We propose three modeling methods using a mobile sensor network to generate high spatio-temporal resolution air pollution maps for urban environments. In our deployment in Lausanne (Switzerland), dedicated sensing nodes are anchored to the public buses and measure multiple air quality parameters including the Lung Deposited Surface Area (LDSA), a state of the art metric for quantifying human exposure to ultrafine particles. In this paper, our focus is on generating LDSA maps. In particular, since the sensor network coverage is spatially and temporally dynamic, we leverage models to estimate the values for the locations and times where the data are not available. We first discretize the area topologically based on the street segments in the city and we then propose the following three prediction models: i) a log-linear regression model based on nine meteorological (e.g., temperature and precipitations) and gaseous (e.g., NO2 and CO) explanatory variables measured at two static stations in the city, ii) a novel network-based log-linear regression model that takes into account the LDSA values of the most correlated streets and also the nine explanatory variables mentioned above, iii) a novel Probabilistic Graphical Model (PGM) in which each street segment is considered as one node of the graph, and inference on conditional joint probability distributions of the nodes results in estimating the values in the nodes of interest. More than 44 millions of geo- and time- stamped LDSA measurements (i.e., more than 14 months of real data) are used in this paper to evaluate the proposed modeling approaches in various time resolutions (hourly, daily, weekly and monthly). The results show that the three approaches bring significant improvements in R2 , RMSE and FAC metrics compared to a baseline K-Nearest Neighbor method.