Transit-Oriented Land Use Planning Using Deep Reinforcement Learning

Bo Yang, Yuebing Liang, Zhan Zhao, Fangyi Ding, Yunhan Zheng, Shenhao Wang, Jinhua ZhaoarXiv 2026 · arXiv

TL;DR

DRL4TOD is a novel framework that applies graph-based deep reinforcement learning to optimize land-use types and development intensity around transit stations. By capturing complex spatial interdependencies through graph neural networks, it consistently outperforms traditional optimization methods, offering planners a flexible tool to balance sustainable urban density with real-world financial feasibility.

What Problem Does This Paper Address?

Transit-Oriented Development (TOD) is the "holy grail" of sustainable urbanism: creating dense, mixed-use neighborhoods within walking distance of public transit to reduce car dependency. However, designing these areas is a combinatorial nightmare.

Planners must decide the land-use type (residential, office, park, etc.) and Floor Area Ratio (FAR) for hundreds of irregular parcels. Every decision is interdependent; a new office tower on one block changes the transit demand and the "feel" of the neighboring blocks. Traditional planning relies on manual trial-and-error or simplified "grid-based" computer models that don't reflect the irregular reality of city parcels. These methods struggle to find quantitatively optimal patterns that balance high transit ridership with the high costs of demolishing and reconstructing existing buildings.

Methods

The authors introduce DRL4TOD, which treats urban planning as a sequential decision-making problem—a Markov Decision Process (MDP) solved via Proximal Policy Optimization (PPO).

  • Graph-Based Spatial Representation: Instead of treating a city like a grid of pixels, the authors model it as a graph. Nodes represent parcels, road intersections, and metro stations, while edges capture physical contiguity and road connectivity.
  • GNN State Encoder: A Graph Neural Network (GNN) uses message-passing layers to "understand" the spatial context. It allows the model to factor in how a parcel’s potential relates to its neighbors and its distance to the transit hub.
  • Multi-Task Policy Network: The DRL agent performs two actions simultaneously: it selects a parcel to modify (discrete action) and assigns a specific FAR value (continuous action sampled from a Beta distribution).
  • Simulation-Driven Rewards: The reward function is multi-objective, balancing Land Use Density, Diversity, and Development Feasibility (minimizing reconstruction costs). Notably, "Transit Capacity Utilization" is calculated using an auxiliary XGBoost model trained on real-world Shanghai metro data to predict ridership based on the generated land-use patterns.

Datasets

The framework was validated using two real-world station areas in Shanghai:

  1. Xintiandi Station: A central, high-density, mixed-use urban hub.
  2. Jinjing Road Station: A more peripheral, residential-heavy suburban context.

The researchers utilized official cadastral data, road networks, and existing land-use records, supplemented by open geospatial sources like OpenStreetMap and Microsoft Building Footprints.

Key Insights and Ideas

  • Topology Over Geometry: The study demonstrates that graph representations are fundamentally superior to raster grids for urban planning. Graphs preserve the "parcel-road-station" relationships that define how people actually navigate a city.
  • The Necessity of GNNs: An ablation study revealed that replacing the GNN with a standard Multilayer Perceptron (MLP) caused performance to tank. Without graph-based reasoning, the model failed to cluster high-intensity uses near stations, proving that "station-aware" features are essential for TOD logic.
  • Planner-in-the-Loop: DRL4TOD isn't meant to replace humans. It's a decision-support tool. As shown in Figure 1, the human planner sets the initial priorities (weights) and reviews the AI-generated alternatives against qualitative factors like "historical character" or "thermal comfort."
Figure 1: Planner-Guided Decision-Support Workflow

Key Results

  • Massive Density Gains: In the Xintiandi case, DRL4TOD achieved a Land Use Density (LUDE) score of 1.6965, nearly double the performance of the strongest baseline (Adaptive Genetic Algorithm at 0.8216).
  • Improved Ridership: The model increased simulated transit ridership by up to 30.2% compared to existing initial plans.
  • Surgical Urban Renewal: Remarkably, DRL4TOD achieved these high-performance outcomes while maintaining higher "Development Feasibility" scores than baselines. This suggests the DRL agent found "smarter" ways to densify that required less total demolition of the existing urban fabric.

Limitations and Future Work

  • Single-Use Parcels: Currently, the model assigns one primary land-use type per parcel. In reality, many TOD buildings are "vertical mixed-use" (e.g., retail on the ground floor with apartments above).
  • Social Equity: The current reward function focuses on physical and economic metrics. It does not explicitly account for social factors like housing affordability, tenure security, or the risk of resident displacement.
  • Static Snapshots: The model optimizes a static "snapshot" of the city. Future versions could incorporate multi-period planning to account for how urban growth and demographic needs evolve over decades.

Summary updated

Abstract

Transit-Oriented Development (TOD) is a key strategy for advancing sustainable urban development by promoting high-density, mixed-use land development within walking distance of public transit stations. However, traditional TOD planning often relies on heuristic methods, making it difficult to efficiently identify quantitatively optimized transit-oriented land use patterns. This paper presents DRL4TOD, a deep reinforcement learning framework that automatically optimizes land use types and floor area ratios (FARs) around transit stations in alignment with TOD principles. DRL4TOD incorporates a graph neural network to capture spatial interdependencies among land parcels, road networks, and metro stations, and uses a DRL-based optimization strategy to assign land-use types and FARs to selected parcels. It jointly optimizes multiple objectives, including land use density, land use diversity, and transit capacity utilization, while minimizing reconstruction costs to account for development feasibility. Applied to two real-world station areas in Shanghai, DRL4TOD consistently outperforms rule-based approaches and genetic algorithms, achieving more effective TOD outcomes with lower reconstruction costs. By adjusting objective weights, the framework enables flexible prioritization, such as placing greater emphasis on cost minimization when financial feasibility is a concern. These results demonstrate DRL4TOD's potential as a decision-support tool for generating coordinated land-use and FAR configurations under planner-specified priorities, rather than prescribing a single fixed design. For policymakers, the framework provides parcel-level, constraint-aware alternatives that make trade-offs among competing objectives explicit and support targeted decisions about development intensity and feasibility.
BibTeX
@misc{yang2026transit,
  author = {Bo Yang and Yuebing Liang and Zhan Zhao and Fangyi Ding and Yunhan Zheng and Shenhao Wang and Jinhua Zhao},
  title = {{Transit-Oriented Land Use Planning Using Deep Reinforcement Learning}},
  year = {2026},
  eprint = {2611.10281},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10281},
}
Figure