Targeting the problem of generating high-resolution air quality maps for cities, we leverage four different sources of data: (i) in-situ air quality measurements produced by our mobile sensor network deployed on public transportation vehicles, (ii) explanatory air-quality and meteorological variables obtained from two static monitoring stations, (iii) land-use data of the city, and (iv) traffic statistics. We propose two novel approaches for estimating the targeted pollutant level at desired time-location pairs, extending also to areas of the city that are beyond the coverage of our mobile sensor network. The first is a log-linear regression model which is built over a virtual dependency graph based on land-use data. The second is a deep learning framework that automatically captures the dependencies of the data based on autoencoders. We have evaluated the two proposed approaches against three canonical modeling techniques considering metrics of coefficient of determination (R-squared), root mean square error (RMSE), and the fraction of predictions within a factor of two of observations (FAC2). Using more than 45 million real measurements in the models, the results show consistently superior performance in respect to the canonical techniques.