Mathematical Evidence-theoretic Framework for Information Fusion of Disaster Scene Big Data

Tuesday, 16 December 2014
ZhiQiang Chen, University of Missouri Kansas City, Kansas City, MO, United States
The remote sensing community and geospatial industries are embracing the paradigm of ‘big data’. This trend is in one hand due to the fact that heterogeneous remote sensors are producing tremendous amounts of earth observation (EO) data every day; on the other hand it is aspired by the promise that big data computing may be the fourth paradigm for scientific discovery. Many traditional techniques have been developed and will continue to be useful to deal with earth-observation big data, for examples, pansharpening, fusion of data with different electromagnetic nature (e.g. color images and SAR images), and use of multi-sensor data for improved land-cover classification. However, two limitations are recognized for these techniques, which include: (1) first, these methods are tightly dependent on a two-dimensional grid scale; and (2) second, temporal, spatial and causal relations are not intelligently treated. These limitations render them insufficient when used to attack the emerging Disaster Scene Big Data (DSBD).

DSBD emerges as a geological or climatic hazard unfolds into a disaster. In the example of an earthquake, disaster data starts accruing as the ground shaking is being monitored. Along the time scale, heterogeneous multi-sensor data arise: EO data with various electromagnetic nature, oblique images, airborne/terrestrial active (Lidar) data, and the recently emerged crowdsourcing data. Neither theoretical models nor effective methods exist to date that can sufficiently fuse these data towards revealing the ‘ground-truth’ of the disaster effects, for example, damage to built objects.

This presentation will present an augmented evidence-theoretic framework based on the classical Dempster–Shafer theory. With a focus on reasoning the ground-truth of build-object damage, causal, correlational and relational evidences will be defined considering their temporal and spatial scales. The newly developed graph-based learning approach will be explored for estimating the belief-plausibility interval of the ground-truth damage. Case-study using recent earthquake disaster data (e.g. the 2011 Christchurch Earthquake) will be discussed.