Using High Resolution Imagery and Neural Networks to Measure Destruction and Reconstruction After a Disaster

Elizabeth Frankenberg , University of North Carolina at Chapel Hill
Keenan Karrigan, University of North Carolina at Chapel Hill
Peter Katz, Duke University
Eric Peshkin, Duke University
Cecep Sumantri, SurveyMETER
Duncan Thomas, Duke University

Climate change coupled with rising sea levels is intensifying the force and frequency of extreme events. Developing tools for measuring damage and pace of recovery and linking measurements to population data is a critical frontier. We combine high resolution satellite imagery, machine learning tools, and population data in the context of the 2004 Indian Ocean Tsunami to demonstrate proof-of-concept for techniques that will enhance research on the short and long-term impacts of extreme events. The Study of the Tsunami Aftermath and Recovery (STAR) provides population data for over ten years. Images are from Digital Globe (2004, 2005, 2007, 2009). We developed a training dataset containing over 9,000,000 manually labelled building pixels. Quality statistics (recall, precision, F1) demonstrate that our method is highly effective, even for chaotic imagery from 2005. Our method is also efficient, classifying 1.1 million pixels per minute versus ~200 pixels per minute by human coders.

See extended abstract

 Presented in Session 46. Innovative Data and Methods for Population and Environment Research