The effective deployment of connected vehicular networks is contingent upon maintaining a desired performance across spatial and temporal domains. In this paper, a graph-based framework, called SMART, is proposed to model and keep track of the spatial and temporal statistics of vehicle-to-infrastructure (V2I) communication latency across a large geographical area. SMART first formulates the spatio-temporal performance of a vehicular network as a graph in which each vertex corresponds to a subregion consisting of a set of neighboring location points with similar statistical features of V2I latency and each edge represents the spatio-correlation between latency statistics of two connected vertices. Motivated by the observation that the complete temporal and spatial latency performance of a vehicular network can be reconstructed from a limited number of vertices and edge relations, we develop a graph reconstruction-based approach using a graph convolutional network integrated with a deep Q-networks algorithm in order to capture the spatial and temporal statistic of feature map pf latency performance for a large-scale vehicular network. Extensive simulations have been conducted based on a five-month latency measurement study on a commercial LTE network. Our results show that the proposed method can significantly improve both the accuracy and efficiency for modeling and reconstructing the latency performance of large vehicular networks.