Semi-Supervised Classification of Collapsed Building Using Remote Sensing, In-Placed Sensors, and Fragility Functions

In this paper, we demonstrate the application of a new methodology for rapid identification of damaged areas to the case of the 2016 Kumamoto earthquake sequence. The method aims to integrate Synthetic Aperture Radar (SAR) imagery, fragility functions, and intense ground motion maps to calibrate a neural network changes detector in SAR images. All the necessary data was collected within a remarkably short period, maintaining relevance to real-world applications. Our findings indicate that minimal labeled data, such as from just six collapsed buildings, is sufficient for optimal neural network calibration. Upon validation with third-party data, this approach achieved an 84% accuracy rate, suggesting its potential as an effective tool during the crucial post-earthquake response phase, as it can provide decision-makers with crucial information to efficiently organize and allocate resources, such as sending aid, food, and rescue squads to the most severely affected areas.