【摘要】 为了在突发自然灾害发生后快速、准确地对灾情损失进行量化评估，克服传统人工调查与传感器监测速度慢、成本高等局限性，以2019年美国加利福尼亚州Ridgecrest地震为例，探讨基于网络社交媒体众包数据对突发自然灾害进行灾情快速评估的可行性；基于时空匹配、多轮关键词筛选及虚假描述剔除技术，从144万条原始推文中清洗并提取25 708条有效灾情推文；结合现有MMI地震灾害评价体系，建立适用于网络社交媒体数据的地震损失4级评价标准；构建、比选7种基于文本向量化和文本分类的地震损失评估模型，对25 708条有效推文进行分析，并与美国地质调查局(USGS)的实际灾后调查结果进行对比验证。结果研究表明：推文数量、推文长度、单词长度、推文反映的损失内容等均与地震的发展显著相关；利用推文反映地震损失的时效性为12～14 h;随着地震发展，损失级别整体呈缓慢线性增长(R~2=0.57);对地震损失进行量化评估，评定等级为1.47,为轻微至中等，与事后USGS的实际调查结果基本吻合；建立的关键词筛选+文本数字化+分类模型框架，验证了利用社交媒体大数据快速量化评估突发自然灾害的可行性，可为网络大数据分析在交通基础韧性、突发自然灾害管理方面的应用提供参考。更多还原

【Abstract】 To enable rapid and accurate quantitative assessment of disaster damage following sudden natural disasters and overcome limitations of traditional field surveys and sensor monitoring, such as slow speed and high cost, this study examined the feasibility of rapidly assessing sudden natural disasters using crowdsourced data from online social media, taking the 2019 Ridgecrest Earthquake in California, USA, as a case study. It cleansed and extracted 25 708 valid disaster-related tweets from 1.44 million raw tweets using spatiotemporal matching, multi-round keyword filtering, and removal of false descriptions. Integrating the existing modified mercalli intensity(MMI) earthquake damage evaluation system, the study established a four-level earthquake damage assessment standard suitable for online social media data. Seven text vectorization and text classification-based earthquake damage assessment models were constructed and compared. These models analyzed the 25 708 valid tweets and were validated against the actual post-disaster survey results from the United States Geological Survey(USGS). The results indicate that tweet volume, tweet length, word length, and the loss content reflected in tweets are all significantly correlated with earthquake development. The timeliness of tweets in reflecting earthquake losses is approximately 12-14 hours. As the earthquake develops, the loss level shows an overall slow linear increase(R~2=0.57); quantitative assessment of earthquake losses yields a rating of 1.47, classified as “slight to moderate”, which aligns well with the post-event actual survey results from USGS. The established “keyword filtering + text vectorization + classification model” framework validates the feasibility of rapidly quantifying sudden natural disasters using social media big data and provides a reference for applying network big data analysis in transportation infrastructure resilience and sudden natural disaster management. 4 tabs, 11 figs, 38 refs.更多还原