Real-Time Ozone Mapping Using a Regression-Interpolation Hybrid Approach, Applied to Tucson, Arizona

Abstract Real-time ozone (O3) maps, intended for public access and mass media, are generated from spatially interpolating (i.e., kriging) sparse monitoring data and are typically characterized by over-smoothed surfaces that inadequately represent local-scale spatial patterns (e.g., averaged over 1 km2). In this paper, a hybrid regression-interpolation methodology is developed to enhance the representation of local-scale spatiotemporal patterns with an application to Tucson, Arizona. The mapping of local patterns is enhanced with pre-interpolation regression modeling of local-scale deviation-from-mean variability, preserving variation in the monitor data that is ubiquitous across the modeling domain (i.e., the areal mean). The model is trained on several years of deviation-from-mean hourly O3 data, and predictor variables are developed using theoretically and empirically derived proxy regression variables. The regression model explains a significant proportion of the variation in the data (r2 = 0.54), with an average error of 7.1 ppb. When augmented with the areal mean, the r2 of the pre-interpolation model increases to 0.847. Model residuals are then spatially interpolated to the extents of the modeling domain. Final concentration estimate maps are the summation of areal mean, regression, and spatially interpolated surfaces, preserving absolute values at monitor locations.