Abstract
Building detection from high-resolution aerial images is a key step in urban planning, disaster management and 3D city modeling. This article extends our earlier study of U-Net-based building segmentation on 20 cm RGB orthophotos of Athens, Greece. Besides a U-Net trained from scratch and a U-Net with a pretrained ResNet-34 encoder, we train bagged ensembles of 20 members for three architectures, including a U-Net++ with a ResNet-50 encoder. All models are trained on the same selected city area and evaluated on a held-out building block from another city suburb, for which ground truth was produced. Inference combines eight-fold test-time augmentation with a reconstruction that keeps only the central region of each patch, stitches soft probability maps instead of binary masks, averages them across members, and applies morphological processing to produce the final reconstructed block. The block is processed in two ways: first, in its original orientation and, second, aligned with the orientation of the training area buildings through appropriate rotation. The best configuration, a bagged U-Net on the rotated block, reaches a block-level IoU of 89.80% and an F1-score of 94.63%. The same bagged model compares favorably to all other single models after improvement achieved through morphological post-processing.