GIS APPROACHES FOR ESTIMATION OF RESIDENTIAL-LEVEL AMBIENT PM CONCENTRATIONS

Abstract

ISEE-43 Introduction: Spatial interpolation methods available in a geographic information system (GIS) are increasingly considered for estimation of geo-coded location-specific ambient air pollutant concentrations using measured data from the U.S. Environmental Protection Agency Aerometric Information Retrieval System (AIRS) database. To accomplish this, important methodological and practical issues still need to be resolved. Aim: This study was conducted to (1) compare the estimations/interpolations of residential-level ambient PM concentrations using national vs. regional data in large-scale population-based studies, (2) perform cross-validations from different kriging models, (3) describe and contrast the most appropriate approaches, and (4) estimate the standard errors of these interpolations. Methods: The year 2000 PM10 and PM2.5 data from the AIRS were used as the source data. Kriging interpolations were performed on 34,000 geo-coded residential locations across the continental U.S. An automated procedure using ArcView GIS and its Geospatial Statistical Analyst extension were employed in this study. Four cross-validation and kriging parameters were chosen to assess the goodness of kriging interpolations: standardized prediction error (SPE), prediction error (PE), root mean square standardized (the prediction variance, RMSS), and standard error (SE) of the predicted PM. Results: Based on these parameters, both national and regional-scale kriging performed satisfactorily, the former slightly better than the latter. The average PE, SPE, and RMSS from cross-validation of exponential, Gaussian and spherical semivariogram models are presented in Table 1, the average standard error of the estimated PM10 and PM2.5 using national scale-kriging. Conclusion: We have developed an automated procedure to perform large-scale interpolations using available GIS techniques and assessed the errors in those interpolations/estimations. National kriging interpolation is a better approach than regional kriging in large-scale studies. The goodness of the interpolations/estimations of residential-level daily PM concentrations in large-scale population-based studies can be evaluated using four cross-validation parameters. Although the three semivariogram models perform similarly, PM generally has a high spatial autocorrelation within short distances and gradually decreases with increasing distance. Moreover, the assumption of a spherical model more closely reflects observed PM spatial distributions. We therefore recommend using the spherical model.