1. A spatially varying coefficient model for mapping PM10 air quality at the European scale
- Author
-
Martijn Schaap, Andrew O. Finley, Alfred Stein, Nicholas A. S. Hamm, and Department of Earth Observation Science
- Subjects
Atmospheric Science ,Meteorology ,Covariance function ,Geostatistics ,Environment ,symbols.namesake ,PM10 ,Urban Development ,Linear regression ,Bayesian hierarchical modeling ,Spatially varying coefficient (SVC) ,Limit (mathematics) ,Built Environment ,Air quality index ,Model evaluation ,General Environmental Science ,Earth / Environmental ,Markov chain Monte Carlo ,CAS - Climate, Air and Sustainability ,symbols ,Environmental science ,ELSS - Earth, Life and Social Sciences ,Scale (map) ,LOTOS-EUROS - Abstract
Particulate matter (PM) air quality in Europe has improved substantially over the past decades, but it still poses a significant threat to human health. Accurate regional scale maps of PM10 concentrations are needed for monitoring progress in mitigation strategies and monitoring compliance with statutory limit values. Chemistry transport models (CTM) use emission databases and simulate the transport and deposition of pollutants. They deliver such maps but are known to be inaccurate. A promising approach is to use geostatistics to model the relationship between the in situ observations and the CTM. This has been shown to be more accurate than using either observations or CTM's alone. This paper presents a spatially varying coefficients (SVC) geostatistical model as an extension of the standard spatially varying intercept (SVI) geostatistical model. SVC allowed the regression coefficient to vary spatially according to a covariance function, the parameters of which were estimated from the data. It was built as a Bayesian hierarchical model and implemented using Markov chain Monte Carlo. The procedure was applied to Airbase PM10 observations and LOTOS-EUROS simulated PM10 for central, southern and eastern Europe. Model-fit diagnostics showed that SVC delivered a better fit to the data than SVI. Mapping the spatially varying coefficients allowed identification of the locations where the CTM performed well or poorly. This could be used for objective CTM evaluation purposes. The posterior predictive simulations were also used to map median PM10 concentrations as well as the probability of exceeding the 50 μg m−3 EU daily PM10 concentration threshold. Although posterior median prediction accuracy was similar for SVI and SVC, SVC better modelled the process and yielded narrower credible intervals. As such, SVC was more appropriate for quantifying uncertainty and for mapping threshold exceedances. The resulting maps may be used to guide air quality assessment and mitigation strategies, including those related to health impacts
- Published
- 2015