The rapid development of deep learning has fundamentally transformed the way urban features are extracted from remotely sensed imagery. Among the most widely adopted architectures for semantic segmentation tasks in geospatial analysis are U-Net and DeepLabV3+, each offering distinct approaches to pixel-level classification. This paper presents a comparative theoretical and methodological assessment of both architectures in the context of urban land cover feature extraction, with specific reference to their implementation within ArcGIS Pro — a leading commercial GIS platform that natively supports deep learning workflows. The analysis focuses on architectural design principles, segmentation performance metrics reported in the literature, and the practical feasibility of deploying each model in an urban GIS environment such as the city of Plovdiv, Bulgaria. Results from existing studies indicate that both architectures achieve competitive accuracy on urban datasets, with U-Net demonstrating strengths in boundary-sensitive tasks and DeepLabV3+ excelling in multi-scale contextual feature capture. The study discusses the implications for GIS practitioners seeking to integrate deep learning into operational workflows using ArcGIS Pro tools.