Deep learning-based road crack segmentation can achieve high benchmark performance on public datasets, but its reliability in UAV-acquired field imagery remains uncertain because real road scenes contain various crack-like non-crack features. This study assessed the field transferability of a public-dataset-trained crack segmentation model using UAV-acquired road images. Nine semantic segmentation model configurations were trained and compared using integrated public crack datasets, and U-Net++ with a ResNet50 backbone was selected for UAV field application. After confidence-based data refinement and training data augmentation, the final model achieved precision, recall, Dice, and IoU values of 0.802, 0.787, 0.794, and 0.659, respectively, on the public validation dataset. Field evaluation was conducted on 148 UAV images and manually corrected UAV ground-truth patches generated from non-overlapping 8 × 8 image grids. The total UAV field evaluation produced precision, recall, Dice, and IoU values of 0.856, 0.806, 0.830, and 0.710, respectively. Major errors were associated with crack-like non-crack factors, including lane markings, manholes, repaired boundaries, shadows, surface contamination, and reflections. Threshold sensitivity analysis indicated that approximately 0.4–0.5 was a practical operating range under the tested UAV conditions. These results show that idealized public-dataset evaluation can overestimate real-world UAV performance and that field transferability evaluation is essential for reliable UAV-based pavement inspection.