Urban Infrastructure Segmentation Using Drone Images

Abstract

Efficient urban planning and infrastructure management require accurate, up-to-date maps of urban environments. Drone-based aerial imaging has emerged as a cost-effective and flexible method for large-scale urban data collection. This paper proposes a semantic segmentation framework for urban infrastructure analysis using high-resolution drone imagery. The proposed system employs a DeepLab v3+ architecture with multi-scale atrous convolutions and an attention gate mechanism to segment urban features including roads, buildings, vegetation, water bodies, and bare land. The model achieves a mean Intersection over Union (mIoU) of 84.7% across all categories on the ISPRS Vaihingen dataset, outperforming baseline FCN and UNet models. The system supports smart city applications, disaster response, and infrastructure monitoring through automated semantic mapping from drone imagery