Automatic building extraction from high-resolution imagery remains constrained by limited training data and domain shifts across geographic regions and spatial resolutions. Although data augmentation is widely applied in semantic segmentation, its capacity to compensate for scarce labeled samples under varying domain conditions remains insufficiently quantified in remotely sensed data. Here, we present a controlled data-centric evaluation to quantify how explicit, label-preserving augmentation influences model generalization under varying domain shifts, rather than proposing a new augmentation algorithm. The experimental design integrates DeepLabV3+ (CNN) and SegFormer (transformer) architectures to assess whether augmentation effects persist across distinct feature-learning paradigms. Four scenarios are constructed, including two intra-domain settings, a resolution shift (0.3 m to 0.1 m), and a geographic shift across heterogeneous urban environments. Training subsets are progressively sampled from 20% to 100% to isolate the interaction between data volume and distributional variability. Geometric, radiometric, and occlusion-based transformations are evaluated individually and in combination. Under cross-domain and low-data regimes, augmentation substantially increases predictive performance. Combined transformations increase mIoU from 0.572 to 0.688 at 20% training data in the resolution shift scenario, while geometric augmentation improves mIoU from 0.444 to 0.533 under geographic transfer. Models trained on 20% augmented data exceed the performance of 100% non-augmented configurations under pronounced domain discrepancies, establishing an operational threshold of data efficiency. Computational analysis indicates negligible overhead (approximately 1 s per epoch) through asynchronous data pipelines. Augmentation functions as a regularization mechanism in intra-domain settings and transitions to a distribution bridging mechanism under cross-domain conditions. Geometric invariance and engineered data diversity partially substitute for manual annotation, enabling improved cross-domain building extraction performance.
Pham et al. (Wed,) studied this question.