Evaluation of MARS for the spatial distribution modeling of carbon monoxide in an urban area

Abstract

Spatial distribution modeling of CO in Tehran can lead to better air pollution management and control, and it is also suitable for exposure assessment and epidemiological studies. In this study MARS (Multi–variate Adaptive Regression Splines) is compared with typical interpolation techniques for spatial distribution modeling of hourly and daily CO concentrations in Tehran, Iran. The measured CO data in 2008 by 16 monitoring stations were used in this study. The Generalized Cross Validation (GCV) and Cross Validation techniques were utilized for the parameter optimization in the MARS and other techniques, respectively. Then the optimized techniques were compared based on the mean absolute of percentage error (MAPE). Although the Cokriging technique presented less MAPE than the Inverse Distance Weighting, Thin Plate Smooth Splines and Kriging techniques, MARS exhibited the least MAPE. In addition, the MARS modeling procedure is easy. Therefore, MARS has merit to be introduced as an appropriate method for spatial distribution modeling. The number of air pollution monitoring stations is very low (16 stations for 22 zones) and the distribution of stations is not suitable for spatial estimation, hence the level of errors was relatively high (more than 60%). Consequently, hourly and daily mapping of CO provides a limited picture of spatial patterns of CO in Tehran, but it is suitable for estimation of relative CO levels in different zones of Tehran. Hence, the map of mean annual CO concentration was generated by averaging daily CO distributions in 2008. It showed that the most polluted regions in Tehran are the central, eastern and southeastern parts, and mean annual CO concentration in these parts (zones 6, 12, 13, 14 and 15) is between 4.2 and 4.6 ppm.