This study set out to develop a land use regression model at sub-neighborhood scale (0.01–1 km) for Portland, Oregon using passive measurements of NO2 at 77 locations. Variables used to develop the model included road and railroad density, traffic volume, and land use with buffers of 50 to 750 m surrounding each measurement site. An initial regression model was able to predict 66% of the variation in NO2. Including wind direction in the regression model increased predictive power by 15%. Iterative random exclusion of 11 sites during model calibration resulted in a 3% variation in predictive power. The regression model was applied to the Portland metropolitan area using 10 m gridded land use layers. This study further validates land use regression for use in North America, and identifies important considerations for their use, such as inclusion of railways, open spaces and meteorological patterns.