Abstract Efficient representation of complex infrastructure systems is essential for tasks such as edge prediction, component classification, and decision-making. However, interactions between these systems and their spatial environments complicate network representation learning. This study introduces the Regional Spatial Graph Convolutional Network (RSGCN), a novel geometric-based multi-modal deep learning model for spatially embedded networks. RSGCN learns from nodes spatial features and is evaluated by embedding and reconstructing various infrastructure networks, including the California Highway Network and the New Jersey Power Network. Compared to GraphSAGE and the Spatial Graph Convolutional Network (SGCN), RSGCN demonstrates superior performance, highlighting the benefits of incorporating regional information for accurate network representations. Content The folder contains the codes and dataset used for the paper "Modeling of spatially embedded networks via regional spatial graph convolutional networks" https://doi.org/10.1111/mice.13286