Spatial modeling of PM2.5 concentrations with a multifactoral radial basis function neural network.

Accurate measurements of PM2.5 concentration over time and space are especially critical for reducing adverse health outcomes. However, sparsely stationary monitoring sites considerably hinder the ability to effectively characterize observed concentrations. Utilizing data on meteorological and land-related factors, this study introduces a radial basis function (RBF) neural network method for estimating PM2.5 concentrations based on sparse observed inputs. The state of Texas in the USA was selected as the study area. Performance of the RBF models was evaluated by statistic indices including mean square error, mean absolute error, mean relative deviation, and the correlation coefficient. Results show that the annual PM2.5 concentrations estimated by the RBF models with meteorological factors and/or land-related factors were markedly closer to the observed concentrations. RBF models with combined meteorological and land-related factors achieved best performance relative to ones with either type of these factors only. It can be concluded that meteorological factors and land-related factors are useful for articulating the variation of PM2.5 concentration in a given study area. With these covariate factors, the RBF neural network can effectively estimate PM2.5 concentrations with acceptable accuracy under the condition of sparse monitoring stations. The improved accuracy of air concentration estimation would greatly benefit epidemiological and environmental studies in characterizing local air pollution and in helping reduce population exposures for areas with limited availability of air quality data.