Artificial intelligence-supported spatial optimization of urban greenway networks

The goal of the study was to construct an AI-driven end-to-end framework to improve the restorative environmental performance of urban greenway networks. Generally, methods for greenway planning may have subjectivity, low optimization efficiency, and difficulty in quantifying multi-dimensional objectives. The framework integrates convolutional neural networks (ResNet-50) convolutional neural networks (CNN) for landscape quality assessment, GraphSAGE-based graph neural networks (GNN) for spatial topology modeling, and proximal policy optimization (PPO)-based deep reinforcement learning (DRL) for multi-objective optimization. A system for restorative assessment was established based on Attention Restoration Theory. This system has four dimensions, being away, fascination, compatibility, and extent. The generalization of the framework was systematically validated in the case of three representative urban scenarios; plains of a medium-sized city, hills of a small city, and a high-density metropolis. The findings show that compared to manual planning, the framework yielded a restorative score improvement of 42.2%, a 72.7% increase in population coverage, and 98.1% enhancement in efficiency (Optimization time from 120 hours to 2.3 hours). Spatial equity improved due to the decrease in the Gini coefficient from 0.42 to 0.28. Strong transferability is evident as migration cost in cross-scenarios is under 5%. The performance dropped by 26% to 41% when any of the modules (CNN, GNN, DRL) were removed. Multi-objective optimization was better than single-objective techniques.  The framework endowed with quantitative decision-support tools, facilitates healthy city construction. It promotes spatial justice by directing physical resources to the most vulnerable sections of the community. Further, the framework provides support for rapid iterative planning of green infrastructure for a smart city.