Comparison between national and local land use regression models for NO2based on routine monitoring data in Japan

Abstract

Background: To evaluate the health effects of air pollution on a national scale, national land use regression (LUR) models is required. However, the traffic volume, population size, or land usage are different between urban and rural areas. The predictability of the national LUR model for rural areas is still not clear. Thus, we built a national and local LUR model for NO2 in Japan, and compared the predictability.Method: We first, built a national-LUR model based on 1678 routine monitoring data for Nitrogen dioxide (NO2). Next, according to the tertile of the prefectural population size, we classified the monitoring sites into three groups. Then we built an urban-LUR (site n=1091), a middle-LUR (n=346), and a rural-LUR model (n=241). We calculated the predictability of each model and compared the predicted concentrations from each model with measured concentrations.Results: The predictability for NO2 concentrations of the national-LUR model was similar to that of the urban-LUR models: adjusted R2 = 0.75 and 0.75, respectively, and root mean square error (RMSE)=3.6 and 3.7 ppb, respectively. Adjusted R2 for the middle-LUR and the rural-LUR models were 0.66 (RMSE=3.2 ppb) and 0.70 (RMSE=2.8 ppb). Although the predictabilities of local LUR models were lower than the national-LUR models, RMSEs of local LUR models were smaller than RMSE in the national-LUR model. Also, the correlation coeffects between predicted and measured concentration in local models were higher than those in the national-LUR model. When concentrations were predicted by national-LUR in the rural area, the correlation coefficient was 0.81. It was 0.84 for the concentrations estimated by the rural-LUR model. Conclusion: The local LUR models might be more represented by the local land use conditions comparing with the national scale. Instead of population size, for improving the local model, the similarity might be evaluated by using a machine learning technique.