Improved retrieval of PM 2.5 from satellite data products using non-linear methods M. Sorek-Hamer a , A.W. Strawa b , R.B. Chatfield b , R. Esswein c , A. Cohen d , D.M. Broday a, * a Civil and Environmental Engineering, Technion, Haifa, Israel b NASA Ames Research Center, Moffett Field, CA, USA c Bay Area Environmental Research Institute, Sonoma, CA, USA d Industrial and Management Engineering, Technion, Haifa, Israel article info Article history: Received 29 April 2013 Received in revised form 31 July 2013 Accepted 2 August 2013 Keywords: PM MODIS OMI GAM MARS abstract Satellite observations may improve the areal coverage of particulate matter (PM) air quality data that nowadays is based on surface measurements. Three statistical methods for retrieving daily PM 2.5 con- centrations from satellite products (MODIS-AOD, OMI-AAI) over the San Joaquin Valley (CA) are compared e Linear Regression (LR), Generalized Additive Models (GAM), and Multivariate Adaptive Regression Splines (MARS). Simple LRs show poor correlations in the western USA (R 2 y 0.2). Both GAM and MARS were found to perform better than the simple LRs, with a slight advantage to the MARS over the GAM (R 2 ¼ 0.71 and R 2 ¼ 0.61, respectively). Since MARS is also characterized by a better compu- tational efficiency than GAM, it can be used for improving PM 2.5 retrievals from satellite aerosol products. Reliable PM 2.5 retrievals can fill in missing surface measurements in areas with sparse ground monitoring coverage and be used for evaluating air quality models and as exposure metrics in epidemiological studies. Ó 2013 Elsevier Ltd. All rights reserved. 1. Introduction Inferring PM 2.5 from satellite data has been an active research field for over a decade (Hoff and Christopher, 2009). Typically, PM is obtained from surface measurements which may be sparse, with an apparent limitation of spatial coverage. Satellite observations can fill in this gap because they have a broader areal coverage than surface monitoring. An aerosol data product that is oftentimes used for estimating ground PM is the Aerosol Optical Depth (AOD) from the MODerate resolution Imaging Spectroradiometer (MODIS) aboard the Terra and Aqua satellites. Ichoku et al. (2002) suggested that MODIS AOD, s, is spatially correlated with hourly averages of AOD observed by ground sunphotometers (AERONET), with a global uncertainty of the MODIS AOD relative to the AERONET AOD of s ¼(0.05 þ 0.15s)(Levy et al., 2010). Moreover, based on 7 years of data Kaufman et al. (2000) concluded that AOD from MODIS instantaneous overpass was highly correlated with daily average AERONET AOD from 50 to 70 globally distributed sites. Thus, whereas hourly PM 2.5 measurements may demonstrate better correlation with the matched AOD, daily average PM 2.5 concen- trations are more useful for environmental health studies and regulations. Hence, daily aerosol products were used throughout this study (Sorek-Hamer et al., 2013). Different methods have been used to predict PM 2.5 concentra- tions from satellite data, particularly AOD, including linear re- gressions (Chu et al., 2003; Engel-Cox et al., 2004; Gupta et al., 2006; Gupta and Christopher, 2008; Choi et al., 2009; Péré et al., 2009), Generalized Linear Models (GLM) (Liu et al., 2005, 2007) and Generalized Additive Models (GAM) (Paciorek et al., 2008; Strawa et al., 2013). Simple linear regressions between PM 2.5 and MODIS AOD show reasonable correlations in some areas, notably the East Coast of the US, however poor correlations have been observed between MODIS AOD and collocated daily/hourly PM 2.5 measure- ments in the western USA (Engel-Cox et al., 2004; Zhang et al., 2009). The latter can be attributed to several possible reasons, including differences in surface reflectance, differences in meteo- rological conditions (e.g. planetary boundary layer height and relative humidity profile along the atmospheric column), presence of aerosols aloft (e.g. long range transport of mineral dust), and geographical differences in PM composition (Engel-Cox et al., 2004; Al-Saadi et al., 2005; Gupta et al., 2006; Pelletier et al., 2007). Although elevated PM 2.5 concentrations (>7 mm/m 3 ) were found to be associated with about 14,000 (range 4300e25,000) premature deaths per year in CA (Tran et al., 2009), ground PM 2.5 measurements in California are sparse and provide only limited * Corresponding author. E-mail address: dbroday@tx.technion.ac.il (D.M. Broday). Contents lists available at ScienceDirect Environmental Pollution journal homepage: www.elsevier.com/locate/envpol 0269-7491/$ e see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.envpol.2013.08.002 Environmental Pollution 182 (2013) 417e423