Empirical calibration of facility wildfire vulnerability index using post-fire damage records and neural additive models

초록

Wildfires increasingly threaten facilities in the wildland-urban interface. Existing vulnerability indices often rely on expert judgment and lack empirical calibration using actual damage data. This study optimizes the facility wildfire vulnerability index using artificial intelligence and empirical damage records. We analyzed a dataset of 5127 buildings from major South Korean wildfires using drone orthoimages. To address the severe class imbalance, we combined neural additive models with the effective number of samples method. The optimization revealed that physical building attributes and road conditions are more critical for structural survival than surrounding forest and suppression factors. Furthermore, we developed an automated assessment system that uses satellite imagery and computer vision to calculate vulnerability scores. This datadriven approach can reduce the need for repeated site visits, enabling a single village to be assessed within approximately 20 min.

키워드