Critical Issues In Land Use Regression Modeling Of PM_2.5 Concentration

ABSTRACT:The lack of in-depth exploration on the space scale dependence of modeling variables, the migration feature and accuracy evaluation methods on PM2.5LUR simulating research results in lots of defects, such as the effection of spatial scale and the migration accuracy of LUR model are not clear, the evaluation system of model accuracy is imperfect. All those defects become a serious impediment for expanding the application scope of LUR model, improving model simulation accuracy and playing the advantage of higher spatial resolution. It is difficult to provide accurate, comprehensive, high-precision PM2.5pollution data for researchers and decision-makers in such a case. Therefore, In this study, simulation of PM2.5annual average concentrations on Eastern Area in2006was chosen as an example. Based on common geographic features characteristic variables, we used correlation analysis and model fit comparison method to analyze the spatial scales dependence characteristics of PM2.5LUR’s modeling variables and the migration characteristics of LUR model, and improve the evaluation of the simulation accuracy of LUR models after the introduction of image information entropy and profile analysis. Results shows that:(1) Based on the correlation coefficient, we determine the contribution rate of characteristic variables for PM2.5pollution. The results show that the pollution contributions of characteristic variables changed significantly with spatial scale. Each characteristic variable would reach the most relevant in a unique space scale. The LUR model based on characteristic variables under optimal spatial scales fitting PM2.5concentration is better than those under any other uniform spatial scales (100m-10km, etc.)(R2fitting maximum of0.37), and the performance is more stable, the simulation result is more accuracy.(2) The comparing result of LUR models under region/sub-region indicates that LUR models in sub-regions would fit better the spatial variation of PM2.5concentrations. Even based on same characteristic variables, the regional LUR model could be quite different. Affected by the diversity of variables’ number and structure between regional LUR models, the migration results was poorer between sub-regions, while the effect of migration from the region to the sub-regions performs better, in part of the sub-region (sub-region1) the migration effect even exceed the local model.(3) Using discrete point and continuous surface based evaluation method to compare the performance of most fitted LUR and region LUR model on simulating PM2.5concentration of eastern United States, the results indicate that traditional point-based evaluation results vary greatly, sometimes even the opposite, affected by inspection site number and spatial distribution patterns (eg, the evaluation results based on test site3was opposite). In contrast, continuous surface based evaluation method can increase the stability of the model evaluation results. The assessment result based on continuous surface method is more reliable.