Comparative Performance of the LUR, ANN, and BME Techniques in the Multiscale Spatiotemporal Mapping of PM2.5 Concentrations in North China

Abstract

Fine particle matter (PM2.5) has been receiving increasing attention by the government due to its considerable adverse health effects, especially in the north part of China. Even though a number of techniques of estimating PM2.5 exposure have been developed, what is still lacking is a systematic comparison of commonly used techniques based on classical statistics, artificial intelligence, and geostatistics. To address this need, the land use regression (LUR), the artificial neural networks (ANN), and the Bayesian maximum entropy (BME) techniques were all used to map the space–time PM2.5 concentration distribution in the highly polluted Jing-Jin-Ji region (Huabei plain of North China) during the period June 2015–May 2016. The tenfold cross-validation analysis and the entropic information theory were used to evaluate numerically the performance of the three techniques at monthly, seasonal, and annual time scales. Our results showed that the performance of each mapping technique was affected by the temporal scale and the degree of spatial heterogeneity. All three techniques were suitable for low temporal resolution (annual) datasets with low spatial variability. BME also showed a noticeable ability to analyze higher temporal resolution (monthly) datasets exhibiting high spatial heterogeneity. BME involved a single dependent variable (PM2.5) and generated complete (full-coverage) space–time PM2.5 maps, whereas LUR and ANN produced incomplete maps because of lacking independent variables (such as satellite data). Due to its self-learning feature, ANN showed better modeling performance than LUR and produced more informative maps. Overall, the ANN and BME techniques perform better than the LUR technique.