insufficient observation density of severe convective weather and provide information support for disaster emergency management. [Methods] Chinese text segmentation and statistical analysis were performed on text data of typical severe convective weather. By integrating meteorological expert knowledge, a disaster-themed corpus tailored to severe convective weather was developed. Semantic information was incorporated into the latent Dirichlet allocation(LDA) topic model and the support vector machine(SVM) classification algorithm to construct a disaster information extraction model under severe convective weather. Taking the severe wind-hail event in Jiangsu on April 30, 2021 as an example, 16 334 original Weibo text messages were collected for simulation experiments. [Results] (1) The constructed disaster information extraction model for severe convective weather demonstrated remarkable effectiveness in identifying and classifying disaster information in Weibo texts. Through primary topic mining, five themes were extracted: weather conditions, public education on disaster prevention, disaster impact, rescue requests, and other information. The secondary classification was performed on “disaster impact” to extract six specific categories: public facilities, power and communication, vehicle traffic, agricultural facilities, casualties, and others. Cross-validation revealed an average accuracy of 92.70% for the primary classification and 90.95% for the secondary classification.(2) The development process of severe convective weather was divided into three stages: warning stage, outbreak stage, and post-disaster stage. All information categories peaked during the outbreak stage. During the outbreak stage, information on public facilities and power and communication was the most prevalent, while during the post-disaster stage, discussions about casualties were the most frequent.(3) The spatial distribution of disaster information quantity was generally consistent with the regions severely affected by the disaster. During severe convective weather events, public facilities faced a higher risk of exposure with damage being the most common. [Conclusion] The extraction model of disaster information based on Weibo for severe convective weather can effectively extract implicit disaster information in Weibo texts, reflect the variation characteristics of disaster events and the focus of public opinion, and provide valuable reference for disaster monitoring and early warning services as well as emergency response command.

关键词

强对流 / 灾情 / 社交媒体 / 微博 / 信息提取 / 主题模型 / 文本分类 / 降雨

Key words

severe convective weather / disaster information / social media / Weibo / information extraction / topic model / text classification / rainfall

基金资助

国家自然科学基金项目(42171081)

中国气象局创新发展专项(CXFZ2024J012)

江苏省气象局面上项目(KM202205)