UrbanNet: Weakly Supervised Deep Learning for Road and Building Segmentation in Urban Areas

Accurate multi-class segmentation of urban infrastructures is crucial, as urban environments present challenges such as occlusions, diverse architectural structures, and complex road conditions. This study presents UrbanNet, a novel FCN-based network designed for robust multi-class urban infrastructures segmentation. The network leverages multi-source data fusion of high-resolution imagery and weakly supervised labels derived from OpenStreetMap. A major limitation of standard CNNs is their reliance on large, high-quality annotated datasets; UrbanNet introduces an efficient training pipeline with dedicated strategies to handle weak labels and improve segmentation robustness. The proposed attention-enhanced architecture refines feature representations and mitigates noise in the labels. This integration offers a practical framework for urban infrastructures extraction from weak labels, supporting efficient and scalable urban analysis.