Optimizing roadside vegetation using deep reinforcement learning to improve thermal environment

Urbanization has accelerated, exacerbating the urban heat island effect, particularly in roadside areas. Adjusting roadside vegetation has emerged as a crucial strategy to ameliorate the thermal conditions along roadsides. However, conventional methods are often limited by fixed vegetation types and layout schemes within a limited area, posing challenges in effectively responding to the diverse optimization needs of complex environments. To address this issue, this study proposes a multitype vegetation optimization model based on reinforcement learning. Using street view images, we extracted large-scale, high resolution tree-shrub-grass multitype vegetation. Simultaneously, the optimization and adjustment of the vegetation structure is formulated as a Markov decision process. A pretrained nonlinear model is used to construct a reward mechanism for the cooling effect, and then a reinforcement learning algorithm with the twin delayed deep deterministic policy gradient (TD3) is used for the optimization strategy. Empirical results reveal that in the primary urban area of Beijing, increasing tree density and reducing grass coverage are the most effective strategies for roadside cooling in high-density built-up areas. Conversely, increasing grass coverage is optimal for roadside cooling at urban peripheries. Furthermore, spatially integrated optimization led to a reduction in local land surface temperature (LST) of approximately 1–3 °C, with an average LST reduction of 1.04 °C across the study area. Densely clustered trees contribute to localized cooling effects, whereas dispersed grasses are more effective at promoting overall cooling. These findings offer valuable insights for the formulation of effective urban vegetation planning and management strategies aimed at enhancing the roadside thermal environment.