Comparison of Different Modeling Strategies for Estimating Long-Term PM2.5 Exposure Using MAIAC (Multiangle Implementation of Atmospheric Correction) AOD in China

The evidence on long-term effects on health is vital for establishing a causal association and assessing the disease burden related to particulate air pollution. However, the remaining problem of aerosol retrievals’ severe absences strikes a serious blow at long-term PM2.5 exposure estimation. The skyrocketed sample size with finer spatial resolution of aerosol products further increases the difficulty to resolve this problem. For finding out a less time-consuming, higher coverage and more precise AOD interpolation, three strategies were proposed and compared in mainland China in 2014–2018, including DI (Directly Interpolate long-term averaged AOD then predict long-term PM2.5), AODS2L (integrate Short-term interpolated AOD into Long-term AOD then predict Long-term PM2.5) and PMS2L (integrate Short-term PM2.5 predicted by Short-term AOD into Long-term PM2.5). AOD coverage varies widely in spatial (0–84.56%) and temporal (10.17–22.57%) distribution, while the places with monsoon climate (Southeast China) and summer (typhoon season) had the lowest coverage. For the monthly or annual PM2.5 estimation, three strategies can fix almost all of missing gaps with the coverage rising from 15.65% to 98.55%–98.79% and provide accurate long-term PM2.5 predictions (monthly CV R2: 0.91, 0.92 and 0.93; annual CV R2: 0.83, 0.85 and 0.92 for DI, AODS2L and PMS2L, respectively). The well-behaved strategies of AODS2L and PMS2L spent much more running time. How to balance the model performance and consuming time is a key point in long-term air pollution research. For the large-scale population-based study, using DI strategy can save time and guarantee certain accuracy. For the individual-based study, PMS2L is the more accurate and flexible for individual exposure but the most time-consuming choice.