Identifikation und Klassifizierung kriegsbedingter Gebäudeschäden unter Verwendung von terrestrischen Bildern und etablierten Modellen zur Objektdetektion und Segmentierung

Beyond the immediate human suffering, Russia’s full-scale invasion of Ukraine has already inflicted massive damage on civilian infrastructure. The scale and nature of this destruction create a pressing need for rapid damage assessment to support recovery planning. Current assessment workflows depend heavily on manual terrestrial inspection, making them slow, resource-intensive, and difficult for local authorities to scale across multiple settlements in a consistent way. Satellite imagery is widely perceived as a useful basis for monitoring war-related damage, but satellite-based approaches, while offering synoptic coverage and scaling well to large areas, lack the spatial granularity and contextual detail needed to capture on-the-ground damage conditions at the level of individual buildings. Herein, we present a deep-learning-based damage assessment pipeline that structures and automates the extraction of damage information from terrestrial imagery. It is intended to be interoperable with other sensor data, such as satellite imagery, by later extension into a more comprehensive solution for automated derivation of robust spatio-temporal evidence of destruction. At its core, the pipeline consists of a YOLOv8-based object detector and a Segment Anything Model (SAM) for automatic extraction of impacted structures from terrestrial imagery, and categorisation as damaged walls, damaged windows, and debris. We used a patch-based data processing strategy to mitigate background dominance. We fine-tuned the models on a manually annotated dataset derived from 380 images. The two-stage architecture first detects and classifies damaged regions, then performs pixel-level segmentation, providing accurate localisation of structural damage. Within our study, we achieved a mean average precision (mAP) of 0.65 for the object detection and a Dice coefficient of 0.91 for the segmentation. This demonstrates that the proposed pipeline is a useful first step towards supporting spatial damage assessment and recovery planning with automated provision of quantitative and qualitative information. Despite the promising results, the current implementation is limited by the relatively small dataset. More research is needed to develop the pipeline into a solution that generalises well to unseen environments and extracts the relevant information in formats and with semantics compatible with the requirements and legal constraints of public administration in Ukraine, or of other specific stakeholders.