Global synthesis of two decades of research on improving PM2.5 estimation models from remote sensing and data science perspectives

Identifying the proxy data and techniques that could yield greater PM2.5 estimation accuracy remains challenging due to the lack of general knowledge on the relative performance of covariates and modeling methods on the PM2.5 modeling accuracy. To bridge a critical gap in the contemporary literature on how to improve satellite -based PM2.5 estimation models, we have systematically reviewed global efforts at PM2.5 estimation between 2000 and 2020 using big data analytics, focusing on a critical evaluation of PM2.5 modeling accuracy im-provements in response to varying methods and covariates, mainly from remote sensing and data science per-spectives. Using an automated literature classification deep learning method, we identified 833 publications with fair inter-comparisons among relevant PM2.5 estimation models. The inter-compared modeling accuracy metrics were then retrieved and aggregated with respect to each modeling factor to form a global research synthesis database. The synthesized results highlighted that greater PM2.5 modeling accuracy (i.e., a mean R2 improvement by 6.2%) can be attained by incorporating aerosol optical depth (AOD); this benefit could even be doubled by applying fine-mode AOD. Further enhancements could be achieved by conducting vertical and humidity cor-rections to AOD (10.41%), as well as including relevant meteorological and socioeconomic factors that are highly associated with PM2.5 variations (12%, on average). By contrast, improving spatial resolution of satellite AOD products and temporal resolution of ground PM2.5 measurements are unlikely to yield better PM2.5 estimation accuracy (improvement of 1.2%). Although advanced machine learning algorithms with better generalization capacity is more cost-effective, the accuracy improvements vary substantially across studies, largely depending on data sources, study regions, and even seasons. The statistical findings underscore the critical importance of domain knowledge and advancements in remote sensing and data science for improving PM2.5 modeling accu-racy. Finally, with inferred knowledge from this quantitative evidence, we discuss the grand challenges and possible solutions for future research directions toward better PM2.5 prediction, mapping, and exposure assessment.