Assessing the social impacts of extreme weather events using social media

自然灾害对英国经济造成重大破坏，但其影响很难预测和观察。该项目将开发利用社交媒体数据准确绘制自然灾害及其影响地图的方法。“社会感知”可以定义为使用来自数字通信（例如手机通话记录、社交媒体、网络搜索和其他在线数据）的未经请求的内容来观察现实世界的事件。社会感知的挑战是从大量非结构化、不完整和可能不准确的用户话语中提取高质量的观察结果。如果能够做到这一点，那么就有很大的机会在目前无法获得观测的领域使用这些数据。一个关键的应用领域是评估自然灾害的影响，这方面的预测通常是基于气象模型，但是用于模型验证的影响观测数据很少。调查人员之前的工作表明，社交媒体可以成功地用作检测和定位野火、洪水和极端降雨事件的数据来源。然而，这些方法还处于发展的早期阶段。本项目将为自然灾害社会感知管道的两个关键方面建立稳健的方法：(1)内容分类。该项目的第一部分将使用机器学习来创建基于文本的分类器，该分类器可以根据内容自动对社交媒体帖子进行分类。(2)位置推断。识别社交媒体内容的地理来源对于危害影响评估至关重要，但只有一小部分社交媒体帖子包含准确的地理标签。该项目的第二部分将开发有效的机器学习方法来推断无地理标记的社交媒体帖子的位置。

Natural hazards cause major disruption to the UK economy, but their impacts are hard to forecast and observe. This project will develop methods for using social media data to accurately map natural hazards and their impacts. "Social sensing" can be defined as observation of real-world events using unsolicited content from digital communications (e.g. mobile phone call records, social media, web searches, and other online data). The challenge of social sensing is to extract high-quality observations from large numbers of unstructured, patchy, and possibly inaccurate user utterances. If this can be achieved then there are significant opportunities to use this data in areas where observations are not currently available.One key application area is assessing the impact of natural hazards, where forecasts are routinely produced based on meteorological models, but for which little impact observation data is available for model validation.Previous work by the investigators has shown that social media can be used successfully as a source of data with which to detect and locate wildfires, floods and extreme rainfall events. However, the methods are at an early stage of development. This project will establish robust methods for two key aspects of the social sensing pipeline for natural hazards:(1) Content classification. The first part of the project will use machine learning to create text-based classifiers that can automatically categorise social media posts based on their content.(2) Location inference. Identifying the geographical origin of social media content is essential for hazard impact evaluations, but only a small fraction of social media posts include accurate geotags.The second part of the project will develop effective machine learning methods for inferring the locations of un-geotagged social media posts.