High-resolution imagery and neural networks link post-tsunami land cover changes to population health and well-being

As extreme events intensify in force and frequency across the globe, relating the damage and subsequent reconstruction to population health and well-being remains a critical frontier. Here we build a convolutional neural network to classify landcover from satellite images of Indonesia before and after the December 2004 Indian Ocean Tsunami and link those measures to population well-being to demonstrate methods that advance analyses of short-term impacts of extreme events and impacts 5 years later. Population data are from the Study of the Tsunami Aftermath and Recovery (STAR) and 2005 and 2010 censuses. We develop manually labelled training data for eight landcover classes and demonstrate the model performs well using standard metrics. Moreover, measures of change over time in landcover correlate strongly with multiple dimensions of well-being from our household survey data and with aggregate population statistics, both immediately after the event and in the subsequent five years.