Assessing ozone networks using positive matrix factorization

• Published in: Environmental progress 2004-07, Vol.23 (2), p.110-119
• Main Authors: Rizzo, Michael J., Scheff, Peter A.
• Format: Article
• Language: English
• ISSN: 0278-4491; 1547-5921
• DOI: 10.1002/ep.10018

In 2001, the United States Environmental Protection Agency (USEPA) began the process of examining the national monitoring networks to assess the contribution of individual monitoring sites in providing useful information to the public and regulatory agencies. One of the first networks to be examined was ozone, with the assessment being initially completed on a national level and then further refined on a regional basis. The goal of the regional analysis was to determine which monitors may be providing redundant information and could, therefore, be removed or relocated to another area in need of additional monitoring data. One technique which was used in the regional analysis of the ozone network was positive matrix factorization (PMF). This technique is similar to classical factor analysis, which allows for a series of related variables to be grouped into a smaller set of independent factors that represent combinations of the original variables. In addition to grouping the data into factors, this novel approach also provides predicted values of the analysis variable. Comparison of the predicted to the actual values not only gave an indication of how well the model fitted the ozone concentrations, but also aided in the determination of the information value of individual monitors. Hourly ozone data were polled from the USEPA's national data archive for a total of 24 states for the prime ozone formation months of May through September for 1996 to 2000. Daily maximum 8‐hour concentrations were calculated for each site according to the methods contained in 40 CFR Part 50 Appendix H. Because PMF requires a complete data record across all sites for all days analyzed, sites that were missing data were interpolated linearly over time. The results of the PMF analysis contained 10 factors representing various areas of the country including the Lake Michigan, Atlantic Coast, North Carolina, St. Louis/Indianapolis, Upper New York State, Ohio, Pennsylvania, Kansas/Southeast Missouri/Arkansas, Minnesota/Northwest Wisconsin, and Kentucky/Tennessee areas. Actual to predicted ratios were calculated for each day at each site and the coefficients of variation (CVs) of the individual ratio distributions were utilized as a metric to determine which sites were consistently being predicted well by PMF. Sites with low CVs were interpreted as being well predicted and considered not to be providing ambient ozone information as valuable as that provided by monitors that were poorly