Active Learning for Improved Damage Detection and Disaster Response

Abstract

From the 2020 Western Wildfires to the 2010 Haiti earthquake, each year natural disasters cost the world thousands of lives lost, trillions of dollars in damage, and irreparable long- term harm to the communities they impact. And because of climate change, the frequency of natural disasters is only projected to increase. In the wake of a natural disaster, conduct- ing fast and accurate damage detection is critical in order for governments to make declara- tions of emergency, mobilize resources, and receive funding. But current damage detection processes are dangerous, costly, and slow, involving manual inspection either on the ground or manually labeling buildings in satellite imagery. Artificial intelligence and computer vision are promising tools to expedite this damage detection process and alleviate human burden. However, existing work in this area does not adequately address two key challenges endemic to disaster response. First, existing damage detection models, which are predominantly fully- supervised convolutional neural nets, do not account for the lack of labeled data that typi- cally exists when a disaster strikes – it is impossible and impractical to label 80%-90% of the buildings in a new disaster that is unfolding. We therefore conduct experiments on gener- alizability, and propose active learning for fine-tuning to reduce the amount of labeled data needed by our damage detection model to only 25%, or less depending on the task. Second, natural disasters disproportionately impact more socially and economically vulnerable com- munities. We must consider this while developing artificial intelligence and computer vision tools, and therefore use the social vulnerability index in evaluating the performance and fair- ness of our models, constructing 12 novel datasets combining imagery and social vulnerabil- ity index in the process. We find that existing methods do in fact lead to disparities amongst groups of different socioeconomic status, and provide suggestions for future work to address this.