Automatic responsive-generation of 3D urban morphology coupled with local climate zones using generative adversarial network

Abstract

Decoupling the intricate relationship between three-dimensional (3D) urban morphology and local climate is paramount importance in the realm of adaptive urban planning. Research based on Local Climate Zone (LCZ) has exhibited promising potential for disentangling the interactive mechanisms between urban morphology and local climate. In this study, generative adversarial networks (GAN) were employed as surrogate models to facilitate the integration of LCZ and urban morphology data from six metropolises for associative training. Through comparative experiments involving quantitative and qualitative evaluations, a rapid and responsive 3D morphology prediction model was developed for LCZ classifications. The results demonstrated the superior convergence capabilities and accuracy of the Pix2pix model (RMSE = 0.187 and R2 = 0.878) when compared to the CycleGAN model (RMSE = 0.344 and R2 = 0.674). The predicted 3D morphologies showcased a pronounced alignment with their respective LCZs, as evidenced by the precise representation of open low-rise buildings and their interrelationships with the coastline in the Sydney sample, as well as the accurate portrayal of the complex urban morphology shaped by compact high-rise buildings and road networks in the Tokyo sample. Utilizing indices calculated from the 3D morphology, the model yielded an average overall accuracy (OA), kappa coefficient, and F1 score of 85.2%, 0.83, and 0.86, respectively, indicating the robustness and adaptability of the model. The framework presented in this research offers practitioners a solid foundation and valuable insights for enhancing the local climate in urban areas through the implementation of 3D generative design techniques.