Accurate extraction of building footprints from unmanned aerial vehicle (UAV) imagery
is essential for urban planning, infrastructure monitoring, and disaster risk management.
However, segmentation performance remains challenged by scale variability,
spectral ambiguity, and partial occlusion in heterogeneous environments. Here, we
examine the integration of dilated convolution, residual learning, and spatial–channel
attention within a unified U-Net framework for building footprint extraction. Rather
than introducing new architectural components, the proposed approach investigates
how these established mechanisms influence segmentation behaviour across diverse
landscape conditions. Multi-rate dilated convolution is incorporated in the bottleneck
stage; residual connections support stable feature propagation, and attention modules
enhance feature discrimination. The model was evaluated on the WHU Building Dataset
and a newly developed UAV dataset (HUMG) representing diverse environmental
conditions in Vietnam. Results showed improved segmentation performance relative
to U-Net, Feature Pyramid Network, and SegFormer, particularly in dense urban, rural,
and vegetation-occluded environments. Although the proposed model incurs higher
computational costs, the additional complexity reflects both increased model capacity
and integrated architectural design. The findings demonstrate that systematic adaptation
of established mechanisms can enhance building segmentation robustness in
high-resolution UAV imagery for offline geospatial applications.
