Efficient post-earthquake reconnaissance planning using adaptive batch-mode active learning

In the aftermath of earthquakes, obtaining a timely and accurate estimate of infrastructure damage is crucial for effective emergency response and recovery planning. Previous research has shown that active learning (AL) methods enable the inference of regional infrastructure damage from sparse on-site inspections through post-earthquake reconnaissance surveys. However, the efficiency of the surveys depends on two key factors: appropriate identification of candidate buildings for inspection and optimized routing between the sequential inspections. This study presents an AL framework that integrates learning and routing objectives, aiming to enhance the efficiency of post-earthquake damage data collection. The proposed approach employs an adaptive batch-mode AL method, using a Gaussian process regression model and an information-theoretic criterion to suggest a batch of candidate buildings for inspection. These buildings are then adaptively visited based on an optimized route schedule. The results demonstrate that the proposed framework reduces the required number of building inspections by up to 53%, decreases resource demands, and improves the predictive performance of the model compared with a baseline method.