Deep reinforcement learning based route selection and accessibility enhancement algorithm for pedestrian green corridors

Pedestrian green corridors support healthier and more sustainable urban mobility; however, routes with the highest green exposure often lack sufficient accessibility for wheelchair users, older adults, and individuals with strollers. Current routing approaches rely on shortest-path heuristics or fixed-weight multi-criteria formulations, which are insufficient for capturing the nonlinear trade-offs among greenery, slope, curb-ramp availability, and sidewalk continuity. This study introduces GreenAccess-RL, a constrained deep reinforcement learning (DRL) framework designed to jointly optimize green exposure and pedestrian accessibility. The framework consists of four integrated components: (i) a public-data graph construction pipeline that combines OpenStreetMap, OpenSidewalks, Sentinel-2 normalized difference vegetation index (NDVI), and shuttle radar topographic mapping (SRTM) elevation data to construct a richly attributed pedestrian network; (ii) a graph neural network (GNN) encoder with proximal policy optimization (PPO) that learns a route policy to maximize cumulative green exposure while meeting hard and soft accessibility constraints; (iii) an accessibility enhancement module that identifies infrastructure bottlenecks and prioritizes low-cost interventions within a fixed budget; and (iv) a cross-city evaluation strategy to support reproducible and transferable urban artificial intelligence research. Experiments conducted in three United States cities (Seattle, Washington D.C., and Portland) indicate that GreenAccess-RL achieves 89–97% higher cumulative NDVI than shortest-path routing and 27–30% higher than the strongest multi-criteria baseline, while maintaining accessibility feasibility rates above 91%. The enhancement module demonstrates that a limited set of targeted interventions can improve accessible green-route quality by up to 18%. Transfer-learning experiments show that a policy trained in one city retains 88–93% of its original performance when applied to an unseen city. These findings confirm that DRL offers a flexible, data-driven framework for jointly optimizing pedestrian greenery, accessibility, and route efficiency, thereby bridging the gap between navigation and urban planning support.