Authors: ZhiQiang Chen*, University of Missouri-Kansas City, Shiming Tang, University of Missouri-Kansas City
Topics: Hazards, Risks, and Disasters, Geographic Information Science and Systems, Spatial Analysis & Modeling
Keywords: Disaster mechanics, Big Data, Crowdsourcing, Deep Learning
Session Type: Paper
Start / End Time: 8:00 AM / 9:40 AM
Room: Forum Room, Omni, West
Presentation File: No File Uploaded
The disaster scenes are extremely dynamic and complex as a function of terrain, built-up structures and infrastructure, human residence, commercial development, and the environment. To this end, the understanding of disaster scenes is directly visual and subjective, even with today’s digital technologies including smartphones, social networks, and cloud infrastructure. On the other hand, thanks to the digital technologies, crowdsourcing becomes much easier today as we witness that thousands of images from the crowd and the citizen scientists (less from the from the professionals, e.g. the journalists) come into sight available for the much broader public (through social networks). Many are designing a centralized approach to collecting these images towards more meaningful exploitation of these images. Among many challenges (such as the geospatial location errors, missing data, and non-structured data), an immediate one is to label or better semantically cluster these images into different categories. In this paper, we undertake a specific task, which is to learn the disaster-scene mechanics (e.g. the type of disaster and the degree of built-up object damage), by exploiting the latest deep learning technology. By taking advantage of transfer learning and a well-developed deep learning framework, we shed light that crowdsourced images can be rapidly classified with a significant accuracy. We further envision that this framework can be exploited if implemented in a high-performance edge-computing device (e.g. smartphones) or centralized cloud infrastructure can provide the crowd a real-time analytics power who may be more incentivized to either improve the crowdsourcing trustfulness and reduce the epistemic uncertainties.