AI for Good

Chapter 14Post-Disaster Building Damage Assessment

—Shahrzad Gholami

Executive Summary

Climate change is accelerating the frequency and severity of natural disasters. Each year, natural disasters affect over 350 million people and cause trillions of dollars in damage. In the wake of those disasters, providing timely and appropriate emergency responses and humanitarian interventions like shelter, medical aid, and food can be challenging. Artificial intelligence frameworks that leverage available satellite imagery can support existing efforts to provide such interventions. Here, we develop a convolutional neural network model that uses high-resolution satellite imagery from before and after disasters to localize buildings and score their damage into four levels, ranging from not damaged to destroyed.

Due to the emergency nature of disaster response efforts, the value of automating damage assessment lies primarily in its speed, rather than its accuracy. When we compared our results to those generated during an international competition, our solution works three times faster than the fastest winning solution and over 50 times faster than the slowest first place solution. Our model achieved a pixel-wise F1 score of 0.74 for the building localization and a pixel-wise harmonic F1 score of 0.60 for damage classification—both indicators that our model was accurate. Further, our model used a simpler architecture compared to other efforts in the competition. To facilitate the use of ...