EAGER: An Exploratory Study of Multi-Hazard Management through Multi-Source Integration of Physical and Social Sensors

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

Natural and man-made disasters can cause significant material damages and human suffering. For example, Superstorm Sandy of 2012 is estimated to have caused more than $68 billion in damages and killed at least 286 people in seven countries. Improving the preparation for, response to, and recovery from disasters can reduce damages, relieve human suffering, and speed up recovery. Among disasters, a multi-hazard is a sequence of disasters in which the first disaster causes the subsequent disasters, making it far more difficult for emergency response teams to handle all of them. For example, the March 11, 2011, Tohoku, Japan, earthquake triggered an unprecedented tsunami, which led to flooding at, and partial meltdown of, the Fukushima Daiichi Nuclear Power Plant. A more frequent example of multi-hazards is landslides, which can be triggered by many causes including earthquakes, rainfall, and man-made environmental changes. While the detection of a single disaster usually only requires one kind of dedicated sensor, for example, seismographs can detect earthquakes reliably, multi-hazards often require a combination of various kinds of sensors for the detection of the multiple events in the sequence. Indeed, the detection of multi-events in general and multi-hazards in particular is a non-trivial problem due to the various kinds of events involved and the large number of combinations that make offline combinatorial analysis impractical. In the case of landslides, their detection is complicated further by the several possible and unrelated causes of landslides (e.g., earthquake and rainfall), each requiring a different kind of sensor. In this project, the team is building a landslide detection system, called LITMUS, that integrates data from two physical sensors -- USGS Global Seismographic Network (GSN), NASA Tropical Rainfall Monitoring Mission (TRMM) -- with data from pervasive social media platforms. This integration of multiple heterogeneous sensors in LITMUS is an illustrative example of successfully applying big data software tools and analytics techniques to solve real-world problems. Specifically, the team is extending geo-tagging to relevant data items, which are filtered in several stages to reduce noise and false positives, and applying machine learning, information retrieval, and semantic web techniques to each data stream. Finally, filtered social media data are being cross-referenced with physical events from the same geo-location to generate supporting evidence for landslide detection. A LITMUS prototype has been detecting more landslides around the world than traditional landslide reporting systems: tests with live streaming data show that the combined result is a list of landslide events that has included the USGS authoritative list, plus many other confirmed landslides around the world.