Jin, Zhihao, Qiang Pu, Nathan Janechek, Huanxin Zhang, Jun Wang, Howard Chang, and Yang Liu. A MAIA-like modeling framework to estimate PM2.5 mass and speciation concentrations with uncertainty. Remote Sensing of Environment 303 (2024): 113995. 

Read Online at https://authors.elsevier.com/a/1iSUj7qzT3DS-

Ambient fine particulate matter (PM2.5) is strongly associated with various adverse health outcomes. However, the lack of extensive PM2.5 measurements, and especially of its components, hinders the assessment of negative health effects caused by PM2.5 in many parts of the world. To address this issue, a new satellite instrument, the Multi-angle Imager for Aerosols (MAIA), with improved design for providing aerosol optical depth (AOD) of high quality, will be helpful in determining concentrations of total and speciated PM2.5. According to the retrieval algorithm of MAIA particulate matter (PM) products, level 2 (L2) PM products are generated based on MAIA AOD on days of observation. Bias-corrected chemical transport model (CTM) outputs are then merged with the L2 PM products to fill their gaps using a Bayesian Model Averaging (BMA) ensemble framework. This process creates the MAIA Level 4 (L4) gap-filled PM products. In this study, we aim to implement the MAIA framework and validate its feasibility after the launch of the MAIA satellite instrument. We used both Bayesian hierarchical model (BHM) and a Bayesian additive regression tree (BART) to predict L2 and CTM-based daily 1 × 1 km2 PM2.5 mass and speciation concentrations, along with prediction uncertainties, over the MAIA Primary Target Area in the Northeastern US in 2018. We then employed the BMA ensemble model to combine the L2 and CTM-based PM2.5 mass predictions to fill gaps in L2 PM2.5 mass and produce Level 4 (L4) gap-filled PM2.5 mass. Our cross-validation experiments showed that both the BHM (R2 ranging from 0.60 to 0.82) and BART (R2 ranging from 0.59 to 0.79) models performed well in predicting CTM-based PM2.5 speciation, with better results for sulfate, organic carbon, and elemental carbon. At the stage of L4 PM2.5 mass predictions, both BHM-based and BART-based BMA ensemble models demonstrated improved performance with their traditional R2 of 0.81 and 0.73, surpassing the input L2 and CTM-based PM2.5 mass. Additionally, our models showed excellent prediction uncertainty control with the coverage rates of 95% posterior prediction confidence interval associated of concentration estimates to be 95% for BHM and 75% for BART across PM2.5 mass and speciation. Results from the proposed modeling techniques contribute to a deeper understanding of the health effects of PM2.5 for future epidemiological studies and provide insights into the MAIA mission for producing improved PM products for health research.

Figure. Framework of the MAIA-like Level 4 gap-filled PM products