Leveraging Automated Machine Learning (AutoML) for Urban Climate Emulation

Junjie Yu, Zhonghua Zheng , Sarah Lindley, Lei Zhao, Chi Wang, Qingyun Wu, Lingcheng Li, David O. Topping, John S. Schreck, David John Gagne, Keith W. Oleson

Urban climate models are critical for understanding and addressing the impacts of urban climate change. Yet, process-based urban climate models face limitations of high-entry barriers and substantial computing resource consumption, prompting the development of data-driven methods. However, the recently developed urban climate emulators, being location-dependent, are less scalable and may overlook geospatial data. In this study, we develop location-independent machine learning emulators for the daily maximum canyon air temperature. To overcome the complexities associated with model selection and hyperparameter optimization in machine learning, we apply automated machine learning (AutoML) to emulation tasks and propose a feature importance analysis framework for the AutoML models. By comparing four types of global urban climate emulators, we found that the location information and urban surface parameters can improve the emulation performance. The results of the AutoML tasks demonstrate that AutoML excels in learning the physics-based urban climate model, achieving a root mean squared error (RMSE) of 0.81 Kelvin for emulators parameterized with location information and urban surface parameters, and an RMSE of 0.91 Kelvin in the temporal extrapolation scenario. The feature importance of the emulators indicates that urban morphological parameters contribute more to the emulators than radiative and thermal parameters. The study serves as a demonstration of the potential that AutoML holds for advancing urban climate research and facilitating urban climate modeling.