Abstract
Following a disaster, responders need to rapidly assess the extent of the damage. The prevailing view is that very-high-resolution optical satellite images can provide more accurate estimates of building damage than lower-resolution synthetic aperture radar (SAR). However, we demonstrate that SAR-based damage proxy maps we produced after the 2023 Kahramanmaraş Türkiye earthquakes outperformed comparable maps derived from optical imagery. This finding held true for both a state-of-the art machine learning method and for visual interpretation. The SAR-based maps achieved an F1 performance score approximately twice as high as the optical-based maps (0.47 vs 0.24, 0.23 and 0.15). Additionally, SAR offers established advantages in both coverage and timeliness: SAR was able to image the entire affected area within ten days, whereas the very-high-resolution optical dataset covered only 5.4% of the SAR-covered area during the same timeframe. Using the largest ground dataset that we know of for any event, we show that the damage distribution captured by our SAR-based maps strongly correlates with the ground observations over a wide range of spatial scales from neighbourhoods (~1 km2) to provinces (~10,000 km2). Based on this experience, we argue that any reliable remote sensing-based damage assessment system should incorporate radar to complement other techniques.
Keywords
earthquake, geophysics, natural hazard