Assessing the transferability of post-disaster building damage assessment using synthetic aperture radar and machine learning

A timely understanding of the scale and distribution of damage to infrastructure such as buildings and dwellings following disasters is crucial for efficient and effective response and recovery. Synthetic Aperture Radar (SAR) is a valuable tool for remotely assessing damage, and when coupled with Machine Learning (ML) has the potential to safely and effectively map damage at an individual building level. For practical disaster response, SAR ML models must generalise to new disaster events beyond their training data. Model transferability to new disaster events in differing contexts has been untested with medium resolution SAR. Therefore, this study evaluates transferability of medium resolution SAR ML models using Sentinel-1 coherence and intensity for mapping damage across 11 disaster events of various types (including earthquakes, hurricanes, wildfires, tsunamis and volcanic eruptions) across different global locations. Models were trained and tested under three scenarios: a reference scenario requiring no transferability, and two transferability scenarios utilising event-based cross-validation: (a) a global scenario and (b) a fire-specific scenario. Model performance decreased substantially under the global transferability scenario compared to the reference scenario, indicating the models struggled to adapt to varying disaster characteristics and geographic contexts. The fire specific scenario outperformed the disaster-agnostic global scenario, suggesting that further +research may benefit from focusing on disaster-type specific models. Although performance across all scenarios was inconsistent, several disaster events saw positive performance, particularly in the fire-specific scenario. Further research is necessary before models should be operationally utilised for disaster response.