Mobile monitoring of air pollutants; performance evaluation of a mixed-model framework in relation to number of drive days

Abstract

Background and Aim. Because of the implicit uncertainty of mobile measurements due to the short-term nature of the measurements, most studies to date have used empirical modeling to stabilize predicted concentrations, thereby losing local spatial information. In a previous paper, we demonstrated that a mixed-model can stabilise the measurements by a land use regression (LUR) model, while allowing street segments to deviate from the LUR prediction based on between-segment variation of the measurements as a random effect. We analysed how many drive-days are needed for the mixed-effect model to improve predictions from LUR models and data-only mapping. Methods. We used black carbon data from the Air View study in Oakland, where every street segment was measured at least 50 times. We selected one drive day per street segment and compared the measurement, the LUR prediction and the mixed-model prediction with the average concentration based on 50 drive days of that street segment. We assumed that driving 50 times on a street segment reflects a robust long-term average concentration. We then sequentially added drive days to the dataset and computed the explained variance (R²) and RMSE. Results. With one drive day on every street segment, the LUR model explained 63% of the variation, with very limited improvement in performance with increasing number of drive-days (65% for 50 drive days). The data-only map predicts less than 30% for one drive-day and more than 90% of the variance after 15 drive days, surpassing the LUR model in explained variance at 4-5 drive days. The mixed model outperformed the data-only and LUR model estimates, with 75% explained variance after 2 drive days and 90% after 12 drive days. Conclusion. The mixed-model improves predictions compared to LUR and data already with two drive days and updates the model when more drives become available.