Because of the complex turbulent flow dynamics caused by interactions between the atmospheric surface layer and the urban topography, forecasting microscale pollutant concentration fields is crucial for monitoring plume dispersion in urban regions. To capture these dynamics, large-eddy simulation (LES) is regarded as a high-fidelity numerical approach, however, it lacks real-time capabilities and remains costly because highly dimensional. Reynolds-averaged Navier-Stokes (RANS) approach is a cruder path to modelling these phenomena based on a statistical description of turbulent phenomena, leading to averaged transport equations. It is much less computationally expensive but also less accurate in representing interactions between the atmospheric boundary-layer flow and the buildings in urban areas. Designing an efficient reduced-order model (ROM) with a level of precision similar to the LES approach and accounting for atmospheric and point-source emission uncertainties is of primary importance. In this paper, we propose a novel multi-fidelity ROM, which combines two levels of data fidelity - LES and RANS - and we evaluate it in the context of atmospheric flow dispersion. RANS data are obtained by injecting detailed flow information from LES into a lower-fidelity tracer transport equation, in the RANS formalism, using physics-constrained machine learning. There are two steps in our approach: i) dimension reduction is performed using a convolutional autoencoder trained using pre-training, and ii) mapping the uncertain parameters to the autoencoder latent space is obtained using co-kriging and an autoregressive model. We consider a case of a simplified two-dimensional flow configuration around a surface-mounted obstacle, where the main physical quantity of interest is the time-average tracer concentration field and its variability with respect to uncertain atmospheric forcing and emission source location. We show that the multi-fidelity approach achieves increased performance compared to the LES and RANSsingle-fidelity approaches at equivalent computational budget. This is a promising approach to designing an efficient ROM applicable to realistic field-scale atmospheric dispersion cases.