EAGER: A Dynamic, Reliability-Weighted, Multi-Pass Probabilistic Framework to Reduce Uncertainty in Crowd-Sourced Post-Disaster Damage Assessments

Stanford University, Stanford CA

Investigators

Abstract

A good estimate of the scale and spatial distribution of damage after a disaster is critical for effective response and recovery. It provides data to make best use of emergency responders, distribute medical, food or other aid, manage debris removal, and advocate for and leverage international aid based on a clear understanding of losses. Modern information technologies provide unprecedented data access to support damage assessment following a disaster, but these data are often very hard to assimilate, gauge its quality and coalesce into effective decision-support tools. In the past decade, crowdsourced and citizen-based analysis of satellite and aerial imagery have become additional mechanisms to analyze data from rapidly evolving disasters due to earthquakes, hurricanes and other hazards. Compared to standard data-gathering methods, these crowdsourcing initiatives have the advantage that the data obtained are typically much more numerous, have much larger spatial coverage and are obtained much more rapidly, but often have significant uncertainty. This has been a significant limitation to the widespread use of crowdsourced data and citizen science interpretation following disaster. This EArly-concept Grant for Exploratory Research (EAGER) project focuses on methodological improvements to crowdsourced post-disaster damage assessment as well as recovery monitoring. Several important impacts are expected, including the identification, measurement and systematic reduction in uncertainty in crowd-sourced damage assessments and the development of a probabilistic framework for quantifying and mapping post-disaster damage. The research marks a methodological paradigm shift in crowdsourced analysis of images, from the direct assignment of deterministic categorical damage states to the definition of full probability distributions of damage. This shift towards a probabilistic framework further enables the disaggregation of uncertainty into its various sources, each of which can be minimized individually. This is achieved by targeting improvements in the three components of the crowdsourced damage assessment process: set-up, imagery analysis, post-processing. Set-up: improving the crowdsourcing task definition and developing a standard visual damage taxonomy. Imagery analysis: improving the identification of damage indicators (e.g. less omissions and miss-categorizations) through multi-pass assessment. Post-processing: developing a weighted aggregation of identified damage indicators into probability distributions of damage. Results from this research are expected to support the development of effective decision-support tools used by emergency responders and disaster recovery planners, promoting more resilient communities in the United States and beyond.