Damage Criticality Assessment from Aerial Imagery Using a Heterogeneous CNN–Vision Transformer Ensemble: A UAV-Assisted Remote Sensing Approach for Post-Earthquake Emergency Response

Advances in remote sensing and unmanned aerial vehicle (UAV) technologies are enabling faster and more effective post-earthquake disaster assessment. Rapid identification of aerial images indicating critical damage is essential for directing emergency response teams to areas where urgent assistance, field inspection, and resource allocation are most needed. However, manually reviewing large volumes of heterogeneous aerial imagery is time-consuming and may delay emergency response operations. This study proposes a heterogeneous CNN–Vision Transformer Voting Ensemble (CTVE) for automated damage criticality assessment from high- and low-altitude aerial imagery. Eight CNN architectures and six vision transformer models are comparatively evaluated on four heterogeneous datasets that differ in altitude, viewing angle, scene characteristics, and damage appearance. Different ensemble combinations are evaluated across the datasets in terms of classification performance and computational time. Based on this evaluation, the selected CTVE combines ConvNeXtBase, DenseNet169, ResNet50, Xception, and DeiT through hard majority voting, achieving a mean accuracy of 94.82% and a mean macro F1-score of 91.44%, with equal weighting across datasets. These results support the effectiveness of the selected heterogeneous ensemble for damage criticality assessment under diverse aerial imaging conditions.