The process of assessing the condition of destroyed buildings using aero photographs acquired from unmanned aerial vehicles (UAVs) has been investigated in this study. The task addressed relates to the fact that post-war monitoring of territories is complicated by the volume of satellite and UAV images, which exceeds the capabilities of expert inspection while the lack of uniform interpretation tools and domain shift reduce the reproducibility of assessment. The proposed two-stage neural-network method combines segmentation of buildings and determining the level of building destruction on a four-level damage scale: "absent", "minor", "significant", "destroyed". The "xView2" corpus was used as the source material, supplemented with authentic marked UAV images. The “You Only Look Once” (YOLO) segmentation models, versions v8s and v11n, were used to extract building contours, and the Vision Transformer (ViT) was used to categorize damage. Experiments were performed in Google Colab (USA) applying PyTorch (USA) and Ultralytics (UK). The Mean Average Precision (mAP) was calculated for segmentation models. The mAP indicators remain acceptable even in complex urban settings. For the ViT classification model, the Precision, Recall, and F1 values above 0.9 were obtained. The values achieved are attributed to the combination of a two-stage architecture and sample balancing. The devised method is applicable to satellite and UAV images; unlike existing solutions, it retains stability under domain shift. The resulting models could be implemented as a basic module in geographic information systems and decision support systems, enabling practical use of this study's results. For correct operation, sufficient resolution and representativeness of the training sample are required
Mazurets et al. (Fri,) studied this question.