A Crowdsourced Knowledge Base for the Damage Assessment of Extreme Events

使用众包志愿者分析遥感图像是一种相对较新的损失评估方法，在2008年四川地震后发展起来，并在2010年海地和2011年新西兰地震期间正式确定。Web 2.0技术的出现和随处可见的免费遥感图像使这种方法成为可能，这些图像是具有高空间、光谱和时间分辨率的天气图。被证明的好处是在提供损失估计方面加速了两到三倍。然而，这种人工众包的损失评估方法的成功与否取决于人群的规模和可靠性。这项研究将集中在一个新的框架，称为Backbone(建立极端事件的众包知识库)，用于极端事件损害评估，利用遥感图像自动发现和分类损害，并建立一个数据驱动的损害特征知识库，可以在未来的事件中重复使用。Backbone是对人工众包方法的转变，在灾难社区是前所未有的，它将众包的力量与计算机科学和图像处理的最新方法结合在一起。它用基于对象的更改检测和分类的自动化方法取代了手动工作，这些方法提高了速度，降低了损失评估的成本，并能很好地根据数据量的增加进行扩展。它将人群从手动注释的任务转移到通过众包主动学习对自动化方法的性能进行质量保证反馈。这提高了评估的准确性，同时将框架的成功与人群的规模和可靠性脱钩，因为从注释者那里征求反馈得分较高，而且仅在困难的案例中。它还结合了多种遥感产品，并进行数据融合，将它们的输出统一到一个共同的损害地图中。随着数据量和产品的激增，这是下一代损害评估的必备特征。使用不同的数据产品，特别是来自高空间分辨率和对天气和太阳光照不太敏感的不可见波段的图像，将更好地区分某些破坏类型。这项研究的更广泛影响是，通过提供用于灾后需求评估(PDNA)的快速和准确损失估计的新方法，降低了极端事件的总体人力和经济成本。知识库的管理在灾难来袭时迅速建立有效的模型，完善评估脆弱性的事件模拟中的损害预测，并促进更好的土地利用规划，以鼓励抗灾社区的发展。这项工作还包括一个基于网络的损失评估模拟器，它绘制了最近地震事件的遥感地震图像，以使更多的公众参与减灾。此外，调查人员将积极从代表性不足的群体中招募研究生和本科生，并在多学科合作中对他们进行指导。机器学习的学生将学习遥感和损害评估，地球工程的学生将学习机器学习和统计数据分析的基础知识。

The use of crowdsourced volunteers to analyze remote sensing imagery is a relatively new damage assessment approach, developed in the wake of the 2008 Sichuan earthquake, and formalized during the 2010 Haiti and 2011 New Zealand earthquakes. This approach is enabled by the advent of Web 2.0 technologies and the ubiquity of free remote sensing images that are synoptic with high spatial-, spectral- and temporal-resolutions. The demonstrated benefit was a speedup by a factor of two or three in the delivery of damage estimates. However, the success of this manual crowdsourced approach for damage assessment is dependent upon the size and reliability of the crowd. This research will focus on a new framework called BACKBOnE (Building A Crowdsourced Knowledge Base of Extreme Events) for extreme event damage assessment utilizing remotely sensed images that automatically finds and classifies damages, and builds a data-driven knowledge base of damage characteristics that can be reused during future events. BACKBOnE is a transformation of the manual crowdsourced approach for damage assessment and is unprecedented in the disasters community, combining the power of crowdsourcing with state-of-the-art methods from computer science and image processing. It replaces the manual effort with automated methods for object-based change detection and classification that increase the speed, reduce the cost of damage assessment, and scale well to increases in data volume. It shifts the crowd from its task of manual annotation to quality assurance feedback on the performance of automated methods via crowdsourced active learning. This improves assessment accuracy, while decoupling the framework's success from the size and reliability of the crowd because feedback is solicited from annotators scored favorably, and only on difficult cases. It also incorporates a multitude of remote sensing products and performs data fusion to unify their outputs into a common map of damage. This is a must-have characteristic of next-generation damage assessment as data volumes and products proliferate. The use of diverse data products, particularly imagery from high spatial resolutions and non-visible bands that are less sensitive to weather and solar illumination, will better discriminate certain damage types.The broader impact of the research is the reduction of the overall human and financial cost of extreme events by contributing new methods for rapid and accurate damage estimates used for Post-Disaster Needs Assessment (PDNA). The curation of a knowledge base builds effective models quickly when a disaster strikes, refines damage predictions in event simulations that assess vulnerability, and fosters better land-use planning that encourages the growth of disaster resilient communities. The work also includes a web-based damage assessment simulator that maps remotely sensed earthquake images from recent earthquake events to engage the greater public in disaster mitigation. In addition, the investigators will actively recruit graduate and undergraduate students from under-represented groups and mentor them within a multi-disciplinary collaboration. Machine learning students will learn about remote sensing and damage assessment, and geoengineering students will learn fundamentals of machine learning and statistical data analysis.