Innovative Techniques
to Combine Data and
Improve Environmental Modeling

Scientific studies have shown associations between air pollutants and negative human health effects, as well as detrimental ecological impacts. However, the interplay of diverse emission sources and complex atmospheric processes makes the air pollution problem difficult to understand. Recently, Battelle developed a hierarchical Bayesian approach to statistical spatial modeling for an air quality assessment project to estimate spatial gradients, or variation, of air pollutants for the U.S. Environmental Protection Agency (EPA). (The term "hierarchical" in this context means that large numbers of variables are not modeled simultaneously, but instead dealt with in layers that build upon one another.) Such models also can define the spatial areas that episodes of unhealthy air quality will affect, or illuminate relationships between different air pollutants.

This innovative approach can be applied in most situations where two or more sources of data, which may differ in bias and precision, are collected to study one process or situation. As one example, Battelle scientists considered monitoring data and model predictions as two spatial representations of fine particulate matter. EPA's Community Multi-Scale Air Quality (CMAQ) Modeling System, which provided predictive modeling data, collected data at many locations in the study area, but its predictions were considered biased and imprecise - i.e., spatially dense but inaccurate. The monitoring data were drawn from the nation-wide Federal Reference Method (FRM) monitoring network, the 'gold standard' of air quality information as measured. Unfortunately, this network of high-quality monitoring measurements provides data from relatively few locations - i.e., spatially sparse but accurate information. By combining both sources, the hierarchical Bayesian approach takes advantage of the complementary strengths of each one.

The graphics below illustrate ambient particulate matter (PM2.5) concentration data that were collected over a large portion of the eastern United States for a two-week period in January 2000. The Bayesian surface is generally higher than the CMAQ surface, indicating that the monitoring data were used to adjust for an innate bias in the CMAQ model. Also note that the numerous local area peaks that were not necessarily accounted for, due to the relative sparseness of the monitoring data, still appear in the Bayesian surface, illustrating that the CMAQ information also has been fully incorporated.

Left Plot of the CMAQ model (surface) and the monitoring observations (spheres).
Right Plot of the Bayesian predictions (surface) and the monitoring observations (spheres).

Other applications of this modeling technique include making predictions about how air pollutants will act within defined geographic areas over a given period of time, using information from more than two inputs, and using it to make several other promising extensions.

For more information, contact Dr. Steve Bortnick (614) 424-7487, bortnick@battelle.org or Dr. David Wendt (614) 424-7653, wendtd@battelle.org.