Datensatz

Zusammenfassung

Typically, disaster response using remote sensing technology requires pre and post-disaster datasets. Unfortunately, pre-disaster data is not always available, and a lack thereof may lead to unreliable results in building damage predictions. Augmenting pre- and post-disaster data can artificially increase the training samples in machine learning (ML) models. In this regard, SAR simulators can provide synthetic data for augmentation purposes. Here, we introduce a general framework for applying SAR simulators to building damage estimation. We used a SAR simulator, RaySAR, and optimized surface parameters, reflection, and a 3D model of the Onagawa Nuclear Power Plant, where the 2011 Tohoku tsunami hit, to assess the reliability of the SAR simulator compared to an actual SAR image. In addition, we constructed 3D models of damaged buildings with multiple damage statuses of the European Macroseismic Scale (EMS-98) to evaluate the suitability of the simulator for representing earthquake and tsunami damage. We strive to improve the original simulator using a more capable rendering algorithm, path tracing, and incorporation of Graphics Processing Units (GPUs) to achieve higher realism in diffuse reflections and shorter computation times. The results are evaluated using similarity measures in the spatial and frequency domains, contrasting simulated with authentic SAR imagery. Preliminary results show that in simulated SAR imagery, damaged and undamaged buildings show distinguishable signatures. We believe that our simulator overcomes the limitations of the diffuse reflection model employed in RaySAR. Conclusively, the proposed framework seems promising for improving damage mapping in the context of future digital twin implementations.