Characterizing meteorology in the southeastern U.S. using Classification and Regression Tree (CART)

Visibility Improvement State and Tribal Association of the Southeast (VISTAS) is one of five Regional Planning Organizations (RPOs) established to coordinate activities associated with the management of regional haze in Class I areas across the U.S. In addition to applying regional meteorological, emissions and air quality models to project visibility improvements, VISTAS has developed a conceptual description of contributions to regional haze in the southeastern U.S. using Classification and Regression Tree (CART) analyses. CART analyses for visibility were performed using meteorological and fine particulate data for 15 Class I areas with IMPROVE (Interagency Monitoring of Protected Visual Environments) monitors in the VISTAS region plus 5 Class I areas outside the VISTAS region. CART analyses were also performed for fine particle mass at 8 sites in the Southeastern Aerosol Research and Characterization Study (SEARCH) network and 16 sites in the Speciated Trends Network (STN) in the VISTAS region; however, these results will not be presented here.

CART provides information about the relationships between meteorological variables (surface and upper air) and visibility, the frequency of occurrence of different types of meteorology and visibility conditions, and the different contributions of PM2.5 components to visibility under different meteorological conditions. CART analyses indicate that high temperatures, high relative humidity, low wind speeds, and elevated PM2.5 on the previous day at upwind urban areas are good indicators of poor visibility days across the Southeast (Figure 1). Relative composition of PM2.5 on poor visibility days varies by site and by visibility class.

CART assigns days into classes based on a dependent variable. For visibility, days were sorted into five classes based on light extinction, with classes representing the 20th, 50th, 80th, and 95th percentiles. Within each class, days are sorted into bins that are defined by values of independent variables, in this case, meteorological variables and fine particle mass at upwind urban areas. The CART decision tree sorts days using the most influential variables first. Days that end up in the same bin are influenced by similar meteorological conditions. Each class is comprised of several bins, each with unique meteorological characteristics. Surface meteorological variables were obtained from the National Climatic Data Center for the period January 2000 – August 2003 at each IMPROVE site or from nearby monitoring sites with similar geographic features (e.g. elevation, latitude, land use) when the IMPROVE site did not have meteorological data. Upper air meteorological variables were obtained for available sites in the southeastern U.S. Using only meteorological inputs, CART analyses assign 65 to 80% of the days to the correct visibility classes. Accuracy is highest for days in the highest and lowest visibility classes. Adding PM2.5 mass on the previous day at upwind urban sites improves the accuracy of classifying days for visibility to 73-88%.

VISTAS has performed 2002 and 2018 annual air quality modeling simulations with CMAQ to evaluate the visibility improvements required by the Regional Haze Rule. In addition to using CART results to evaluate the appropriateness of 2002 to represent other years, CART has also been used to develop weighing factors for each modeled day of the year for each site with observations to allow VISTAS to evaluate the overall impact of emission reductions on air quality over longer time periods. Specifically, these weighting factors will allow the air quality model to represent the 5-year regional haze baseline period with only one year of modeling. CART can also be used to select specific days and episodes to represent the most frequently occurring meteorological conditions on the most polluted days. CART can be optimized across multiple sites to select periods from 2002 for more in-depth analysis of model performance, modeled air quality responses to emissions changes, and/or source attributions.