Extending VHR Optical/SAR Building Damage Assessment Benchmarks with Multi-Resolution TerraSAR-X Data: Impact of Sensor Variety on Model Robustness

Kurzfassung

Synthetic aperture radar (SAR) satellite sensors can provide essential all-weather, day-and-night capabilities for rapid building damage assessment (BDA) in time-critical natural disaster response scenarios. Besides the sensor-inherent challenges, limited availability of very high resolution (VHR) SAR datasets for training suitable machine learning models, however, has hampered wide adoption in rapid mapping operations so far. Recent datasets based on bi-temporal VHR satellite data address the aspect of multi-modal BDA (e.g. BRIGHT or DisasterM3) by pairing pre-event optical imagery with post-event SAR acquisitions. While these datasets offer a rich geographic variety, their associated benchmark results revealed strong limitations in transferability to unseen regions and disaster events, e.g. due to inconsistent damage patterns in the SAR backscatter signatures, especially in the case of zero-shot cross-event scenarios. Furthermore, their multi-sensor, multi-resolution generalization capabilities beyond the included X-band sensors (Capella, Umbra) remain unclear. In this study, we investigate an extended multi-modal (optical-SAR) VHR BDA dataset that expands the existing BRIGHT dataset with TerraSAR-X/TanDEM-X (TSX/TDX) data for ten additional disaster events to increase both sensor and geographic variability. We examine the effectiveness of transfer learning and domain adaptation techniques and present preliminary benchmarking results. Besides analyzing the transferability to another VHR X-band SAR sensor and extending the geographic variety of publicly available benchmark datasets by additional globally distributed events, this work further examines the limitations of transferability in multi-resolution settings, including data from different TSX/TDX imaging modes like StripMap (up to 3m) and (High Resolution) SpotLight (up to 1m).