Transit-oriented land use planning for regional metro networks via deep reinforcement learning

Transit-oriented development (TOD) mitigates traffic congestion and emissions by promoting high-density, mixed-use development around transit stations. While extensive research exists on single-station catchments, TOD planning at the metro network level remains underexplored. Recent studies emphasize the importance of coordinating land use across stations, yet a decision-support framework for network-scale TOD optimization is still absent. This study introduces a deep reinforcement learning framework for regional metro network TOD planning (DRL-RMN). It formulates network-level land use planning as a sequential decision-making process, incorporating a graph neural network to encode spatial dynamics and a masked hierarchical action mechanism to account for development constraints. Multiple planning objectives are simultaneously optimized, including population coverage, network opportunity accessibility, and spatial equity. Results from Shanghai and Beijing reveal that DRL-RMN achieves a superior balance between efficiency and equity compared to heuristic benchmarks. An interpretable machine learning model is applied to understand the planning results, revealing that recommended office development tends to cluster around stations with high closeness centrality, whereas the pattern of residential growth is more city-specific. We further demonstrate the flexibility of DRL-RMN by accommodating varied planning objectives (e.g., prioritizing efficiency vs. equity) and budget constraints (e.g., high vs. low). Altogether, our approach offers planners a scalable, interpretable tool for integrating land use and transport systems at the network level.