Urban areas are known to modify the interaction between land surfaces and the lower atmosphere, and therefore impact the momentum and scalar transport, the patterns of temperature and precipitation, and so forth. The overarching goals of this work are investigating the local and non-local impact induced by different urban forms in the lower atmosphere, understanding the uncertainties of urban form representation in high-resolution numerical models, and exploring the feasibility of adopting a machine learning-based method as an alternative way to address the uncertainties. Chapter 1 presents a general review of the current status of urban form representation in the land surface model and the challenges in representing urban land surface coupling. Chapter 2 investigates the impact of urban structures on precipitation distribution through a cross-city analysis using high-resolution observation data. The results present evidence that the spatial pattern of precipitation can be influenced by the interaction between background regional conditions and urban forms. This implies the importance of understanding and capturing the impact induced by urban structures and motivates the next chapters to explore more in numerical simulations. Chapter 3 examines how the presence of urban structures affects the error growth of initial perturbation and corresponding predictability in streamwise velocity and passive scalar fields through identical-twin simulations. This study is intended to address the response of the system to a perturbation, i.e., initial uncertainty. The result reveals that urban structures can decrease the predictability of the passive scalar and destroy the similarity between the error statistics of the velocity and the passive scalar. Chapter 4 focuses on the uncertainties induced by the parameters and investigates the impact of urban forms with different streamwise heterogeneities but the same urban canopy parameters (i.e., mean building height and fraction) on the lower atmosphere. The non-local effects of three different urban forms highlight the importance of capturing urban forms, especially the height variation of buildings, in the Numerical Weather Prediction (NWP) models. Chapter 5 explores the feasibility of using the machine learning method to address the uncertainties and capture the impact of different urban forms. The U-net architecture is trained using the large eddy simulation output to learn the mapping from the urban canopy structure to the velocity field. The prediction error for velocity and aerodynamic drag coefficient shows the prediction capability of the trained model and highlights the potential of combining computational fluid dynamics modeling and machine learning to develop city-specific parameterizations for NWP models. The conclusions and outlook are in the last chapter.