Using self-organizing maps to develop ambient air quality classifications: a time series example

• Type of Material: Article
• Authors: John L Pearce, Emory University; Lance Waller, Emory University; Howard Chang, Emory University; Mitchel Klein, Emory University; James A Mulholland, Georgia Institute of Technology; Jeremy Sarnat, Emory University; Stefanie Ebelt Sarnat, Emory University; Matthew J Strickland, Emory University; Paige Tolbert, Emory University
• Language: English
• Date: 2014
• Publisher: BMC (part of Springer Nature)
• Publication Version: Final Published Version
• Copyright Statement: © 2014 Pearce et al.; licensee BioMed Central Ltd.
• License: Creative Commons Attribution 2.0 Generic
• Final Published Version (URL): http://www.ehjournal.net/content/13/1/56
• Title of Journal or Parent Work: Environmental Health
• ISSN: 1476-069X
• Volume: 13
• Issue: 56
• Grant/Funding Information: This publication was made possible, in part, by US Environmental Protection Agency grant R834799.; Research reported in this publication was supported by the National Institute of Environmental Health Sciences of the National Institutes of Health under Award Numbers T32ES016160 and K99ES023475.
• Keywords: Cluster analysis; Classification; Air pollution; Kohonen Map
• Research Categories: Health Sciences, Public Health; Environmental Sciences

Abstract

Background Development of exposure metrics that capture features of the multipollutant environment are needed to investigate health effects of pollutant mixtures. This is a complex problem that requires development of new methodologies. Objective Present a self-organizing map (SOM) framework for creating ambient air quality classifications that group days with similar multipollutant profiles. Methods Eight years of day-level data from Atlanta, GA, for ten ambient air pollutants collected at a central monitor location were classified using SOM into a set of day types based on their day-level multipollutant profiles. We present strategies for using SOM to develop a multipollutant metric of air quality and compare results with more traditional techniques. Results Our analysis found that 16 types of days reasonably describe the day-level multipollutant combinations that appear most frequently in our data. Multipollutant day types ranged from conditions when all pollutants measured low to days exhibiting relatively high concentrations for either primary or secondary pollutants or both. The temporal nature of class assignments indicated substantial heterogeneity in day type frequency distributions (~1%-14%), relatively short-term durations (<2 day persistence), and long-term and seasonal trends. Meteorological summaries revealed strong day type weather dependencies and pollutant concentration summaries provided interesting scenarios for further investigation. Comparison with traditional methods found SOM produced similar classifications with added insight regarding between-class relationships. Conclusion We find SOM to be an attractive framework for developing ambient air quality classification because the approach eases interpretation of results by allowing users to visualize classifications on an organized map. The presented approach provides an appealing tool for developing multipollutant metrics of air quality that can be used to support multipollutant health studies.