Urban scene understanding requires both global geographic context and local structural detail. Earth Observation (EO) imagery supports large-scale land-cover and land-use (LCLU) mapping, but in urban areas it often merges heterogeneous surfaces into broad built-up classes. Vehicle-based sensors such as LiDAR and cameras recover these local structures, but their maps can drift and often remain in a local coordinate frame. We present urban graph, which combines overhead EO priors, vehicle observations, and fixed roadside anchors in a hierarchical semantic scene graph. Coarse georeferenced regions from EO data are updated with local observations, while a factor graph jointly optimises SLAM constraints and global geodetic constraints. The resulting graph is projected back to the overhead layer to separate coarse urban classes into finer semantic components. Experiments in CARLA show improved global alignment, reduced drift, and more detailed projection of local semantics into EO space.