The rapid advancement of advanced air mobility (AAM) and urban air mobility (UAM) has increased the demand for the precise localization of drones in complex urban environments, where global navigation satellite systems (GNSS) sensors often suffer from degradation due to buildings, interference, and multipath effects. Vision-based localization using aerial imagery has therefore emerged as a promising alternative, relying on the reliable segmentation of quasi-static objects such as buildings and roads. In this study, we address the challenge of achieving both high segmentation accuracy and real-time performance under resource-constrained drone environments. We propose a coverage-aware, data-centric training strategy that systematically organizes training datasets on the basis of the proportion of building and road pixels. In addition, we develop lightweight segmentation models by adapting backbone networks and model scales of YOLO11-seg and DeepLabV3+. Extensive experiments demonstrate that midrange-coverage datasets (20–40%) yield the most stable and balanced performance, while removing low-coverage samples improves data efficiency without notable performance degradation. Furthermore, all proposed models achieve real-time inference exceeding 60 frames per second (FPS), confirming their suitability for onboard deployment. Overall, the results of this study provide practical guidelines for jointly optimizing training data composition and model efficiency in AAM/UAM applications.